summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorerdgeist <erdgeist@erdgeist.org>2025-08-15 12:42:40 +0200
committererdgeist <erdgeist@erdgeist.org>2025-08-15 12:42:40 +0200
commit30325d24d107dbf133da39f7c96d1510fd1c9449 (patch)
tree932baa5b2a4475821f16dccf9e3e05011daa6d92
parent9022d768021bbe15c7815cc6f8b64218b46f0e10 (diff)
Bump to codec2 version 1.2.0erdgeist-bump-to-1.2.0
Diffstat
-rw-r--r--Makefile9
-rw-r--r--_kiss_fft_guts.h234
-rw-r--r--bpf.h120
-rw-r--r--bpfb.h119
-rw-r--r--codebook.c82
-rw-r--r--codebookd.c82
-rw-r--r--codebookge.c10
-rw-r--r--codebookjmv.c (renamed from codebookjvm.c)28
-rw-r--r--codebooklspmelvq.c247
-rw-r--r--codebookmel.c145
-rw-r--r--codebooknewamp1.c16
-rw-r--r--codebooknewamp1_energy.c8
-rw-r--r--codebooknewamp2.c2
-rw-r--r--codec2.c3331
-rw-r--r--codec2.h102
-rw-r--r--codec2_fft.c112
-rw-r--r--codec2_fft.h111
-rw-r--r--codec2_internal.h129
-rw-r--r--comp_prim.h109
-rw-r--r--defines.h89
-rw-r--r--interp.c247
-rw-r--r--interp.h16
-rw-r--r--kiss_fft.c714
-rw-r--r--kiss_fft.h60
-rw-r--r--kiss_fftr.c272
-rw-r--r--kiss_fftr.h25
-rw-r--r--lpc.c223
-rw-r--r--lpc.h8
-rw-r--r--lsp.c354
-rw-r--r--lsp.h2
-rw-r--r--machdep.h12
-rw-r--r--main.c2
-rw-r--r--mbest.c194
-rw-r--r--mbest.h23
-rw-r--r--newamp1.c770
-rw-r--r--newamp1.h86
-rw-r--r--newamp2.c570
-rw-r--r--newamp2.h89
-rw-r--r--nlp.c609
-rw-r--r--nlp.h5
-rw-r--r--os.h82
-rw-r--r--pack.c104
-rw-r--r--phase.c242
-rw-r--r--phase.h6
-rw-r--r--postfilter.c58
-rw-r--r--postfilter.h2
-rw-r--r--quantise.c1992
-rw-r--r--quantise.h116
-rw-r--r--sine.c591
-rw-r--r--sine.h13
-rwxr-xr-xstripdown.sh17
-rw-r--r--version.h8
52 files changed, 4804 insertions, 7793 deletions
diff --git a/Makefile b/Makefile
index a716871..2cc38c4 100644
--- a/Makefile
+++ b/Makefile
@@ -1,11 +1,10 @@
1CFLAGS = -Wall -Wno-strict-overflow -std=gnu11 -fPIC -g -O2 -I. 1CFLAGS = -Wall -Wno-strict-overflow -std=gnu11 -fPIC -g -O2 -I. -lm
2CFLAGS += -DHORUS_L2_RX -DINTERLEAVER -DRUN_TIME_TABLES -DSCRAMBLER -Dcodec2_EXPORTS 2CFLAGS += -DHORUS_L2_RX -DINTERLEAVER -DRUN_TIME_TABLES -DSCRAMBLER -Dcodec2_EXPORTS
3CFLAGS += -Wno-incompatible-pointer-types-discards-qualifiers 3CFLAGS += -Wno-incompatible-pointer-types-discards-qualifiers
4 4
5LIBSRC=codebook.c codebookd.c codebookge.c codebookmel.c codebooklspmelvq.c codebookjvm.c \ 5LIBSRC=codebook.c codebookd.c codebookge.c codebookjmv.c codebooknewamp1_energy.c \
6 codebooknewamp1.c codebooknewamp1_energy.c codebooknewamp2.c codebooknewamp2_energy.c \ 6 codebooknewamp1.c codec2.c codec2_fft.c interp.c kiss_fft.c kiss_fftr.c \
7 codec2.c codec2_fft.c interp.c kiss_fft.c kiss_fftr.c lpc.c lsp.c mbest.c \ 7 lpc.c lsp.c mbest.c newamp1.c nlp.c pack.c phase.c postfilter.c quantise.c sine.c
8 newamp1.c newamp2.c nlp.c pack.c phase.c postfilter.c quantise.c sine.c
9 8
10all: main 9all: main
11 10
diff --git a/_kiss_fft_guts.h b/_kiss_fft_guts.h
index 1c62a34..16305af 100644
--- a/_kiss_fft_guts.h
+++ b/_kiss_fft_guts.h
@@ -3,33 +3,49 @@ Copyright (c) 2003-2010, Mark Borgerding
3 3
4All rights reserved. 4All rights reserved.
5 5
6Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: 6Redistribution and use in source and binary forms, with or without modification,
7 7are permitted provided that the following conditions are met:
8 * Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. 8
9 * Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. 9 * Redistributions of source code must retain the above copyright notice,
10 * Neither the author nor the names of any contributors may be used to endorse or promote products derived from this software without specific prior written permission. 10this list of conditions and the following disclaimer.
11 11 * Redistributions in binary form must reproduce the above copyright notice,
12THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 12this list of conditions and the following disclaimer in the documentation and/or
13other materials provided with the distribution.
14 * Neither the author nor the names of any contributors may be used to
15endorse or promote products derived from this software without specific prior
16written permission.
17
18THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
19ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
20WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
21DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
22ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
23(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
24LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
25ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
27SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
13*/ 28*/
14 29
15/* kiss_fft.h 30/* kiss_fft.h
16 defines kiss_fft_scalar as either short or a float type 31 defines kiss_fft_scalar as either short or a float type
17 and defines 32 and defines
18 typedef struct { kiss_fft_scalar r; kiss_fft_scalar i; }kiss_fft_cpx; */ 33 typedef struct { kiss_fft_scalar r; kiss_fft_scalar i; }kiss_fft_cpx; */
19#include "kiss_fft.h"
20#include <limits.h> 34#include <limits.h>
21 35
36#include "kiss_fft.h"
37
22#define MAXFACTORS 32 38#define MAXFACTORS 32
23/* e.g. an fft of length 128 has 4 factors 39/* e.g. an fft of length 128 has 4 factors
24 as far as kissfft is concerned 40 as far as kissfft is concerned
25 4*4*4*2 41 4*4*4*2
26 */ 42 */
27 43
28struct kiss_fft_state{ 44struct kiss_fft_state {
29 int nfft; 45 int nfft;
30 int inverse; 46 int inverse;
31 int factors[2*MAXFACTORS]; 47 int factors[2 * MAXFACTORS];
32 kiss_fft_cpx twiddles[1]; 48 kiss_fft_cpx twiddles[1];
33}; 49};
34 50
35/* 51/*
@@ -42,123 +58,137 @@ struct kiss_fft_state{
42 C_ADDTO( res , a) : res += a 58 C_ADDTO( res , a) : res += a
43 * */ 59 * */
44#ifdef FIXED_POINT 60#ifdef FIXED_POINT
45#if (FIXED_POINT==32) 61#if (FIXED_POINT == 32)
46# define FRACBITS 31 62#define FRACBITS 31
47# define SAMPPROD int64_t 63#define SAMPPROD int64_t
48#define SAMP_MAX 2147483647 64#define SAMP_MAX 2147483647
49#else 65#else
50# define FRACBITS 15 66#define FRACBITS 15
51# define SAMPPROD int32_t 67#define SAMPPROD int32_t
52#define SAMP_MAX 32767 68#define SAMP_MAX 32767
53#endif 69#endif
54 70
55#define SAMP_MIN -SAMP_MAX 71#define SAMP_MIN -SAMP_MAX
56 72
57#if defined(CHECK_OVERFLOW) 73#if defined(CHECK_OVERFLOW)
58# define CHECK_OVERFLOW_OP(a,op,b) \ 74#define CHECK_OVERFLOW_OP(a, op, b) \
59 if ( (SAMPPROD)(a) op (SAMPPROD)(b) > SAMP_MAX || (SAMPPROD)(a) op (SAMPPROD)(b) < SAMP_MIN ) { \ 75 if ((SAMPPROD)(a)op(SAMPPROD)(b) > SAMP_MAX || \
60 fprintf(stderr,"WARNING:overflow @ " __FILE__ "(%d): (%d " #op" %d) = %ld\n",__LINE__,(a),(b),(SAMPPROD)(a) op (SAMPPROD)(b) ); } 76 (SAMPPROD)(a)op(SAMPPROD)(b) < SAMP_MIN) { \
77 fprintf(stderr, \
78 "WARNING:overflow @ " __FILE__ "(%d): (%d " #op " %d) = %ld\n", \
79 __LINE__, (a), (b), (SAMPPROD)(a)op(SAMPPROD)(b)); \
80 }
61#endif 81#endif
62 82
63 83#define smul(a, b) ((SAMPPROD)(a) * (b))
64# define smul(a,b) ( (SAMPPROD)(a)*(b) ) 84#define sround(x) (kiss_fft_scalar)(((x) + (1 << (FRACBITS - 1))) >> FRACBITS)
65# define sround( x ) (kiss_fft_scalar)( ( (x) + (1<<(FRACBITS-1)) ) >> FRACBITS ) 85
66 86#define S_MUL(a, b) sround(smul(a, b))
67# define S_MUL(a,b) sround( smul(a,b) ) 87
68 88#define C_MUL(m, a, b) \
69# define C_MUL(m,a,b) \ 89 do { \
70 do{ (m).r = sround( smul((a).r,(b).r) - smul((a).i,(b).i) ); \ 90 (m).r = sround(smul((a).r, (b).r) - smul((a).i, (b).i)); \
71 (m).i = sround( smul((a).r,(b).i) + smul((a).i,(b).r) ); }while(0) 91 (m).i = sround(smul((a).r, (b).i) + smul((a).i, (b).r)); \
72 92 } while (0)
73# define DIVSCALAR(x,k) \ 93
74 (x) = sround( smul( x, SAMP_MAX/k ) ) 94#define DIVSCALAR(x, k) (x) = sround(smul(x, SAMP_MAX / k))
75 95
76# define C_FIXDIV(c,div) \ 96#define C_FIXDIV(c, div) \
77 do { DIVSCALAR( (c).r , div); \ 97 do { \
78 DIVSCALAR( (c).i , div); }while (0) 98 DIVSCALAR((c).r, div); \
79 99 DIVSCALAR((c).i, div); \
80# define C_MULBYSCALAR( c, s ) \ 100 } while (0)
81 do{ (c).r = sround( smul( (c).r , s ) ) ;\ 101
82 (c).i = sround( smul( (c).i , s ) ) ; }while(0) 102#define C_MULBYSCALAR(c, s) \
83 103 do { \
84#else /* not FIXED_POINT*/ 104 (c).r = sround(smul((c).r, s)); \
85 105 (c).i = sround(smul((c).i, s)); \
86# define S_MUL(a,b) ( (a)*(b) ) 106 } while (0)
87#define C_MUL(m,a,b) \ 107
88 do{ (m).r = (a).r*(b).r - (a).i*(b).i;\ 108#else /* not FIXED_POINT*/
89 (m).i = (a).r*(b).i + (a).i*(b).r; }while(0) 109
90# define C_FIXDIV(c,div) /* NOOP */ 110#define S_MUL(a, b) ((a) * (b))
91# define C_MULBYSCALAR( c, s ) \ 111#define C_MUL(m, a, b) \
92 do{ (c).r *= (s);\ 112 do { \
93 (c).i *= (s); }while(0) 113 (m).r = (a).r * (b).r - (a).i * (b).i; \
114 (m).i = (a).r * (b).i + (a).i * (b).r; \
115 } while (0)
116#define C_FIXDIV(c, div) /* NOOP */
117#define C_MULBYSCALAR(c, s) \
118 do { \
119 (c).r *= (s); \
120 (c).i *= (s); \
121 } while (0)
94#endif 122#endif
95 123
96#ifndef CHECK_OVERFLOW_OP 124#ifndef CHECK_OVERFLOW_OP
97# define CHECK_OVERFLOW_OP(a,op,b) /* noop */ 125#define CHECK_OVERFLOW_OP(a, op, b) /* noop */
98#endif 126#endif
99 127
100#define C_ADD( res, a,b)\ 128#define C_ADD(res, a, b) \
101 do { \ 129 do { \
102 CHECK_OVERFLOW_OP((a).r,+,(b).r)\ 130 CHECK_OVERFLOW_OP((a).r, +, (b).r) \
103 CHECK_OVERFLOW_OP((a).i,+,(b).i)\ 131 CHECK_OVERFLOW_OP((a).i, +, (b).i) \
104 (res).r=(a).r+(b).r; (res).i=(a).i+(b).i; \ 132 (res).r = (a).r + (b).r; \
105 }while(0) 133 (res).i = (a).i + (b).i; \
106#define C_SUB( res, a,b)\ 134 } while (0)
107 do { \ 135#define C_SUB(res, a, b) \
108 CHECK_OVERFLOW_OP((a).r,-,(b).r)\ 136 do { \
109 CHECK_OVERFLOW_OP((a).i,-,(b).i)\ 137 CHECK_OVERFLOW_OP((a).r, -, (b).r) \
110 (res).r=(a).r-(b).r; (res).i=(a).i-(b).i; \ 138 CHECK_OVERFLOW_OP((a).i, -, (b).i) \
111 }while(0) 139 (res).r = (a).r - (b).r; \
112#define C_ADDTO( res , a)\ 140 (res).i = (a).i - (b).i; \
113 do { \ 141 } while (0)
114 CHECK_OVERFLOW_OP((res).r,+,(a).r)\ 142#define C_ADDTO(res, a) \
115 CHECK_OVERFLOW_OP((res).i,+,(a).i)\ 143 do { \
116 (res).r += (a).r; (res).i += (a).i;\ 144 CHECK_OVERFLOW_OP((res).r, +, (a).r) \
117 }while(0) 145 CHECK_OVERFLOW_OP((res).i, +, (a).i) \
118 146 (res).r += (a).r; \
119#define C_SUBFROM( res , a)\ 147 (res).i += (a).i; \
120 do {\ 148 } while (0)
121 CHECK_OVERFLOW_OP((res).r,-,(a).r)\ 149
122 CHECK_OVERFLOW_OP((res).i,-,(a).i)\ 150#define C_SUBFROM(res, a) \
123 (res).r -= (a).r; (res).i -= (a).i; \ 151 do { \
124 }while(0) 152 CHECK_OVERFLOW_OP((res).r, -, (a).r) \
125 153 CHECK_OVERFLOW_OP((res).i, -, (a).i) \
154 (res).r -= (a).r; \
155 (res).i -= (a).i; \
156 } while (0)
126 157
127#ifdef FIXED_POINT 158#ifdef FIXED_POINT
128# define KISS_FFT_COS(phase) floorf(.5+SAMP_MAX * cosf (phase)) 159#define KISS_FFT_COS(phase) floorf(.5 + SAMP_MAX * cosf(phase))
129# define KISS_FFT_SIN(phase) floorf(.5+SAMP_MAX * sinf (phase)) 160#define KISS_FFT_SIN(phase) floorf(.5 + SAMP_MAX * sinf(phase))
130# define HALF_OF(x) ((x)>>1) 161#define HALF_OF(x) ((x) >> 1)
131#elif defined(USE_SIMD) 162#elif defined(USE_SIMD)
132# define KISS_FFT_COS(phase) _mm_set1_ps( cosf(phase) ) 163#define KISS_FFT_COS(phase) _mm_set1_ps(cosf(phase))
133# define KISS_FFT_SIN(phase) _mm_set1_ps( sinf(phase) ) 164#define KISS_FFT_SIN(phase) _mm_set1_ps(sinf(phase))
134# define HALF_OF(x) ((x)*_mm_set1_ps(.5)) 165#define HALF_OF(x) ((x)*_mm_set1_ps(.5))
135#else 166#else
136# define KISS_FFT_COS(phase) (kiss_fft_scalar) cosf(phase) 167#define KISS_FFT_COS(phase) (kiss_fft_scalar) cosf(phase)
137# define KISS_FFT_SIN(phase) (kiss_fft_scalar) sinf(phase) 168#define KISS_FFT_SIN(phase) (kiss_fft_scalar) sinf(phase)
138# define HALF_OF(x) ((x)*.5) 169#define HALF_OF(x) ((x)*.5)
139#endif 170#endif
140 171
141#define kf_cexp(x,phase) \ 172#define kf_cexp(x, phase) \
142 do{ \ 173 do { \
143 (x)->r = KISS_FFT_COS(phase);\ 174 (x)->r = KISS_FFT_COS(phase); \
144 (x)->i = KISS_FFT_SIN(phase);\ 175 (x)->i = KISS_FFT_SIN(phase); \
145 }while(0) 176 } while (0)
146
147 177
148/* a debugging function */ 178/* a debugging function */
149#define pcpx(c)\ 179#define pcpx(c) \
150 fprintf(stderr,"%g + %gi\n",(double)((c)->r),(double)((c)->i) ) 180 fprintf(stderr, "%g + %gi\n", (double)((c)->r), (double)((c)->i))
151
152 181
153#ifdef KISS_FFT_USE_ALLOCA 182#ifdef KISS_FFT_USE_ALLOCA
154// define this to allow use of alloca instead of malloc for temporary buffers 183// define this to allow use of alloca instead of malloc for temporary buffers
155// Temporary buffers are used in two case: 184// Temporary buffers are used in two case:
156// 1. FFT sizes that have "bad" factors. i.e. not 2,3 and 5 185// 1. FFT sizes that have "bad" factors. i.e. not 2,3 and 5
157// 2. "in-place" FFTs. Notice the quotes, since kissfft does not really do an in-place transform. 186// 2. "in-place" FFTs. Notice the quotes, since kissfft does not really do an
187// in-place transform.
158#include <alloca.h> 188#include <alloca.h>
159#define KISS_FFT_TMP_ALLOC(nbytes) alloca(nbytes) 189#define KISS_FFT_TMP_ALLOC(nbytes) alloca(nbytes)
160#define KISS_FFT_TMP_FREE(ptr) 190#define KISS_FFT_TMP_FREE(ptr)
161#else 191#else
162#define KISS_FFT_TMP_ALLOC(nbytes) KISS_FFT_MALLOC(nbytes) 192#define KISS_FFT_TMP_ALLOC(nbytes) KISS_FFT_MALLOC(nbytes)
163#define KISS_FFT_TMP_FREE(ptr) KISS_FFT_FREE(ptr) 193#define KISS_FFT_TMP_FREE(ptr) KISS_FFT_FREE(ptr)
164#endif 194#endif
diff --git a/bpf.h b/bpf.h
index 0b0a6e2..d74e059 100644
--- a/bpf.h
+++ b/bpf.h
@@ -1,106 +1,18 @@
1#define BPF_N 101 1#define BPF_N 101
2 2
3float bpf[]={ 3float bpf[] = {
4 0.002174, 4 0.002174, 0.003245, 0.002147, 0.001866, 0.002764, 0.000567, -0.001641,
5 0.003245, 5 -0.000565, -0.002415, -0.005837, -0.003620, -0.002828, -0.006268, -0.002787,
6 0.002147, 6 0.001963, -0.001234, 0.001446, 0.009200, 0.005331, 0.003521, 0.011821,
7 0.001866, 7 0.006951, -0.002015, 0.005137, 0.001828, -0.013390, -0.007058, -0.003273,
8 0.002764, 8 -0.020458, -0.014321, 0.001751, -0.012891, -0.009730, 0.018993, 0.008544,
9 0.000567, 9 0.000534, 0.035755, 0.029074, -0.001192, 0.030852, 0.030983, -0.029834,
10 -0.001641, 10 -0.009550, 0.011945, -0.081971, -0.082875, 0.000423, -0.133526, -0.211778,
11 -0.000565, 11 0.182628, 0.514906, 0.182628, -0.211778, -0.133526, 0.000423, -0.082875,
12 -0.002415, 12 -0.081971, 0.011945, -0.009550, -0.029834, 0.030983, 0.030852, -0.001192,
13 -0.005837, 13 0.029074, 0.035755, 0.000534, 0.008544, 0.018993, -0.009730, -0.012891,
14 -0.003620, 14 0.001751, -0.014321, -0.020458, -0.003273, -0.007058, -0.013390, 0.001828,
15 -0.002828, 15 0.005137, -0.002015, 0.006951, 0.011821, 0.003521, 0.005331, 0.009200,
16 -0.006268, 16 0.001446, -0.001234, 0.001963, -0.002787, -0.006268, -0.002828, -0.003620,
17 -0.002787, 17 -0.005837, -0.002415, -0.000565, -0.001641, 0.000567, 0.002764, 0.001866,
18 0.001963, 18 0.002147, 0.003245, 0.002174};
19 -0.001234,
20 0.001446,
21 0.009200,
22 0.005331,
23 0.003521,
24 0.011821,
25 0.006951,
26 -0.002015,
27 0.005137,
28 0.001828,
29 -0.013390,
30 -0.007058,
31 -0.003273,
32 -0.020458,
33 -0.014321,
34 0.001751,
35 -0.012891,
36 -0.009730,
37 0.018993,
38 0.008544,
39 0.000534,
40 0.035755,
41 0.029074,
42 -0.001192,
43 0.030852,
44 0.030983,
45 -0.029834,
46 -0.009550,
47 0.011945,
48 -0.081971,
49 -0.082875,
50 0.000423,
51 -0.133526,
52 -0.211778,
53 0.182628,
54 0.514906,
55 0.182628,
56 -0.211778,
57 -0.133526,
58 0.000423,
59 -0.082875,
60 -0.081971,
61 0.011945,
62 -0.009550,
63 -0.029834,
64 0.030983,
65 0.030852,
66 -0.001192,
67 0.029074,
68 0.035755,
69 0.000534,
70 0.008544,
71 0.018993,
72 -0.009730,
73 -0.012891,
74 0.001751,
75 -0.014321,
76 -0.020458,
77 -0.003273,
78 -0.007058,
79 -0.013390,
80 0.001828,
81 0.005137,
82 -0.002015,
83 0.006951,
84 0.011821,
85 0.003521,
86 0.005331,
87 0.009200,
88 0.001446,
89 -0.001234,
90 0.001963,
91 -0.002787,
92 -0.006268,
93 -0.002828,
94 -0.003620,
95 -0.005837,
96 -0.002415,
97 -0.000565,
98 -0.001641,
99 0.000567,
100 0.002764,
101 0.001866,
102 0.002147,
103 0.003245,
104 0.002174
105};
106
diff --git a/bpfb.h b/bpfb.h
index 73b1a69..e3a958f 100644
--- a/bpfb.h
+++ b/bpfb.h
@@ -1,105 +1,18 @@
1#define BPFB_N 101 1#define BPFB_N 101
2 2
3float bpfb[]={ 3float bpfb[] = {
4 0.003795, 4 0.003795, 0.006827, 0.002261, 0.002523, 0.005758, -0.000264, -0.000674,
5 0.006827, 5 0.003113, -0.004144, -0.004923, 0.000043, -0.008017, -0.008711, -0.001802,
6 0.002261, 6 -0.010210, -0.010428, -0.000899, -0.009413, -0.009072, 0.003469, -0.005335,
7 0.002523, 7 -0.004828, 0.010724, 0.000941, 0.000708, 0.018957, 0.007084, 0.004825,
8 0.005758, 8 0.025418, 0.010147, 0.004452, 0.027434, 0.007550, -0.002861, 0.023483,
9 -0.000264, 9 -0.001944, -0.018138, 0.014122, -0.017583, -0.040768, 0.002598, -0.036604,
10 -0.000674, 10 -0.069541, -0.004273, -0.054876, -0.107289, 0.010068, -0.068052, -0.200119,
11 0.003113, 11 0.207287, 0.597150, 0.207287, -0.200119, -0.068052, 0.010068, -0.107289,
12 -0.004144, 12 -0.054876, -0.004273, -0.069541, -0.036604, 0.002598, -0.040768, -0.017583,
13 -0.004923, 13 0.014122, -0.018138, -0.001944, 0.023483, -0.002861, 0.007550, 0.027434,
14 0.000043, 14 0.004452, 0.010147, 0.025418, 0.004825, 0.007084, 0.018957, 0.000708,
15 -0.008017, 15 0.000941, 0.010724, -0.004828, -0.005335, 0.003469, -0.009072, -0.009413,
16 -0.008711, 16 -0.000899, -0.010428, -0.010210, -0.001802, -0.008711, -0.008017, 0.000043,
17 -0.001802, 17 -0.004923, -0.004144, 0.003113, -0.000674, -0.000264, 0.005758, 0.002523,
18 -0.010210, 18 0.002261, 0.006827, 0.003795}; \ No newline at end of file
19 -0.010428,
20 -0.000899,
21 -0.009413,
22 -0.009072,
23 0.003469,
24 -0.005335,
25 -0.004828,
26 0.010724,
27 0.000941,
28 0.000708,
29 0.018957,
30 0.007084,
31 0.004825,
32 0.025418,
33 0.010147,
34 0.004452,
35 0.027434,
36 0.007550,
37 -0.002861,
38 0.023483,
39 -0.001944,
40 -0.018138,
41 0.014122,
42 -0.017583,
43 -0.040768,
44 0.002598,
45 -0.036604,
46 -0.069541,
47 -0.004273,
48 -0.054876,
49 -0.107289,
50 0.010068,
51 -0.068052,
52 -0.200119,
53 0.207287,
54 0.597150,
55 0.207287,
56 -0.200119,
57 -0.068052,
58 0.010068,
59 -0.107289,
60 -0.054876,
61 -0.004273,
62 -0.069541,
63 -0.036604,
64 0.002598,
65 -0.040768,
66 -0.017583,
67 0.014122,
68 -0.018138,
69 -0.001944,
70 0.023483,
71 -0.002861,
72 0.007550,
73 0.027434,
74 0.004452,
75 0.010147,
76 0.025418,
77 0.004825,
78 0.007084,
79 0.018957,
80 0.000708,
81 0.000941,
82 0.010724,
83 -0.004828,
84 -0.005335,
85 0.003469,
86 -0.009072,
87 -0.009413,
88 -0.000899,
89 -0.010428,
90 -0.010210,
91 -0.001802,
92 -0.008711,
93 -0.008017,
94 0.000043,
95 -0.004923,
96 -0.004144,
97 0.003113,
98 -0.000674,
99 -0.000264,
100 0.005758,
101 0.002523,
102 0.002261,
103 0.006827,
104 0.003795
105}; \ No newline at end of file
diff --git a/codebook.c b/codebook.c
index 0b491ee..50602b7 100644
--- a/codebook.c
+++ b/codebook.c
@@ -1,14 +1,18 @@
1/* THIS IS A GENERATED FILE. Edit generate_codebook.c and its input */ 1/* THIS IS A GENERATED FILE. Edit generate_codebook.c and its input */
2 2
3/* 3/*
4 * This intermediary file and the files that used to create it are under 4 * This intermediary file and the files that used to create it are under
5 * The LGPL. See the file COPYING. 5 * The LGPL. See the file COPYING.
6 */ 6 */
7 7
8#include "defines.h" 8#include "defines.h"
9 9
10 /* codebook/lsp1.txt */ 10 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp1.txt */
11#ifdef __EMBEDDED__
11static const float codes0[] = { 12static const float codes0[] = {
13#else
14static float codes0[] = {
15#endif
12 225, 16 225,
13 250, 17 250,
14 275, 18 275,
@@ -26,8 +30,12 @@ static const float codes0[] = {
26 575, 30 575,
27 600 31 600
28}; 32};
29 /* codebook/lsp2.txt */ 33 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp2.txt */
34#ifdef __EMBEDDED__
30static const float codes1[] = { 35static const float codes1[] = {
36#else
37static float codes1[] = {
38#endif
31 325, 39 325,
32 350, 40 350,
33 375, 41 375,
@@ -45,8 +53,12 @@ static const float codes1[] = {
45 675, 53 675,
46 700 54 700
47}; 55};
48 /* codebook/lsp3.txt */ 56 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp3.txt */
57#ifdef __EMBEDDED__
49static const float codes2[] = { 58static const float codes2[] = {
59#else
60static float codes2[] = {
61#endif
50 500, 62 500,
51 550, 63 550,
52 600, 64 600,
@@ -64,8 +76,12 @@ static const float codes2[] = {
64 1200, 76 1200,
65 1250 77 1250
66}; 78};
67 /* codebook/lsp4.txt */ 79 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp4.txt */
80#ifdef __EMBEDDED__
68static const float codes3[] = { 81static const float codes3[] = {
82#else
83static float codes3[] = {
84#endif
69 700, 85 700,
70 800, 86 800,
71 900, 87 900,
@@ -83,8 +99,12 @@ static const float codes3[] = {
83 2100, 99 2100,
84 2200 100 2200
85}; 101};
86 /* codebook/lsp5.txt */ 102 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp5.txt */
103#ifdef __EMBEDDED__
87static const float codes4[] = { 104static const float codes4[] = {
105#else
106static float codes4[] = {
107#endif
88 950, 108 950,
89 1050, 109 1050,
90 1150, 110 1150,
@@ -102,8 +122,12 @@ static const float codes4[] = {
102 2350, 122 2350,
103 2450 123 2450
104}; 124};
105 /* codebook/lsp6.txt */ 125 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp6.txt */
126#ifdef __EMBEDDED__
106static const float codes5[] = { 127static const float codes5[] = {
128#else
129static float codes5[] = {
130#endif
107 1100, 131 1100,
108 1200, 132 1200,
109 1300, 133 1300,
@@ -121,8 +145,12 @@ static const float codes5[] = {
121 2500, 145 2500,
122 2600 146 2600
123}; 147};
124 /* codebook/lsp7.txt */ 148 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp7.txt */
149#ifdef __EMBEDDED__
125static const float codes6[] = { 150static const float codes6[] = {
151#else
152static float codes6[] = {
153#endif
126 1500, 154 1500,
127 1600, 155 1600,
128 1700, 156 1700,
@@ -140,8 +168,12 @@ static const float codes6[] = {
140 2900, 168 2900,
141 3000 169 3000
142}; 170};
143 /* codebook/lsp8.txt */ 171 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp8.txt */
172#ifdef __EMBEDDED__
144static const float codes7[] = { 173static const float codes7[] = {
174#else
175static float codes7[] = {
176#endif
145 2300, 177 2300,
146 2400, 178 2400,
147 2500, 179 2500,
@@ -151,8 +183,12 @@ static const float codes7[] = {
151 2900, 183 2900,
152 3000 184 3000
153}; 185};
154 /* codebook/lsp9.txt */ 186 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp9.txt */
187#ifdef __EMBEDDED__
155static const float codes8[] = { 188static const float codes8[] = {
189#else
190static float codes8[] = {
191#endif
156 2500, 192 2500,
157 2600, 193 2600,
158 2700, 194 2700,
@@ -162,8 +198,12 @@ static const float codes8[] = {
162 3100, 198 3100,
163 3200 199 3200
164}; 200};
165 /* codebook/lsp10.txt */ 201 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp10.txt */
202#ifdef __EMBEDDED__
166static const float codes9[] = { 203static const float codes9[] = {
204#else
205static float codes9[] = {
206#endif
167 2900, 207 2900,
168 3100, 208 3100,
169 3300, 209 3300,
@@ -171,70 +211,70 @@ static const float codes9[] = {
171}; 211};
172 212
173const struct lsp_codebook lsp_cb[] = { 213const struct lsp_codebook lsp_cb[] = {
174 /* codebook/lsp1.txt */ 214 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp1.txt */
175 { 215 {
176 1, 216 1,
177 4, 217 4,
178 16, 218 16,
179 codes0 219 codes0
180 }, 220 },
181 /* codebook/lsp2.txt */ 221 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp2.txt */
182 { 222 {
183 1, 223 1,
184 4, 224 4,
185 16, 225 16,
186 codes1 226 codes1
187 }, 227 },
188 /* codebook/lsp3.txt */ 228 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp3.txt */
189 { 229 {
190 1, 230 1,
191 4, 231 4,
192 16, 232 16,
193 codes2 233 codes2
194 }, 234 },
195 /* codebook/lsp4.txt */ 235 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp4.txt */
196 { 236 {
197 1, 237 1,
198 4, 238 4,
199 16, 239 16,
200 codes3 240 codes3
201 }, 241 },
202 /* codebook/lsp5.txt */ 242 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp5.txt */
203 { 243 {
204 1, 244 1,
205 4, 245 4,
206 16, 246 16,
207 codes4 247 codes4
208 }, 248 },
209 /* codebook/lsp6.txt */ 249 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp6.txt */
210 { 250 {
211 1, 251 1,
212 4, 252 4,
213 16, 253 16,
214 codes5 254 codes5
215 }, 255 },
216 /* codebook/lsp7.txt */ 256 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp7.txt */
217 { 257 {
218 1, 258 1,
219 4, 259 4,
220 16, 260 16,
221 codes6 261 codes6
222 }, 262 },
223 /* codebook/lsp8.txt */ 263 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp8.txt */
224 { 264 {
225 1, 265 1,
226 3, 266 3,
227 8, 267 8,
228 codes7 268 codes7
229 }, 269 },
230 /* codebook/lsp9.txt */ 270 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp9.txt */
231 { 271 {
232 1, 272 1,
233 3, 273 3,
234 8, 274 8,
235 codes8 275 codes8
236 }, 276 },
237 /* codebook/lsp10.txt */ 277 /* /Users/erdgeist/Coding/codec2/src/codebook/lsp10.txt */
238 { 278 {
239 1, 279 1,
240 2, 280 2,
diff --git a/codebookd.c b/codebookd.c
index 5816402..5ca8651 100644
--- a/codebookd.c
+++ b/codebookd.c
@@ -1,14 +1,18 @@
1/* THIS IS A GENERATED FILE. Edit generate_codebook.c and its input */ 1/* THIS IS A GENERATED FILE. Edit generate_codebook.c and its input */
2 2
3/* 3/*
4 * This intermediary file and the files that used to create it are under 4 * This intermediary file and the files that used to create it are under
5 * The LGPL. See the file COPYING. 5 * The LGPL. See the file COPYING.
6 */ 6 */
7 7
8#include "defines.h" 8#include "defines.h"
9 9
10 /* codebook/dlsp1.txt */ 10 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp1.txt */
11#ifdef __EMBEDDED__
11static const float codes0[] = { 12static const float codes0[] = {
13#else
14static float codes0[] = {
15#endif
12 25, 16 25,
13 50, 17 50,
14 75, 18 75,
@@ -42,8 +46,12 @@ static const float codes0[] = {
42 775, 46 775,
43 800 47 800
44}; 48};
45 /* codebook/dlsp2.txt */ 49 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp2.txt */
50#ifdef __EMBEDDED__
46static const float codes1[] = { 51static const float codes1[] = {
52#else
53static float codes1[] = {
54#endif
47 25, 55 25,
48 50, 56 50,
49 75, 57 75,
@@ -77,8 +85,12 @@ static const float codes1[] = {
77 775, 85 775,
78 800 86 800
79}; 87};
80 /* codebook/dlsp3.txt */ 88 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp3.txt */
89#ifdef __EMBEDDED__
81static const float codes2[] = { 90static const float codes2[] = {
91#else
92static float codes2[] = {
93#endif
82 25, 94 25,
83 50, 95 50,
84 75, 96 75,
@@ -112,8 +124,12 @@ static const float codes2[] = {
112 775, 124 775,
113 800 125 800
114}; 126};
115 /* codebook/dlsp4.txt */ 127 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp4.txt */
128#ifdef __EMBEDDED__
116static const float codes3[] = { 129static const float codes3[] = {
130#else
131static float codes3[] = {
132#endif
117 25, 133 25,
118 50, 134 50,
119 75, 135 75,
@@ -147,8 +163,12 @@ static const float codes3[] = {
147 1350, 163 1350,
148 1400 164 1400
149}; 165};
150 /* codebook/dlsp5.txt */ 166 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp5.txt */
167#ifdef __EMBEDDED__
151static const float codes4[] = { 168static const float codes4[] = {
169#else
170static float codes4[] = {
171#endif
152 25, 172 25,
153 50, 173 50,
154 75, 174 75,
@@ -182,8 +202,12 @@ static const float codes4[] = {
182 1350, 202 1350,
183 1400 203 1400
184}; 204};
185 /* codebook/dlsp6.txt */ 205 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp6.txt */
206#ifdef __EMBEDDED__
186static const float codes5[] = { 207static const float codes5[] = {
208#else
209static float codes5[] = {
210#endif
187 25, 211 25,
188 50, 212 50,
189 75, 213 75,
@@ -217,8 +241,12 @@ static const float codes5[] = {
217 1350, 241 1350,
218 1400 242 1400
219}; 243};
220 /* codebook/dlsp7.txt */ 244 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp7.txt */
245#ifdef __EMBEDDED__
221static const float codes6[] = { 246static const float codes6[] = {
247#else
248static float codes6[] = {
249#endif
222 25, 250 25,
223 50, 251 50,
224 75, 252 75,
@@ -252,8 +280,12 @@ static const float codes6[] = {
252 775, 280 775,
253 800 281 800
254}; 282};
255 /* codebook/dlsp8.txt */ 283 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp8.txt */
284#ifdef __EMBEDDED__
256static const float codes7[] = { 285static const float codes7[] = {
286#else
287static float codes7[] = {
288#endif
257 25, 289 25,
258 50, 290 50,
259 75, 291 75,
@@ -287,8 +319,12 @@ static const float codes7[] = {
287 775, 319 775,
288 800 320 800
289}; 321};
290 /* codebook/dlsp9.txt */ 322 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp9.txt */
323#ifdef __EMBEDDED__
291static const float codes8[] = { 324static const float codes8[] = {
325#else
326static float codes8[] = {
327#endif
292 25, 328 25,
293 50, 329 50,
294 75, 330 75,
@@ -322,8 +358,12 @@ static const float codes8[] = {
322 775, 358 775,
323 800 359 800
324}; 360};
325 /* codebook/dlsp10.txt */ 361 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp10.txt */
362#ifdef __EMBEDDED__
326static const float codes9[] = { 363static const float codes9[] = {
364#else
365static float codes9[] = {
366#endif
327 25, 367 25,
328 50, 368 50,
329 75, 369 75,
@@ -359,70 +399,70 @@ static const float codes9[] = {
359}; 399};
360 400
361const struct lsp_codebook lsp_cbd[] = { 401const struct lsp_codebook lsp_cbd[] = {
362 /* codebook/dlsp1.txt */ 402 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp1.txt */
363 { 403 {
364 1, 404 1,
365 5, 405 5,
366 32, 406 32,
367 codes0 407 codes0
368 }, 408 },
369 /* codebook/dlsp2.txt */ 409 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp2.txt */
370 { 410 {
371 1, 411 1,
372 5, 412 5,
373 32, 413 32,
374 codes1 414 codes1
375 }, 415 },
376 /* codebook/dlsp3.txt */ 416 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp3.txt */
377 { 417 {
378 1, 418 1,
379 5, 419 5,
380 32, 420 32,
381 codes2 421 codes2
382 }, 422 },
383 /* codebook/dlsp4.txt */ 423 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp4.txt */
384 { 424 {
385 1, 425 1,
386 5, 426 5,
387 32, 427 32,
388 codes3 428 codes3
389 }, 429 },
390 /* codebook/dlsp5.txt */ 430 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp5.txt */
391 { 431 {
392 1, 432 1,
393 5, 433 5,
394 32, 434 32,
395 codes4 435 codes4
396 }, 436 },
397 /* codebook/dlsp6.txt */ 437 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp6.txt */
398 { 438 {
399 1, 439 1,
400 5, 440 5,
401 32, 441 32,
402 codes5 442 codes5
403 }, 443 },
404 /* codebook/dlsp7.txt */ 444 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp7.txt */
405 { 445 {
406 1, 446 1,
407 5, 447 5,
408 32, 448 32,
409 codes6 449 codes6
410 }, 450 },
411 /* codebook/dlsp8.txt */ 451 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp8.txt */
412 { 452 {
413 1, 453 1,
414 5, 454 5,
415 32, 455 32,
416 codes7 456 codes7
417 }, 457 },
418 /* codebook/dlsp9.txt */ 458 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp9.txt */
419 { 459 {
420 1, 460 1,
421 5, 461 5,
422 32, 462 32,
423 codes8 463 codes8
424 }, 464 },
425 /* codebook/dlsp10.txt */ 465 /* /Users/erdgeist/Coding/codec2/src/codebook/dlsp10.txt */
426 { 466 {
427 1, 467 1,
428 5, 468 5,
diff --git a/codebookge.c b/codebookge.c
index 5bcb623..8bb0329 100644
--- a/codebookge.c
+++ b/codebookge.c
@@ -1,14 +1,18 @@
1/* THIS IS A GENERATED FILE. Edit generate_codebook.c and its input */ 1/* THIS IS A GENERATED FILE. Edit generate_codebook.c and its input */
2 2
3/* 3/*
4 * This intermediary file and the files that used to create it are under 4 * This intermediary file and the files that used to create it are under
5 * The LGPL. See the file COPYING. 5 * The LGPL. See the file COPYING.
6 */ 6 */
7 7
8#include "defines.h" 8#include "defines.h"
9 9
10 /* codebook/gecb.txt */ 10 /* /Users/erdgeist/Coding/codec2/src/codebook/gecb.txt */
11#ifdef __EMBEDDED__
11static const float codes0[] = { 12static const float codes0[] = {
13#else
14static float codes0[] = {
15#endif
12 2.71, 12.0184, 16 2.71, 12.0184,
13 0.04675, -2.73881, 17 0.04675, -2.73881,
14 0.120993, 8.38895, 18 0.120993, 8.38895,
@@ -268,7 +272,7 @@ static const float codes0[] = {
268}; 272};
269 273
270const struct lsp_codebook ge_cb[] = { 274const struct lsp_codebook ge_cb[] = {
271 /* codebook/gecb.txt */ 275 /* /Users/erdgeist/Coding/codec2/src/codebook/gecb.txt */
272 { 276 {
273 2, 277 2,
274 8, 278 8,
diff --git a/codebookjvm.c b/codebookjmv.c
index 86f5b83..75af664 100644
--- a/codebookjvm.c
+++ b/codebookjmv.c
@@ -1,14 +1,18 @@
1/* THIS IS A GENERATED FILE. Edit generate_codebook.c and its input */ 1/* THIS IS A GENERATED FILE. Edit generate_codebook.c and its input */
2 2
3/* 3/*
4 * This intermediary file and the files that used to create it are under 4 * This intermediary file and the files that used to create it are under
5 * The LGPL. See the file COPYING. 5 * The LGPL. See the file COPYING.
6 */ 6 */
7 7
8#include "defines.h" 8#include "defines.h"
9 9
10 /* codebook/lspjvm1.txt */ 10 /* /Users/erdgeist/Coding/codec2/src/codebook/lspjmv1.txt */
11#ifdef __EMBEDDED__
11static const float codes0[] = { 12static const float codes0[] = {
13#else
14static float codes0[] = {
15#endif
12 0.435217, 0.668864, 1.0103, 1.22042, 1.50398, 1.78468, 2.13546, 2.35747, 2.61891, 2.73804, 16 0.435217, 0.668864, 1.0103, 1.22042, 1.50398, 1.78468, 2.13546, 2.35747, 2.61891, 2.73804,
13 0.179285, 0.33316, 0.500638, 0.79695, 1.03999, 1.23497, 1.6523, 1.84823, 2.62556, 2.80497, 17 0.179285, 0.33316, 0.500638, 0.79695, 1.03999, 1.23497, 1.6523, 1.84823, 2.62556, 2.80497,
14 0.268785, 0.356576, 0.595753, 1.04434, 1.24938, 1.42868, 1.68699, 1.86469, 2.33991, 2.5138, 18 0.268785, 0.356576, 0.595753, 1.04434, 1.24938, 1.42868, 1.68699, 1.86469, 2.33991, 2.5138,
@@ -522,8 +526,12 @@ static const float codes0[] = {
522 0.427795, 0.519003, 0.771284, 0.93724, 1.08662, 1.60988, 1.87875, 2.05279, 2.53412, 2.65715, 526 0.427795, 0.519003, 0.771284, 0.93724, 1.08662, 1.60988, 1.87875, 2.05279, 2.53412, 2.65715,
523 0.22437, 0.317969, 0.439666, 0.812931, 1.3985, 1.62663, 1.79418, 2.114, 2.30916, 2.49684 527 0.22437, 0.317969, 0.439666, 0.812931, 1.3985, 1.62663, 1.79418, 2.114, 2.30916, 2.49684
524}; 528};
525 /* codebook/lspjvm2.txt */ 529 /* /Users/erdgeist/Coding/codec2/src/codebook/lspjmv2.txt */
530#ifdef __EMBEDDED__
526static const float codes1[] = { 531static const float codes1[] = {
532#else
533static float codes1[] = {
534#endif
527 0.005167, -0.03731, -0.002159, 0.016849, 0.130396, 535 0.005167, -0.03731, -0.002159, 0.016849, 0.130396,
528 0.039445, 0.03168, -0.074412, -0.031499, 0.060536, 536 0.039445, 0.03168, -0.074412, -0.031499, 0.060536,
529 0.019479, -0.030564, -0.048137, -0.056279, -0.027829, 537 0.019479, -0.030564, -0.048137, -0.056279, -0.027829,
@@ -1037,8 +1045,12 @@ static const float codes1[] = {
1037 -0.033256, -0.053774, 0.049001, -0.002339, 0.013545, 1045 -0.033256, -0.053774, 0.049001, -0.002339, 0.013545,
1038 -0.006432, -0.012089, -0.086842, 0.104105, 0.061991 1046 -0.006432, -0.012089, -0.086842, 0.104105, 0.061991
1039}; 1047};
1040 /* codebook/lspjvm3.txt */ 1048 /* /Users/erdgeist/Coding/codec2/src/codebook/lspjmv3.txt */
1049#ifdef __EMBEDDED__
1041static const float codes2[] = { 1050static const float codes2[] = {
1051#else
1052static float codes2[] = {
1053#endif
1042 0.007066, 0.075781, -0.070082, -0.092014, -0.066477, 1054 0.007066, 0.075781, -0.070082, -0.092014, -0.066477,
1043 0.09051, 0.106622, 0.025911, -0.01676, 0.003724, 1055 0.09051, 0.106622, 0.025911, -0.01676, 0.003724,
1044 -0.024628, 0.058332, 0.012876, 0.059557, -0.002092, 1056 -0.024628, 0.058332, 0.012876, 0.059557, -0.002092,
@@ -1553,22 +1565,22 @@ static const float codes2[] = {
1553 0.086151, -0.113571, -0.019466, -0.009167, 0.003662 1565 0.086151, -0.113571, -0.019466, -0.009167, 0.003662
1554}; 1566};
1555 1567
1556const struct lsp_codebook lsp_cbjvm[] = { 1568const struct lsp_codebook lsp_cbjmv[] = {
1557 /* codebook/lspjvm1.txt */ 1569 /* /Users/erdgeist/Coding/codec2/src/codebook/lspjmv1.txt */
1558 { 1570 {
1559 10, 1571 10,
1560 9, 1572 9,
1561 512, 1573 512,
1562 codes0 1574 codes0
1563 }, 1575 },
1564 /* codebook/lspjvm2.txt */ 1576 /* /Users/erdgeist/Coding/codec2/src/codebook/lspjmv2.txt */
1565 { 1577 {
1566 5, 1578 5,
1567 9, 1579 9,
1568 512, 1580 512,
1569 codes1 1581 codes1
1570 }, 1582 },
1571 /* codebook/lspjvm3.txt */ 1583 /* /Users/erdgeist/Coding/codec2/src/codebook/lspjmv3.txt */
1572 { 1584 {
1573 5, 1585 5,
1574 9, 1586 9,
diff --git a/codebooklspmelvq.c b/codebooklspmelvq.c
deleted file mode 100644
index f49ff36..0000000
--- a/codebooklspmelvq.c
+++ /dev/null
@@ -1,247 +0,0 @@
1/* THIS IS A GENERATED FILE. Edit generate_codebook.c and its input */
2
3/*
4 * This intermediary file and the files that used to create it are under
5 * The LGPL. See the file COPYING.
6 */
7
8#include "defines.h"
9
10 /* /Users/erdgeist/Coding/codec2/src/codebook/lspmelvq1.txt */
11#ifdef __EMBEDDED__
12static const float codes0[] = {
13#else
14static float codes0[] = {
15#endif
16 640.185, 893.139, 1393.85, 1494.06, 1656.26, 1700.82,
17 558.516, 716.906, 1057.36, 1182.62, 1503.8, 1610.03,
18 615.108, 769.622, 1120.16, 1222.84, 1366.46, 1465.05,
19 663.458, 736.833, 964.764, 1043.54, 1623.17, 1681.29,
20 487.957, 577.723, 1017.43, 1177.04, 1552.1, 1615.69,
21 536.099, 733.407, 1292.28, 1406.09, 1577.7, 1637.49,
22 473.015, 542.559, 877.397, 1285.82, 1591.04, 1647.44,
23 525.343, 652.014, 1206.83, 1493.96, 1647.97, 1698.3,
24 510.887, 572.868, 945.226, 1445.68, 1678.17, 1705.53,
25 534.915, 721.265, 1275.92, 1415.76, 1648.5, 1695.73,
26 865.189, 1047, 1267.14, 1389.32, 1646.57, 1696.97,
27 608.033, 869.887, 1300.95, 1432.87, 1639.74, 1689.41,
28 554.972, 649.352, 866.845, 979.873, 1645.31, 1695.39,
29 696.079, 813.97, 1102.49, 1219.79, 1536.49, 1621.94,
30 553.879, 691.097, 1200.84, 1339.34, 1629.08, 1683.5,
31 778.561, 997.776, 1258.63, 1390.34, 1601.99, 1657.86,
32 713.107, 778.893, 992.875, 1051.95, 1497.45, 1650.66,
33 490.27, 598.18, 1116.02, 1244.13, 1622.26, 1672.21,
34 448.556, 512.085, 1271.7, 1448.18, 1579.37, 1642.48,
35 465.688, 535.312, 1099.19, 1535.79, 1684.29, 1710.9,
36 812.222, 1087.53, 1470.44, 1559.73, 1692.18, 1726.38,
37 428.174, 489.426, 1160.33, 1409.3, 1597.5, 1651.3,
38 712.81, 957.56, 1433.02, 1516.37, 1675.39, 1710.06,
39 717.255, 934.073, 1305.2, 1436.72, 1647.2, 1693.09,
40 492.888, 580.393, 1339.52, 1461.07, 1592, 1653.42,
41 550.467, 675.888, 990.888, 1177.44, 1615.64, 1658,
42 714.528, 801.792, 1072.94, 1146.08, 1637.06, 1706.58,
43 513.819, 590.989, 798.691, 895.755, 1557.76, 1624.56,
44 436.653, 518.413, 1281.99, 1547.08, 1685.12, 1717.76,
45 681.854, 758.354, 1046.65, 1120.92, 1412.25, 1603.42,
46 873.962, 1118.49, 1376.61, 1465.07, 1665.38, 1707.18,
47 553.529, 634.092, 1144.77, 1284.8, 1542.87, 1620.63,
48 448.532, 519.097, 1054.57, 1319.66, 1591.26, 1649.85,
49 742.267, 885.293, 1152.83, 1318.51, 1569.36, 1631.45,
50 529.03, 654.522, 1355.76, 1511.75, 1662.39, 1706.4,
51 463.794, 597.77, 1176.05, 1366.13, 1629.37, 1678.01,
52 626.936, 706.66, 1058.04, 1323.62, 1473.28, 1599.68,
53 477.322, 615.5, 1488.89, 1550.5, 1683.1, 1712.34,
54 547.442, 815.442, 1313.38, 1486.96, 1671.97, 1717.4,
55 610.671, 819.955, 1219.11, 1363.66, 1592.05, 1654.31,
56 547.414, 746.54, 1438.43, 1517.72, 1659.64, 1695.57,
57 604.823, 821.146, 1137.94, 1358.29, 1598.94, 1655.64,
58 525.935, 616.739, 1060.13, 1427.33, 1593.35, 1657.48,
59 622.5, 762.143, 1318.65, 1410.96, 1618.12, 1680.06,
60 436.917, 516.583, 1390.29, 1475.86, 1594.71, 1633.74,
61 792.487, 1031.24, 1362.62, 1472.68, 1649.26, 1697.35,
62 457.707, 526.207, 865.966, 1120.47, 1564.83, 1625.28,
63 526.39, 624.21, 1269.65, 1374.23, 1558.07, 1620.96,
64 483.768, 573.505, 1440.66, 1512.43, 1622.37, 1671.31,
65 953.061, 1194.03, 1416.67, 1515.82, 1678.82, 1718.21,
66 499.947, 627.358, 1299.94, 1394.23, 1643.17, 1685.33,
67 648.723, 838.181, 1225.5, 1383.45, 1637.46, 1691.67,
68 672.588, 1022.6, 1346.21, 1443.75, 1651.15, 1695.43,
69 581.833, 674.944, 955.167, 1020.5, 1370.5, 1503.11,
70 536.143, 652.531, 1243.84, 1315.27, 1425.1, 1505.73,
71 560.558, 786.65, 1224.66, 1373.98, 1630.06, 1682.68,
72 591.926, 783.722, 982, 1140.81, 1581.61, 1625.28,
73 548.537, 644.524, 940.451, 1048.74, 1557.89, 1609.48,
74 516.916, 723.253, 1135.95, 1309.19, 1600.08, 1669.48,
75 426.521, 506.077, 1457.73, 1535.02, 1641.35, 1678.45,
76 645.189, 776.595, 1020.78, 1240.07, 1597, 1648.7,
77 637.105, 941.474, 1242.21, 1372.49, 1646.7, 1694.8,
78 691.228, 788.141, 1202.12, 1294.89, 1626.97, 1681.77,
79 699.08, 886.655, 1300.9, 1399.28, 1579.94, 1646.79
80};
81 /* /Users/erdgeist/Coding/codec2/src/codebook/lspmelvq2.txt */
82#ifdef __EMBEDDED__
83static const float codes1[] = {
84#else
85static float codes1[] = {
86#endif
87 4.44342, 51.1708, 7.45726, -27.4373, -18.5056, -18.1989,
88 51.58, 63.3166, 57.7796, -44.1591, -18.4834, -6.93392,
89 -20.2795, -21.7454, 4.66947, 52.1569, 30.4367, 36.8582,
90 -29.1104, -5.63933, -3.45383, -63.0261, -20.4423, -19.0485,
91 2.91622, 40.8374, 16.579, -51.8461, 38.5045, 18.1728,
92 -20.6977, -11.4022, -36.6173, -16.6116, -56.8965, -24.301,
93 -20.2385, 26.6332, 33.1191, 27.6284, -36.7493, -25.6041,
94 54.9871, 0.71748, 23.0674, -22.1031, 11.6643, 10.9938,
95 -62.7215, 21.7547, 21.2907, -7.64891, -4.5533, -9.71777,
96 -9.56338, -3.85841, 25.0454, -9.45216, 6.05017, 5.35043,
97 47.3823, 56.6122, -27.0315, -24.67, 4.86343, -0.225495,
98 26.3997, 26.7857, -1.66167, 62.8366, -19.7653, -8.55169,
99 -8.77648, -9.04545, -7.88996, 28.2433, -35.0963, -21.9709,
100 14.8423, 25.4563, -56.1262, -50.2934, -22.9393, -15.8113,
101 4.94186, 27.7798, 8.34579, 10.8553, -3.12587, -3.97807,
102 12.6426, -52.2317, 37.2487, -57.2067, -14.5125, -5.54035,
103 13.485, 15.3246, -23.9644, -21.3135, 19.5779, 14.1597,
104 -55.3543, -45.2077, 10.5185, 43.0461, -24.9859, -19.3484,
105 27.7226, 32.1882, 20.0321, 24.3328, -72.8194, -51.1823,
106 -31.3818, -5.25745, -43.7806, 14.1312, 17.6392, 9.81024,
107 -48.26, -26.2973, -44.1428, -31.9001, 22.5085, -0.467938,
108 7.37202, -7.79071, -12.5732, 27.1074, 9.34052, 14.4477,
109 14.5295, 8.82597, 57.0009, -16.3234, -32.4142, -21.0224,
110 32.4616, 48.6062, 38.5452, 9.77182, 1.82856, 11.5063,
111 -43.8275, -22.6263, -29.8278, 13.6115, 9.66849, -63.5218,
112 -11.9967, 2.74308, -73.6375, -20.9809, -4.11839, 7.71405,
113 24.0162, -2.29513, -6.80983, -26.4043, -21.8529, -16.3381,
114 -16.1484, 35.9086, -3.0837, 3.83958, 42.3003, 17.5003,
115 54.1225, -48.7513, -14.8712, -38.1256, 15.2903, 7.33079,
116 53.0929, 13.9221, 10.6536, 24.345, -16.5952, -16.0365,
117 33.1415, 38.5714, -26.0251, 22.021, 15.6866, 13.2593,
118 15.7194, -49.4061, 31.4552, 10.1896, 0.219911, 1.62902,
119 4.10868, 14.2755, 58.475, -1.16668, 52.6265, 43.4938,
120 358.653, -112.587, 85.9867, 52.08, -52.88, -249.24,
121 -12.5792, -6.64039, -33.0106, 1.51449, 50.3259, 61.6091,
122 -6.81685, 14.3146, 14.1563, 53.6363, 83.6051, 38.174,
123 -18.0006, 41.3575, -46.3736, 8.47794, -10.4611, -11.3847,
124 28.7711, 31.4689, -39.5744, 1.43977, -37.7309, -30.5309,
125 -34.4692, -11.0778, 44.4681, -60.0045, -44.9729, -34.8364,
126 38.7401, 41.0529, -7.92946, 20.0279, 74.4246, 54.6498,
127 81.8008, 133.531, -5.47375, 34.4759, 75.7417, 52.6112,
128 -61.5067, -76.6328, -47.2471, -43.5928, -9.46878, 0.832598,
129 -11.544, -44.7728, -13.9257, -3.32122, 24.816, 21.9064,
130 2.83945, -45.0522, -36.8776, -14.7577, -11.9912, -8.75366,
131 -8.97657, -14.0499, 39.2628, 48.8038, -14.3789, -4.79625,
132 31.6292, 32.0168, 5.52278, -79.8411, -24.1789, -15.7597,
133 15.893, 10.4667, -43.9057, -20.4885, 80.7594, 6.8023,
134 22.3285, -69.4942, -5.16156, 48.8868, 4.01995, -3.04376,
135 31.4062, -20.0952, -55.1259, 0.505988, 20.86, 15.0816,
136 -12.9696, 37.2548, 18.1009, 51.0767, 18.2053, 19.8474,
137 -45.4855, -45.3454, -4.22795, -15.8693, -6.07272, -1.99631,
138 -9.25462, 45.0241, 70.2879, -44.4484, 2.63165, 4.15336,
139 -51.1589, -27.7586, 19.7999, -9.87882, 63.4125, 48.3124,
140 -8.94084, -13.4815, -44.9602, 52.839, -17.2582, -1.59439,
141 23.18, -3.00374, 33.162, 31.5424, 29.111, 25.9327,
142 -28.4665, -10.6868, 88.7543, -3.88024, 17.5881, 13.7336,
143 -10.6578, 7.43021, -20.945, -17.6491, -11.2815, -10.8001,
144 -35.6182, -6.90301, 3.92829, 26.9695, 2.8894, 5.01685,
145 -3.78263, -19.2095, -10.5568, -67.7468, 20.3733, 29.552,
146 53.8109, -33.8374, -27.8223, -7.73289, -31.3462, -23.5678,
147 -38.8762, -17.7679, 11.457, 28.4607, 50.7006, -19.2488,
148 -44.6024, -60.2806, 56.35, 21.7545, 5.23677, 7.66608,
149 -3.80758, -39.1425, 28.8305, -3.09285, -42.7534, -25.1803,
150 45.2559, 28.8742, 60.0723, 64.461, 3.51203, -27.883
151};
152 /* /Users/erdgeist/Coding/codec2/src/codebook/lspmelvq3.txt */
153#ifdef __EMBEDDED__
154static const float codes2[] = {
155#else
156static float codes2[] = {
157#endif
158 -9.63558, 27.5501, 15.4445, -4.34872, -1.8587, 1.27054,
159 -15.5343, 23.0515, -15.4436, -2.01887, -6.19433, -13.5085,
160 8.38867, -1.60998, -32.3903, 10.1765, 1.9467, 12.2454,
161 -3.06306, -9.55983, 14.367, -1.87159, 19.6192, -3.78366,
162 -18.495, -13.3811, -23.8928, -20.3745, -6.87856, -17.4887,
163 15.4925, 14.707, -0.0697855, 15.7541, 12.6051, -3.19768,
164 9.60466, -4.56494, 10.1616, 15.5594, 6.89224, -31.5602,
165 -11.5625, -23.0872, 34.9163, 12.3052, 7.67426, -1.26298,
166 2.5486, -3.90251, -19.1132, 7.6131, -31.0016, 12.4759,
167 2.74156, 12.4124, -39.5057, -0.325024, -22.9186, -28.5606,
168 6.30148, 15.4402, -2.6284, -20.1603, 5.22906, -12.3451,
169 -6.91862, 16.6335, -1.65064, 2.99602, -23.9479, -1.43947,
170 -14.3907, -31.417, 10.1113, 1.70013, -21.5733, 4.736,
171 -1.67171, 6.22751, -13.7187, 21.0936, -9.69243, -10.5756,
172 2.15266, 21.2198, -13.0171, -1.43135, 18.8831, 10.6664,
173 13.8913, 27.3565, 0.472838, -7.40477, -14.8705, 25.7448,
174 28.402, -2.05484, -9.32712, -17.3169, 15.643, 6.96908,
175 -15.863, -17.8482, -24.9238, 12.5574, 7.17566, 0.0161972,
176 5.99291, -41.0228, 1.95791, -6.78012, 9.20162, 4.6234,
177 -6.33629, -7.61679, 27.7318, -8.9214, -14.1931, 7.88247,
178 -12.2367, -21.245, -2.5927, 13.7776, -2.7864, -24.5072,
179 23.15, -9.93687, -2.92559, 3.88086, 11.2667, 11.1998,
180 9.67437, -9.4269, 6.2582, 36.5694, 2.88654, 1.82052,
181 15.385, -2.88243, 19.8377, -14.5111, 5.92264, -2.55757,
182 -20.9648, 3.76147, 18.5074, -13.5547, -7.84261, -19.98,
183 24.6032, 16.989, -19.1622, 1.35535, -0.0122027, 0.166227,
184 9.98886, -6.89666, -20.5111, -2.89196, 2.58467, -17.049,
185 -9.17761, -23.7209, 12.6088, -18.2654, 2.17718, -13.0865,
186 -9.73326, -12.2682, 6.80914, 20.3469, -10.2912, 4.85191,
187 -5.19406, 6.78014, 18.0099, 14.3782, -0.124328, -10.0141,
188 -4.69806, 6.71393, -19.1371, 8.19814, 23.3987, -10.3316,
189 1.02965, -4.63654, 21.9822, 11.5088, -30.9617, -20.6354,
190 -33.2824, 31.4666, -11.4837, 11.5144, -1.36834, 9.44599,
191 -37.5431, -3.16317, -2.09497, -2.62712, 40.4714, -33.0527,
192 -8.79595, -15.5174, -15.7916, 7.97003, 37.2542, 40.7063,
193 -14.7261, -12.6884, 2.42105, -10.2686, 25.9033, 14.8525,
194 22.9598, -16.6224, -3.64949, 4.44269, -22.3897, 13.6968,
195 -10.874, 4.18931, -24.2284, -3.63764, -15.1379, 40.9515,
196 28.2393, -8.63225, -12.544, 28.8282, -0.987894, -4.9824,
197 -25.0777, -0.481678, -3.37082, 5.55114, -9.89898, -8.07628,
198 23.3581, 3.12034, -8.63348, 0.63042, -18.2216, -22.1886,
199 44.9505, 19.8267, 23.7129, 8.58075, 7.80458, 1.78796,
200 -8.13112, -2.1262, -7.12776, -25.0529, -16.7287, 8.41402,
201 2.01965, 19.2579, 20.0963, 5.99199, 28.1098, 5.96128,
202 2.42493, -6.33216, -26.5858, -23.8607, 8.27049, 3.05805,
203 0.0153248, 0.446112, -4.92759, 19.0023, 22.7346, 15.5451,
204 -7.39591, 40.285, 10.8414, 25.7961, -8.81069, -13.5,
205 5.81306, 11.2384, 6.93765, -9.43067, 9.51418, 22.9709,
206 -9.17611, -16.4993, -1.56929, -2.9111, 4.17113, 10.9228,
207 10.0376, -27.4993, -8.25332, -1.715, -11.5063, -10.467,
208 23.6637, -13.8338, 14.7284, 8.00341, -2.71881, -8.80708,
209 27.101, 2.42801, 11.4599, -24.1577, -20.9901, 4.52358,
210 16.8065, 19.3315, 11.1219, 13.3391, -13.1522, 0.91428,
211 -25.2603, 6.04837, 12.1994, 21.9372, 14.8795, 6.93368,
212 -1.24639, -7.96856, 16.4064, -2.36409, -25.9093, 46.0938,
213 8.81687, 24.8004, 11.4475, -13.261, -19.8693, -28.4793,
214 15.3175, -12.5335, 24.1778, 10.5133, 22.2244, 9.93191,
215 -18.7982, 38.939, -20.0631, -35.4052, 7.5879, -0.599373,
216 -18.1388, 9.5843, 17.4375, -21.057, 8.1634, 6.31216,
217 -61.5486, -8.71159, 19.7502, -25.2365, 3.56558, -1.64973,
218 -26.9863, 0.291017, -12.5337, -10.505, 11.0614, 4.84072,
219 -0.630579, -3.87056, -0.707795, -13.1306, -19.6548, -16.2436,
220 4.87022, -5.90744, 3.46971, -40.0866, 16.8741, 10.3333,
221 -1.00985, 16.234, -0.475836, 28.3848, 1.75473, 28.2608
222};
223
224const struct lsp_codebook lspmelvq_cb[] = {
225 /* /Users/erdgeist/Coding/codec2/src/codebook/lspmelvq1.txt */
226 {
227 6,
228 6,
229 64,
230 codes0
231 },
232 /* /Users/erdgeist/Coding/codec2/src/codebook/lspmelvq2.txt */
233 {
234 6,
235 6,
236 64,
237 codes1
238 },
239 /* /Users/erdgeist/Coding/codec2/src/codebook/lspmelvq3.txt */
240 {
241 6,
242 6,
243 64,
244 codes2
245 },
246 { 0, 0, 0, 0 }
247};
diff --git a/codebookmel.c b/codebookmel.c
deleted file mode 100644
index a548db5..0000000
--- a/codebookmel.c
+++ /dev/null
@@ -1,145 +0,0 @@
1/* THIS IS A GENERATED FILE. Edit generate_codebook.c and its input */
2
3/*
4 * This intermediary file and the files that used to create it are under
5 * The LGPL. See the file COPYING.
6 */
7
8#include "defines.h"
9
10 /* /Users/erdgeist/Coding/codec2/src/codebook/mel1.txt */
11#ifdef __EMBEDDED__
12static const float codes0[] = {
13#else
14static float codes0[] = {
15#endif
16 550,
17 600,
18 650,
19 700,
20 750,
21 800,
22 850,
23 900
24};
25 /* /Users/erdgeist/Coding/codec2/src/codebook/mel2.txt */
26#ifdef __EMBEDDED__
27static const float codes1[] = {
28#else
29static float codes1[] = {
30#endif
31 50,
32 100,
33 200,
34 300
35};
36 /* /Users/erdgeist/Coding/codec2/src/codebook/mel3.txt */
37#ifdef __EMBEDDED__
38static const float codes2[] = {
39#else
40static float codes2[] = {
41#endif
42 800,
43 850,
44 900,
45 950,
46 1000,
47 1050,
48 1100,
49 1150,
50 1200,
51 1250,
52 1300,
53 1350,
54 1400,
55 1450,
56 1500,
57 1650
58};
59 /* /Users/erdgeist/Coding/codec2/src/codebook/mel4.txt */
60#ifdef __EMBEDDED__
61static const float codes3[] = {
62#else
63static float codes3[] = {
64#endif
65 25,
66 50,
67 75,
68 100,
69 125,
70 150,
71 175,
72 250
73};
74 /* /Users/erdgeist/Coding/codec2/src/codebook/mel5.txt */
75#ifdef __EMBEDDED__
76static const float codes4[] = {
77#else
78static float codes4[] = {
79#endif
80 1350,
81 1400,
82 1450,
83 1500,
84 1550,
85 1600,
86 1650,
87 1700
88};
89 /* /Users/erdgeist/Coding/codec2/src/codebook/mel6.txt */
90#ifdef __EMBEDDED__
91static const float codes5[] = {
92#else
93static float codes5[] = {
94#endif
95 25,
96 50,
97 100,
98 150
99};
100
101const struct lsp_codebook mel_cb[] = {
102 /* /Users/erdgeist/Coding/codec2/src/codebook/mel1.txt */
103 {
104 1,
105 3,
106 8,
107 codes0
108 },
109 /* /Users/erdgeist/Coding/codec2/src/codebook/mel2.txt */
110 {
111 1,
112 2,
113 4,
114 codes1
115 },
116 /* /Users/erdgeist/Coding/codec2/src/codebook/mel3.txt */
117 {
118 1,
119 4,
120 16,
121 codes2
122 },
123 /* /Users/erdgeist/Coding/codec2/src/codebook/mel4.txt */
124 {
125 1,
126 3,
127 8,
128 codes3
129 },
130 /* /Users/erdgeist/Coding/codec2/src/codebook/mel5.txt */
131 {
132 1,
133 3,
134 8,
135 codes4
136 },
137 /* /Users/erdgeist/Coding/codec2/src/codebook/mel6.txt */
138 {
139 1,
140 2,
141 4,
142 codes5
143 },
144 { 0, 0, 0, 0 }
145};
diff --git a/codebooknewamp1.c b/codebooknewamp1.c
index 6de33ee..4705acb 100644
--- a/codebooknewamp1.c
+++ b/codebooknewamp1.c
@@ -7,8 +7,12 @@
7 7
8#include "defines.h" 8#include "defines.h"
9 9
10 /* /home/david/codec2-dev/src/codebook/train_120_1.txt */ 10 /* /Users/erdgeist/Coding/codec2/src/codebook/train_120_1.txt */
11#ifdef __EMBEDDED__
11static const float codes0[] = { 12static const float codes0[] = {
13#else
14static float codes0[] = {
15#endif
12 6.7484, 7.6125, 6.0332, 5.2789, 1.5239, 2.2353, 2.0748, 0.5289, 0.8748, 2.5432, -2.2863, -3.191, -0.0434, -1.9857, -3.3605, 0.7069, -5.9493, -0.5672, -0.6798, -18.0977, 16 6.7484, 7.6125, 6.0332, 5.2789, 1.5239, 2.2353, 2.0748, 0.5289, 0.8748, 2.5432, -2.2863, -3.191, -0.0434, -1.9857, -3.3605, 0.7069, -5.9493, -0.5672, -0.6798, -18.0977,
13 4.0503, 3.9086, 2.9225, 2.3773, 0.658, -0.4363, -0.0644, 2.4063, 1.3428, 2.4542, 0.5275, 0.982, -1.3277, 0.6811, 0.0273, -0.1838, -0.0222, -0.6478, -2.2405, -17.4152, 17 4.0503, 3.9086, 2.9225, 2.3773, 0.658, -0.4363, -0.0644, 2.4063, 1.3428, 2.4542, 0.5275, 0.982, -1.3277, 0.6811, 0.0273, -0.1838, -0.0222, -0.6478, -2.2405, -17.4152,
14 13.3284, 12.1212, 10.6531, 9.8214, 11.0388, 15.812, 19.711, 16.5488, 16.1068, 15.8771, 7.2553, 4.2486, -6.0036, -12.5476, -20.1299, -28.2803, -25.3971, -21.7907, -11.5143, -26.859, 18 13.3284, 12.1212, 10.6531, 9.8214, 11.0388, 15.812, 19.711, 16.5488, 16.1068, 15.8771, 7.2553, 4.2486, -6.0036, -12.5476, -20.1299, -28.2803, -25.3971, -21.7907, -11.5143, -26.859,
@@ -522,8 +526,12 @@ static const float codes0[] = {
522 18.9039, 18.0327, 13.6297, 9.7561, 7.4319, 5.3744, 2.9646, 1.7415, 0.6716, 0.6859, 2.7446, 0.5342, -4.3187, -6.1877, -6.9311, -9.8734, -11.9041, -9.1281, -8.9591, -25.169, 526 18.9039, 18.0327, 13.6297, 9.7561, 7.4319, 5.3744, 2.9646, 1.7415, 0.6716, 0.6859, 2.7446, 0.5342, -4.3187, -6.1877, -6.9311, -9.8734, -11.9041, -9.1281, -8.9591, -25.169,
523 6.4513, 6.7843, 9.4438, 12.9549, 16.4801, 14.9093, 8.2954, 1.6877, -2.5193, -4.4922, -6.7348, -6.9853, -4.6783, 0.7181, 1.0483, -5.1128, -1.1161, -7.5305, -12.398, -27.2058 527 6.4513, 6.7843, 9.4438, 12.9549, 16.4801, 14.9093, 8.2954, 1.6877, -2.5193, -4.4922, -6.7348, -6.9853, -4.6783, 0.7181, 1.0483, -5.1128, -1.1161, -7.5305, -12.398, -27.2058
524}; 528};
525 /* /home/david/codec2-dev/src/codebook/train_120_2.txt */ 529 /* /Users/erdgeist/Coding/codec2/src/codebook/train_120_2.txt */
530#ifdef __EMBEDDED__
526static const float codes1[] = { 531static const float codes1[] = {
532#else
533static float codes1[] = {
534#endif
527 -1.2668, -1.2477, -0.0681, 3.8419, -0.0693, -1.7919, -1.5943, 0.8402, 0.155, -3.1526, -3.0204, 0.7337, -0.2603, 1.659, 0.023, 5.5893, -2.4959, 1.9604, -0.6348, 0.7999, 535 -1.2668, -1.2477, -0.0681, 3.8419, -0.0693, -1.7919, -1.5943, 0.8402, 0.155, -3.1526, -3.0204, 0.7337, -0.2603, 1.659, 0.023, 5.5893, -2.4959, 1.9604, -0.6348, 0.7999,
528 2.6673, 1.4923, 1.1408, -0.7478, -1.0755, -1.3421, -0.4884, 0.2535, 0.2951, 1.5088, 1.9447, -2.662, 2.3751, 1.5298, 4.1357, -6.563, -2.1766, -2.7535, -0.7706, 1.2362, 536 2.6673, 1.4923, 1.1408, -0.7478, -1.0755, -1.3421, -0.4884, 0.2535, 0.2951, 1.5088, 1.9447, -2.662, 2.3751, 1.5298, 4.1357, -6.563, -2.1766, -2.7535, -0.7706, 1.2362,
529 -0.6415, -0.4348, -1.2178, -0.987, -1.1057, -2.1421, -2.3594, -0.4977, -4.1484, -4.712, 5.5425, 3.0695, 3.661, 2.9729, 2.4379, -1.6136, -1.3052, 2.1342, 1.9164, -0.5692, 537 -0.6415, -0.4348, -1.2178, -0.987, -1.1057, -2.1421, -2.3594, -0.4977, -4.1484, -4.712, 5.5425, 3.0695, 3.661, 2.9729, 2.4379, -1.6136, -1.3052, 2.1342, 1.9164, -0.5692,
@@ -1039,14 +1047,14 @@ static const float codes1[] = {
1039}; 1047};
1040 1048
1041const struct lsp_codebook newamp1vq_cb[] = { 1049const struct lsp_codebook newamp1vq_cb[] = {
1042 /* /home/david/codec2-dev/src/codebook/train_120_1.txt */ 1050 /* /Users/erdgeist/Coding/codec2/src/codebook/train_120_1.txt */
1043 { 1051 {
1044 20, 1052 20,
1045 9, 1053 9,
1046 512, 1054 512,
1047 codes0 1055 codes0
1048 }, 1056 },
1049 /* /home/david/codec2-dev/src/codebook/train_120_2.txt */ 1057 /* /Users/erdgeist/Coding/codec2/src/codebook/train_120_2.txt */
1050 { 1058 {
1051 20, 1059 20,
1052 9, 1060 9,
diff --git a/codebooknewamp1_energy.c b/codebooknewamp1_energy.c
index 6ecdd7c..7e6a092 100644
--- a/codebooknewamp1_energy.c
+++ b/codebooknewamp1_energy.c
@@ -7,8 +7,12 @@
7 7
8#include "defines.h" 8#include "defines.h"
9 9
10 /* /home/david/codec2-dev/src/codebook/newamp1_energy_q.txt */ 10 /* /Users/erdgeist/Coding/codec2/src/codebook/newamp1_energy_q.txt */
11#ifdef __EMBEDDED__
11static const float codes0[] = { 12static const float codes0[] = {
13#else
14static float codes0[] = {
15#endif
12 10, 16 10,
13 12.5, 17 12.5,
14 15, 18 15,
@@ -28,7 +32,7 @@ static const float codes0[] = {
28}; 32};
29 33
30const struct lsp_codebook newamp1_energy_cb[] = { 34const struct lsp_codebook newamp1_energy_cb[] = {
31 /* /home/david/codec2-dev/src/codebook/newamp1_energy_q.txt */ 35 /* /Users/erdgeist/Coding/codec2/src/codebook/newamp1_energy_q.txt */
32 { 36 {
33 1, 37 1,
34 4, 38 4,
diff --git a/codebooknewamp2.c b/codebooknewamp2.c
index c8389b0..f3e37dd 100644
--- a/codebooknewamp2.c
+++ b/codebooknewamp2.c
@@ -519,7 +519,7 @@ const struct lsp_codebook newamp2vq_cb[] = {
519 /* /Users/erdgeist/Coding/codec2/src/codebook/codes_450.txt */ 519 /* /Users/erdgeist/Coding/codec2/src/codebook/codes_450.txt */
520 { 520 {
521 41, 521 41,
522 8.96578, 522 9,
523 500, 523 500,
524 codes0 524 codes0
525 }, 525 },
diff --git a/codec2.c b/codec2.c
index 2697356..b27626a 100644
--- a/codec2.c
+++ b/codec2.c
@@ -26,33 +26,31 @@
26 along with this program; if not, see <http://www.gnu.org/licenses/>. 26 along with this program; if not, see <http://www.gnu.org/licenses/>.
27*/ 27*/
28 28
29#include "codec2.h"
30
29#include <assert.h> 31#include <assert.h>
32#include <math.h>
33#include <stdbool.h>
30#include <stdio.h> 34#include <stdio.h>
31#include <stdlib.h> 35#include <stdlib.h>
32#include <stdbool.h>
33#include <string.h> 36#include <string.h>
34#include <math.h>
35 37
36#include "defines.h" 38#include "bpf.h"
39#include "bpfb.h"
37#include "codec2_fft.h" 40#include "codec2_fft.h"
38#include "sine.h" 41#include "codec2_internal.h"
39#include "nlp.h" 42#include "debug_alloc.h"
43#include "defines.h"
40#include "dump.h" 44#include "dump.h"
41#include "lpc.h"
42#include "quantise.h"
43#include "phase.h"
44#include "interp.h" 45#include "interp.h"
45#include "postfilter.h" 46#include "lpc.h"
46#include "codec2.h"
47#include "lsp.h" 47#include "lsp.h"
48#include "newamp2.h"
49#include "codec2_internal.h"
50#include "machdep.h" 48#include "machdep.h"
51#include "bpf.h" 49#include "nlp.h"
52#include "bpfb.h" 50#include "phase.h"
53#include "c2wideband.h" 51#include "postfilter.h"
54 52#include "quantise.h"
55#include "debug_alloc.h" 53#include "sine.h"
56 54
57/*---------------------------------------------------------------------------* \ 55/*---------------------------------------------------------------------------* \
58 56
@@ -62,32 +60,30 @@
62 60
63void analyse_one_frame(struct CODEC2 *c2, MODEL *model, short speech[]); 61void analyse_one_frame(struct CODEC2 *c2, MODEL *model, short speech[]);
64void synthesise_one_frame(struct CODEC2 *c2, short speech[], MODEL *model, 62void synthesise_one_frame(struct CODEC2 *c2, short speech[], MODEL *model,
65 COMP Aw[], float gain); 63 COMP Aw[], float gain);
66void codec2_encode_3200(struct CODEC2 *c2, unsigned char * bits, short speech[]); 64void codec2_encode_3200(struct CODEC2 *c2, unsigned char *bits, short speech[]);
67void codec2_decode_3200(struct CODEC2 *c2, short speech[], const unsigned char * bits); 65void codec2_decode_3200(struct CODEC2 *c2, short speech[],
68void codec2_encode_2400(struct CODEC2 *c2, unsigned char * bits, short speech[]); 66 const unsigned char *bits);
69void codec2_decode_2400(struct CODEC2 *c2, short speech[], const unsigned char * bits); 67void codec2_encode_2400(struct CODEC2 *c2, unsigned char *bits, short speech[]);
70void codec2_encode_1600(struct CODEC2 *c2, unsigned char * bits, short speech[]); 68void codec2_decode_2400(struct CODEC2 *c2, short speech[],
71void codec2_decode_1600(struct CODEC2 *c2, short speech[], const unsigned char * bits); 69 const unsigned char *bits);
72void codec2_encode_1400(struct CODEC2 *c2, unsigned char * bits, short speech[]); 70void codec2_encode_1600(struct CODEC2 *c2, unsigned char *bits, short speech[]);
73void codec2_decode_1400(struct CODEC2 *c2, short speech[], const unsigned char * bits); 71void codec2_decode_1600(struct CODEC2 *c2, short speech[],
74void codec2_encode_1300(struct CODEC2 *c2, unsigned char * bits, short speech[]); 72 const unsigned char *bits);
75void codec2_decode_1300(struct CODEC2 *c2, short speech[], const unsigned char * bits, float ber_est); 73void codec2_encode_1400(struct CODEC2 *c2, unsigned char *bits, short speech[]);
76void codec2_encode_1200(struct CODEC2 *c2, unsigned char * bits, short speech[]); 74void codec2_decode_1400(struct CODEC2 *c2, short speech[],
77void codec2_decode_1200(struct CODEC2 *c2, short speech[], const unsigned char * bits); 75 const unsigned char *bits);
78void codec2_encode_700(struct CODEC2 *c2, unsigned char * bits, short speech[]); 76void codec2_encode_1300(struct CODEC2 *c2, unsigned char *bits, short speech[]);
79void codec2_decode_700(struct CODEC2 *c2, short speech[], const unsigned char * bits); 77void codec2_decode_1300(struct CODEC2 *c2, short speech[],
80void codec2_encode_700b(struct CODEC2 *c2, unsigned char * bits, short speech[]); 78 const unsigned char *bits, float ber_est);
81void codec2_decode_700b(struct CODEC2 *c2, short speech[], const unsigned char * bits); 79void codec2_encode_1200(struct CODEC2 *c2, unsigned char *bits, short speech[]);
82void codec2_encode_700c(struct CODEC2 *c2, unsigned char * bits, short speech[]); 80void codec2_decode_1200(struct CODEC2 *c2, short speech[],
83void codec2_decode_700c(struct CODEC2 *c2, short speech[], const unsigned char * bits); 81 const unsigned char *bits);
84void codec2_encode_450(struct CODEC2 *c2, unsigned char * bits, short speech[]); 82void codec2_encode_700c(struct CODEC2 *c2, unsigned char *bits, short speech[]);
85void codec2_decode_450(struct CODEC2 *c2, short speech[], const unsigned char * bits); 83void codec2_decode_700c(struct CODEC2 *c2, short speech[],
86void codec2_decode_450pwb(struct CODEC2 *c2, short speech[], const unsigned char * bits); 84 const unsigned char *bits);
87static void ear_protection(float in_out[], int n); 85static void ear_protection(float in_out[], int n);
88 86
89
90
91/*---------------------------------------------------------------------------*\ 87/*---------------------------------------------------------------------------*\
92 88
93 FUNCTIONS 89 FUNCTIONS
@@ -108,247 +104,170 @@ static void ear_protection(float in_out[], int n);
108 104
109\*---------------------------------------------------------------------------*/ 105\*---------------------------------------------------------------------------*/
110 106
111 107struct CODEC2 *codec2_create(int mode) {
112//Don't create CODEC2_MODE_450PWB for Encoding as it has undefined behavior ! 108 struct CODEC2 *c2;
113struct CODEC2 * codec2_create(int mode) 109 int i, l;
114{ 110
115 struct CODEC2 *c2; 111 // ALL POSSIBLE MODES MUST BE CHECKED HERE!
116 int i,l; 112 // we test if the desired mode is enabled at compile time
117 113 // and return NULL if not
118 // ALL POSSIBLE MODES MUST BE CHECKED HERE! 114
119 // we test if the desired mode is enabled at compile time 115 if (false == (CODEC2_MODE_ACTIVE(CODEC2_MODE_3200, mode) ||
120 // and return NULL if not 116 CODEC2_MODE_ACTIVE(CODEC2_MODE_2400, mode) ||
121 117 CODEC2_MODE_ACTIVE(CODEC2_MODE_1600, mode) ||
122 if (false == ( CODEC2_MODE_ACTIVE(CODEC2_MODE_3200, mode) 118 CODEC2_MODE_ACTIVE(CODEC2_MODE_1400, mode) ||
123 || CODEC2_MODE_ACTIVE(CODEC2_MODE_2400, mode) 119 CODEC2_MODE_ACTIVE(CODEC2_MODE_1300, mode) ||
124 || CODEC2_MODE_ACTIVE(CODEC2_MODE_1600, mode) 120 CODEC2_MODE_ACTIVE(CODEC2_MODE_1200, mode) ||
125 || CODEC2_MODE_ACTIVE(CODEC2_MODE_1400, mode) 121 CODEC2_MODE_ACTIVE(CODEC2_MODE_700C, mode))) {
126 || CODEC2_MODE_ACTIVE(CODEC2_MODE_1300, mode) 122 return NULL;
127 || CODEC2_MODE_ACTIVE(CODEC2_MODE_1200, mode) 123 }
128 || CODEC2_MODE_ACTIVE(CODEC2_MODE_700, mode) 124
129 || CODEC2_MODE_ACTIVE(CODEC2_MODE_700B, mode) 125 c2 = (struct CODEC2 *)MALLOC(sizeof(struct CODEC2));
130 || CODEC2_MODE_ACTIVE(CODEC2_MODE_700C, mode) 126 if (c2 == NULL) return NULL;
131 || CODEC2_MODE_ACTIVE(CODEC2_MODE_450, mode) 127
132 || CODEC2_MODE_ACTIVE(CODEC2_MODE_450PWB, mode) 128 c2->mode = mode;
133 ) ) 129
134 { 130 /* store constants in a few places for convenience */
135 return NULL; 131
136 } 132 c2->c2const = c2const_create(8000, N_S);
137 133 c2->Fs = c2->c2const.Fs;
138 c2 = (struct CODEC2*)MALLOC(sizeof(struct CODEC2)); 134 int n_samp = c2->n_samp = c2->c2const.n_samp;
139 if (c2 == NULL) 135 int m_pitch = c2->m_pitch = c2->c2const.m_pitch;
140 return NULL; 136
141 137 c2->Pn = (float *)MALLOC(2 * n_samp * sizeof(float));
142 c2->mode = mode; 138 if (c2->Pn == NULL) {
143 139 return NULL;
144 /* store constants in a few places for convenience */ 140 }
145 141 c2->Sn_ = (float *)MALLOC(2 * n_samp * sizeof(float));
146 if( CODEC2_MODE_ACTIVE(CODEC2_MODE_450PWB, mode) == 0){ 142 if (c2->Sn_ == NULL) {
147 c2->c2const = c2const_create(8000, N_S); 143 FREE(c2->Pn);
148 }else{ 144 return NULL;
149 c2->c2const = c2const_create(16000, N_S); 145 }
150 } 146 c2->w = (float *)MALLOC(m_pitch * sizeof(float));
151 c2->Fs = c2->c2const.Fs; 147 if (c2->w == NULL) {
152 int n_samp = c2->n_samp = c2->c2const.n_samp; 148 FREE(c2->Pn);
153 int m_pitch = c2->m_pitch = c2->c2const.m_pitch; 149 FREE(c2->Sn_);
154 150 return NULL;
155 c2->Pn = (float*)MALLOC(2*n_samp*sizeof(float)); 151 }
156 if (c2->Pn == NULL) { 152 c2->Sn = (float *)MALLOC(m_pitch * sizeof(float));
157 return NULL; 153 if (c2->Sn == NULL) {
158 } 154 FREE(c2->Pn);
159 c2->Sn_ = (float*)MALLOC(2*n_samp*sizeof(float)); 155 FREE(c2->Sn_);
160 if (c2->Sn_ == NULL) { 156 FREE(c2->w);
161 FREE(c2->Pn); 157 return NULL;
162 return NULL; 158 }
163 } 159
164 c2->w = (float*)MALLOC(m_pitch*sizeof(float)); 160 for (i = 0; i < m_pitch; i++) c2->Sn[i] = 1.0;
165 if (c2->w == NULL) { 161 c2->hpf_states[0] = c2->hpf_states[1] = 0.0;
166 FREE(c2->Pn); 162 for (i = 0; i < 2 * n_samp; i++) c2->Sn_[i] = 0;
167 FREE(c2->Sn_); 163 c2->fft_fwd_cfg = codec2_fft_alloc(FFT_ENC, 0, NULL, NULL);
168 return NULL; 164 c2->fftr_fwd_cfg = codec2_fftr_alloc(FFT_ENC, 0, NULL, NULL);
169 } 165 make_analysis_window(&c2->c2const, c2->fft_fwd_cfg, c2->w, c2->W);
170 c2->Sn = (float*)MALLOC(m_pitch*sizeof(float)); 166 make_synthesis_window(&c2->c2const, c2->Pn);
171 if (c2->Sn == NULL) { 167 c2->fftr_inv_cfg = codec2_fftr_alloc(FFT_DEC, 1, NULL, NULL);
172 FREE(c2->Pn); 168 c2->prev_f0_enc = 1 / P_MAX_S;
173 FREE(c2->Sn_); 169 c2->bg_est = 0.0;
174 FREE(c2->w); 170 c2->ex_phase = 0.0;
175 return NULL; 171
176 } 172 for (l = 1; l <= MAX_AMP; l++) c2->prev_model_dec.A[l] = 0.0;
177 173 c2->prev_model_dec.Wo = TWO_PI / c2->c2const.p_max;
178 for(i=0; i<m_pitch; i++) 174 c2->prev_model_dec.L = PI / c2->prev_model_dec.Wo;
179 c2->Sn[i] = 1.0; 175 c2->prev_model_dec.voiced = 0;
180 c2->hpf_states[0] = c2->hpf_states[1] = 0.0; 176
181 for(i=0; i<2*n_samp; i++) 177 for (i = 0; i < LPC_ORD; i++) {
182 c2->Sn_[i] = 0; 178 c2->prev_lsps_dec[i] = i * PI / (LPC_ORD + 1);
183 c2->fft_fwd_cfg = codec2_fft_alloc(FFT_ENC, 0, NULL, NULL); 179 }
184 c2->fftr_fwd_cfg = codec2_fftr_alloc(FFT_ENC, 0, NULL, NULL); 180 c2->prev_e_dec = 1;
185 make_analysis_window(&c2->c2const, c2->fft_fwd_cfg, c2->w,c2->W); 181
186 make_synthesis_window(&c2->c2const, c2->Pn); 182 c2->nlp = nlp_create(&c2->c2const);
187 c2->fftr_inv_cfg = codec2_fftr_alloc(FFT_DEC, 1, NULL, NULL); 183 if (c2->nlp == NULL) {
188 quantise_init(); 184 return NULL;
189 c2->prev_f0_enc = 1/P_MAX_S; 185 }
190 c2->bg_est = 0.0; 186
191 c2->ex_phase = 0.0; 187 c2->lpc_pf = 1;
192 188 c2->bass_boost = 1;
193 for(l=1; l<=MAX_AMP; l++) 189 c2->beta = LPCPF_BETA;
194 c2->prev_model_dec.A[l] = 0.0; 190 c2->gamma = LPCPF_GAMMA;
195 c2->prev_model_dec.Wo = TWO_PI/c2->c2const.p_max; 191
196 c2->prev_model_dec.L = PI/c2->prev_model_dec.Wo; 192 c2->xq_enc[0] = c2->xq_enc[1] = 0.0;
197 c2->prev_model_dec.voiced = 0; 193 c2->xq_dec[0] = c2->xq_dec[1] = 0.0;
198 194
199 for(i=0; i<LPC_ORD; i++) { 195 c2->smoothing = 0;
200 c2->prev_lsps_dec[i] = i*PI/(LPC_ORD+1); 196 c2->se = 0.0;
201 } 197 c2->nse = 0;
202 c2->prev_e_dec = 1; 198 c2->user_rate_K_vec_no_mean_ = NULL;
203 199 c2->post_filter_en = true;
204 c2->nlp = nlp_create(&c2->c2const); 200
205 if (c2->nlp == NULL) { 201 c2->bpf_buf = (float *)MALLOC(sizeof(float) * (BPF_N + 4 * c2->n_samp));
206 return NULL; 202 assert(c2->bpf_buf != NULL);
207 } 203 for (i = 0; i < BPF_N + 4 * c2->n_samp; i++) c2->bpf_buf[i] = 0.0;
208 204
209 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_700B, mode)) 205 c2->softdec = NULL;
210 c2->gray = 0; // natural binary better for trellis decoding (hopefully added later) 206 c2->gray = 1;
211 else 207
212 c2->gray = 1; 208 /* newamp1 initialisation */
213 209
214 c2->lpc_pf = 1; c2->bass_boost = 1; c2->beta = LPCPF_BETA; c2->gamma = LPCPF_GAMMA; 210 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_700C, c2->mode)) {
215 211 mel_sample_freqs_kHz(c2->rate_K_sample_freqs_kHz, NEWAMP1_K, ftomel(200.0),
216 c2->xq_enc[0] = c2->xq_enc[1] = 0.0; 212 ftomel(3700.0));
217 c2->xq_dec[0] = c2->xq_dec[1] = 0.0; 213 int k;
218 214 for (k = 0; k < NEWAMP1_K; k++) {
219 c2->smoothing = 0; 215 c2->prev_rate_K_vec_[k] = 0.0;
220 c2->se = 0.0; c2->nse = 0; 216 c2->eq[k] = 0.0;
221 c2->user_rate_K_vec_no_mean_ = NULL; 217 }
222 c2->post_filter_en = 1; 218 c2->eq_en = false;
223 219 c2->Wo_left = 0.0;
224 c2->bpf_buf = (float*)MALLOC(sizeof(float)*(BPF_N+4*c2->n_samp)); 220 c2->voicing_left = 0;
225 assert(c2->bpf_buf != NULL); 221 c2->phase_fft_fwd_cfg = codec2_fft_alloc(NEWAMP1_PHASE_NFFT, 0, NULL, NULL);
226 for(i=0; i<BPF_N+4*c2->n_samp; i++) 222 c2->phase_fft_inv_cfg = codec2_fft_alloc(NEWAMP1_PHASE_NFFT, 1, NULL, NULL);
227 c2->bpf_buf[i] = 0.0; 223 }
228 224
229 c2->softdec = NULL; 225 c2->fmlfeat = NULL;
230 226 c2->fmlmodel = NULL;
231 /* newamp1 initialisation */ 227
232 228 // make sure that one of the two decode function pointers is empty
233 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_700C, c2->mode)) { 229 // for the encode function pointer this is not required since we always set it
234 mel_sample_freqs_kHz(c2->rate_K_sample_freqs_kHz, NEWAMP1_K, ftomel(200.0), ftomel(3700.0) ); 230 // to a meaningful value
235 int k; 231
236 for(k=0; k<NEWAMP1_K; k++) { 232 c2->decode = NULL;
237 c2->prev_rate_K_vec_[k] = 0.0; 233 c2->decode_ber = NULL;
238 c2->eq[k] = 0.0; 234
239 } 235 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_3200, c2->mode)) {
240 c2->eq_en = 0; 236 c2->encode = codec2_encode_3200;
241 c2->Wo_left = 0.0; 237 c2->decode = codec2_decode_3200;
242 c2->voicing_left = 0;; 238 }
243 c2->phase_fft_fwd_cfg = codec2_fft_alloc(NEWAMP1_PHASE_NFFT, 0, NULL, NULL); 239
244 c2->phase_fft_inv_cfg = codec2_fft_alloc(NEWAMP1_PHASE_NFFT, 1, NULL, NULL); 240 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_2400, c2->mode)) {
245 } 241 c2->encode = codec2_encode_2400;
246 242 c2->decode = codec2_decode_2400;
247 /* newamp2 initialisation */ 243 }
248 244
249 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_450, c2->mode)) { 245 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_1600, c2->mode)) {
250 n2_mel_sample_freqs_kHz(c2->n2_rate_K_sample_freqs_kHz, NEWAMP2_K); 246 c2->encode = codec2_encode_1600;
251 int k; 247 c2->decode = codec2_decode_1600;
252 for(k=0; k<NEWAMP2_K; k++) { 248 }
253 c2->n2_prev_rate_K_vec_[k] = 0.0; 249
254 } 250 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_1400, c2->mode)) {
255 c2->Wo_left = 0.0; 251 c2->encode = codec2_encode_1400;
256 c2->voicing_left = 0;; 252 c2->decode = codec2_decode_1400;
257 c2->phase_fft_fwd_cfg = codec2_fft_alloc(NEWAMP2_PHASE_NFFT, 0, NULL, NULL); 253 }
258 c2->phase_fft_inv_cfg = codec2_fft_alloc(NEWAMP2_PHASE_NFFT, 1, NULL, NULL); 254
259 } 255 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_1300, c2->mode)) {
260 /* newamp2 PWB initialisation */ 256 c2->encode = codec2_encode_1300;
261 257 c2->decode_ber = codec2_decode_1300;
262 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_450PWB, c2->mode)) { 258 }
263 n2_mel_sample_freqs_kHz(c2->n2_pwb_rate_K_sample_freqs_kHz, NEWAMP2_16K_K); 259
264 int k; 260 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_1200, c2->mode)) {
265 for(k=0; k<NEWAMP2_16K_K; k++) { 261 c2->encode = codec2_encode_1200;
266 c2->n2_pwb_prev_rate_K_vec_[k] = 0.0; 262 c2->decode = codec2_decode_1200;
267 } 263 }
268 c2->Wo_left = 0.0; 264
269 c2->voicing_left = 0;; 265 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_700C, c2->mode)) {
270 c2->phase_fft_fwd_cfg = codec2_fft_alloc(NEWAMP2_PHASE_NFFT, 0, NULL, NULL); 266 c2->encode = codec2_encode_700c;
271 c2->phase_fft_inv_cfg = codec2_fft_alloc(NEWAMP2_PHASE_NFFT, 1, NULL, NULL); 267 c2->decode = codec2_decode_700c;
272 } 268 }
273 269
274 c2->fmlfeat = NULL; 270 return c2;
275
276 // make sure that one of the two decode function pointers is empty
277 // for the encode function pointer this is not required since we always set it
278 // to a meaningful value
279
280 c2->decode = NULL;
281 c2->decode_ber = NULL;
282
283 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_3200, c2->mode))
284 {
285 c2->encode = codec2_encode_3200;
286 c2->decode = codec2_decode_3200;
287 }
288
289 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_2400, c2->mode))
290 {
291 c2->encode = codec2_encode_2400;
292 c2->decode = codec2_decode_2400;
293 }
294
295 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1600, c2->mode))
296 {
297 c2->encode = codec2_encode_1600;
298 c2->decode = codec2_decode_1600;
299 }
300
301 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1400, c2->mode))
302 {
303 c2->encode = codec2_encode_1400;
304 c2->decode = codec2_decode_1400;
305 }
306
307 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1300, c2->mode))
308 {
309 c2->encode = codec2_encode_1300;
310 c2->decode_ber = codec2_decode_1300;
311 }
312
313 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1200, c2->mode))
314 {
315 c2->encode = codec2_encode_1200;
316 c2->decode = codec2_decode_1200;
317 }
318
319 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_700, c2->mode))
320 {
321 c2->encode = codec2_encode_700;
322 c2->decode = codec2_decode_700;
323 }
324
325 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_700B, c2->mode))
326 {
327 c2->encode = codec2_encode_700b;
328 c2->decode = codec2_decode_700b;
329 }
330
331 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_700C, c2->mode))
332 {
333 c2->encode = codec2_encode_700c;
334 c2->decode = codec2_decode_700c;
335 }
336
337 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_450, c2->mode))
338 {
339 c2->encode = codec2_encode_450;
340 c2->decode = codec2_decode_450;
341 }
342
343 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_450PWB, c2->mode))
344 {
345 //Encode PWB doesnt make sense
346 c2->encode = codec2_encode_450;
347 c2->decode = codec2_decode_450pwb;
348 }
349
350
351 return c2;
352} 271}
353 272
354/*---------------------------------------------------------------------------*\ 273/*---------------------------------------------------------------------------*\
@@ -361,31 +280,22 @@ struct CODEC2 * codec2_create(int mode)
361 280
362\*---------------------------------------------------------------------------*/ 281\*---------------------------------------------------------------------------*/
363 282
364void codec2_destroy(struct CODEC2 *c2) 283void codec2_destroy(struct CODEC2 *c2) {
365{ 284 assert(c2 != NULL);
366 assert(c2 != NULL); 285 FREE(c2->bpf_buf);
367 FREE(c2->bpf_buf); 286 nlp_destroy(c2->nlp);
368 nlp_destroy(c2->nlp); 287 codec2_fft_free(c2->fft_fwd_cfg);
369 codec2_fft_free(c2->fft_fwd_cfg); 288 codec2_fftr_free(c2->fftr_fwd_cfg);
370 codec2_fftr_free(c2->fftr_fwd_cfg); 289 codec2_fftr_free(c2->fftr_inv_cfg);
371 codec2_fftr_free(c2->fftr_inv_cfg); 290 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_700C, c2->mode)) {
372 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_700C, c2->mode)) { 291 codec2_fft_free(c2->phase_fft_fwd_cfg);
373 codec2_fft_free(c2->phase_fft_fwd_cfg); 292 codec2_fft_free(c2->phase_fft_inv_cfg);
374 codec2_fft_free(c2->phase_fft_inv_cfg); 293 }
375 } 294 FREE(c2->Pn);
376 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_450, c2->mode)) { 295 FREE(c2->Sn);
377 codec2_fft_free(c2->phase_fft_fwd_cfg); 296 FREE(c2->w);
378 codec2_fft_free(c2->phase_fft_inv_cfg); 297 FREE(c2->Sn_);
379 } 298 FREE(c2);
380 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_450PWB, c2->mode)) {
381 codec2_fft_free(c2->phase_fft_fwd_cfg);
382 codec2_fft_free(c2->phase_fft_inv_cfg);
383 }
384 FREE(c2->Pn);
385 FREE(c2->Sn);
386 FREE(c2->w);
387 FREE(c2->Sn_);
388 FREE(c2);
389} 299}
390 300
391/*---------------------------------------------------------------------------*\ 301/*---------------------------------------------------------------------------*\
@@ -399,32 +309,31 @@ void codec2_destroy(struct CODEC2 *c2)
399\*---------------------------------------------------------------------------*/ 309\*---------------------------------------------------------------------------*/
400 310
401int codec2_bits_per_frame(struct CODEC2 *c2) { 311int codec2_bits_per_frame(struct CODEC2 *c2) {
402 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_3200, c2->mode)) 312 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_3200, c2->mode)) return 64;
403 return 64; 313 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_2400, c2->mode)) return 48;
404 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_2400, c2->mode)) 314 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_1600, c2->mode)) return 64;
405 return 48; 315 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_1400, c2->mode)) return 56;
406 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1600, c2->mode)) 316 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_1300, c2->mode)) return 52;
407 return 64; 317 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_1200, c2->mode)) return 48;
408 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1400, c2->mode)) 318 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_700C, c2->mode)) return 28;
409 return 56; 319
410 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1300, c2->mode)) 320 return 0; /* shouldn't get here */
411 return 52;
412 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1200, c2->mode))
413 return 48;
414 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_700, c2->mode))
415 return 28;
416 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_700B, c2->mode))
417 return 28;
418 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_700C, c2->mode))
419 return 28;
420 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_450, c2->mode))
421 return 18;
422 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_450PWB, c2->mode))
423 return 18;
424
425 return 0; /* shouldn't get here */
426} 321}
427 322
323/*---------------------------------------------------------------------------*\
324
325 FUNCTION....: codec2_bytes_per_frame
326 DATE CREATED: April 2021
327
328 Returns the number of bytes per frame. Useful for allocated storage for
329 codec2_encode()/codec2_decode(). Note the number of bits may not be a
330 multiple of 8, therefore some bits in the last byte may be unused.
331
332\*---------------------------------------------------------------------------*/
333
334int codec2_bytes_per_frame(struct CODEC2 *c2) {
335 return (codec2_bits_per_frame(c2) + 7) / 8;
336}
428 337
429/*---------------------------------------------------------------------------*\ 338/*---------------------------------------------------------------------------*\
430 339
@@ -437,60 +346,61 @@ int codec2_bits_per_frame(struct CODEC2 *c2) {
437\*---------------------------------------------------------------------------*/ 346\*---------------------------------------------------------------------------*/
438 347
439int codec2_samples_per_frame(struct CODEC2 *c2) { 348int codec2_samples_per_frame(struct CODEC2 *c2) {
440 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_3200, c2->mode)) 349 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_3200, c2->mode)) return 160;
441 return 160; 350 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_2400, c2->mode)) return 160;
442 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_2400, c2->mode)) 351 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_1600, c2->mode)) return 320;
443 return 160; 352 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_1400, c2->mode)) return 320;
444 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1600, c2->mode)) 353 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_1300, c2->mode)) return 320;
445 return 320; 354 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_1200, c2->mode)) return 320;
446 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1400, c2->mode)) 355 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_700C, c2->mode)) return 320;
447 return 320; 356 return 0; /* shouldn't get here */
448 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1300, c2->mode))
449 return 320;
450 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1200, c2->mode))
451 return 320;
452 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_700, c2->mode))
453 return 320;
454 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_700B, c2->mode))
455 return 320;
456 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_700C, c2->mode))
457 return 320;
458 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_450, c2->mode))
459 return 320;
460 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_450PWB, c2->mode))
461 return 640;
462 return 0; /* shouldnt get here */
463} 357}
464 358
465void codec2_encode(struct CODEC2 *c2, unsigned char *bits, short speech[]) 359/*---------------------------------------------------------------------------*\
466{ 360
467 assert(c2 != NULL); 361 FUNCTION....: codec2_encode
468 assert(c2->encode != NULL); 362 AUTHOR......: David Rowe
363 DATE CREATED: Nov 14 2011
364
365 Take an input buffer of speech samples, and compress them to a packed buffer
366 of bytes.
469 367
470 c2->encode(c2, bits, speech); 368\*---------------------------------------------------------------------------*/
471 369
472} 370void codec2_encode(struct CODEC2 *c2, unsigned char *bytes, short speech[]) {
371 assert(c2 != NULL);
372 assert(c2->encode != NULL);
473 373
474void codec2_decode(struct CODEC2 *c2, short speech[], const unsigned char *bits) 374 c2->encode(c2, bytes, speech);
475{
476 codec2_decode_ber(c2, speech, bits, 0.0);
477} 375}
478 376
479void codec2_decode_ber(struct CODEC2 *c2, short speech[], const unsigned char *bits, float ber_est) 377/*---------------------------------------------------------------------------*\
480{
481 assert(c2 != NULL);
482 assert(c2->decode != NULL || c2->decode_ber != NULL);
483 378
484 if (c2->decode != NULL) 379 FUNCTION....: codec2_decode
485 { 380 AUTHOR......: David Rowe
486 c2->decode(c2, speech, bits); 381 DATE CREATED: Nov 14 2011
487 } 382
488 else 383 Take an input packed buffer of bytes, and decode them to a buffer of speech
489 { 384 samples.
490 c2->decode_ber(c2, speech, bits, ber_est); 385
491 } 386\*---------------------------------------------------------------------------*/
387
388void codec2_decode(struct CODEC2 *c2, short speech[],
389 const unsigned char *bytes) {
390 codec2_decode_ber(c2, speech, bytes, 0.0);
492} 391}
493 392
393void codec2_decode_ber(struct CODEC2 *c2, short speech[],
394 const unsigned char *bits, float ber_est) {
395 assert(c2 != NULL);
396 assert(c2->decode != NULL || c2->decode_ber != NULL);
397
398 if (c2->decode != NULL) {
399 c2->decode(c2, speech, bits);
400 } else {
401 c2->decode_ber(c2, speech, bits, ber_est);
402 }
403}
494 404
495/*---------------------------------------------------------------------------*\ 405/*---------------------------------------------------------------------------*\
496 406
@@ -503,60 +413,60 @@ void codec2_decode_ber(struct CODEC2 *c2, short speech[], const unsigned char *b
503 The codec2 algorithm actually operates internally on 10ms (80 413 The codec2 algorithm actually operates internally on 10ms (80
504 sample) frames, so we run the encoding algorithm twice. On the 414 sample) frames, so we run the encoding algorithm twice. On the
505 first frame we just send the voicing bits. On the second frame we 415 first frame we just send the voicing bits. On the second frame we
506 send all model parameters. Compared to 2400 we use a larger number 416 send all model parameters. Compared to 2400 we encode the LSP
507 of bits for the LSPs and non-VQ pitch and energy. 417 differences, a larger number of bits for the LSP(d)s and scalar
418 (non-VQ) quantisation for pitch and energy.
508 419
509 The bit allocation is: 420 The bit allocation is:
510 421
511 Parameter bits/frame 422 Parameter bits/frame
512 -------------------------------------- 423 ------------------------------------------------------
513 Harmonic magnitudes (LSPs) 50 424 Harmonic magnitudes (LSP differerences) 50
514 Pitch (Wo) 7 425 Pitch (Wo) 7
515 Energy 5 426 Energy 5
516 Voicing (10ms update) 2 427 Voicing (10ms update) 2
517 TOTAL 64 428 TOTAL 64
518 429
519\*---------------------------------------------------------------------------*/ 430\*---------------------------------------------------------------------------*/
520 431
521void codec2_encode_3200(struct CODEC2 *c2, unsigned char * bits, short speech[]) 432void codec2_encode_3200(struct CODEC2 *c2, unsigned char *bits,
522{ 433 short speech[]) {
523 MODEL model; 434 MODEL model;
524 float ak[LPC_ORD+1]; 435 float ak[LPC_ORD + 1];
525 float lsps[LPC_ORD]; 436 float lsps[LPC_ORD];
526 float e; 437 float e;
527 int Wo_index, e_index; 438 int Wo_index, e_index;
528 int lspd_indexes[LPC_ORD]; 439 int lspd_indexes[LPC_ORD];
529 int i; 440 int i;
530 unsigned int nbit = 0; 441 unsigned int nbit = 0;
531 442
532 assert(c2 != NULL); 443 assert(c2 != NULL);
533 444
534 memset(bits, '\0', ((codec2_bits_per_frame(c2) + 7) / 8)); 445 memset(bits, '\0', ((codec2_bits_per_frame(c2) + 7) / 8));
535 446
536 /* first 10ms analysis frame - we just want voicing */ 447 /* first 10ms analysis frame - we just want voicing */
537 448
538 analyse_one_frame(c2, &model, speech); 449 analyse_one_frame(c2, &model, speech);
539 pack(bits, &nbit, model.voiced, 1); 450 pack(bits, &nbit, model.voiced, 1);
540 451
541 /* second 10ms analysis frame */ 452 /* second 10ms analysis frame */
542 453
543 analyse_one_frame(c2, &model, &speech[c2->n_samp]); 454 analyse_one_frame(c2, &model, &speech[c2->n_samp]);
544 pack(bits, &nbit, model.voiced, 1); 455 pack(bits, &nbit, model.voiced, 1);
545 Wo_index = encode_Wo(&c2->c2const, model.Wo, WO_BITS); 456 Wo_index = encode_Wo(&c2->c2const, model.Wo, WO_BITS);
546 pack(bits, &nbit, Wo_index, WO_BITS); 457 pack(bits, &nbit, Wo_index, WO_BITS);
547 458
548 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD); 459 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD);
549 e_index = encode_energy(e, E_BITS); 460 e_index = encode_energy(e, E_BITS);
550 pack(bits, &nbit, e_index, E_BITS); 461 pack(bits, &nbit, e_index, E_BITS);
551 462
552 encode_lspds_scalar(lspd_indexes, lsps, LPC_ORD); 463 encode_lspds_scalar(lspd_indexes, lsps, LPC_ORD);
553 for(i=0; i<LSPD_SCALAR_INDEXES; i++) { 464 for (i = 0; i < LSPD_SCALAR_INDEXES; i++) {
554 pack(bits, &nbit, lspd_indexes[i], lspd_bits(i)); 465 pack(bits, &nbit, lspd_indexes[i], lspd_bits(i));
555 } 466 }
556 assert(nbit == (unsigned)codec2_bits_per_frame(c2)); 467 assert(nbit == (unsigned)codec2_bits_per_frame(c2));
557} 468}
558 469
559
560/*---------------------------------------------------------------------------*\ 470/*---------------------------------------------------------------------------*\
561 471
562 FUNCTION....: codec2_decode_3200 472 FUNCTION....: codec2_decode_3200
@@ -567,77 +477,75 @@ void codec2_encode_3200(struct CODEC2 *c2, unsigned char * bits, short speech[])
567 477
568\*---------------------------------------------------------------------------*/ 478\*---------------------------------------------------------------------------*/
569 479
570void codec2_decode_3200(struct CODEC2 *c2, short speech[], const unsigned char * bits) 480void codec2_decode_3200(struct CODEC2 *c2, short speech[],
571{ 481 const unsigned char *bits) {
572 MODEL model[2]; 482 MODEL model[2];
573 int lspd_indexes[LPC_ORD]; 483 int lspd_indexes[LPC_ORD];
574 float lsps[2][LPC_ORD]; 484 float lsps[2][LPC_ORD];
575 int Wo_index, e_index; 485 int Wo_index, e_index;
576 float e[2]; 486 float e[2];
577 float snr; 487 float snr;
578 float ak[2][LPC_ORD+1]; 488 float ak[2][LPC_ORD + 1];
579 int i,j; 489 int i, j;
580 unsigned int nbit = 0; 490 unsigned int nbit = 0;
581 COMP Aw[FFT_ENC]; 491 COMP Aw[FFT_ENC];
582 492
583 assert(c2 != NULL); 493 assert(c2 != NULL);
584 494
585 /* only need to zero these out due to (unused) snr calculation */ 495 /* only need to zero these out due to (unused) snr calculation */
586 496
587 for(i=0; i<2; i++) 497 for (i = 0; i < 2; i++)
588 for(j=1; j<=MAX_AMP; j++) 498 for (j = 1; j <= MAX_AMP; j++) model[i].A[j] = 0.0;
589 model[i].A[j] = 0.0;
590 499
591 /* unpack bits from channel ------------------------------------*/ 500 /* unpack bits from channel ------------------------------------*/
592 501
593 /* this will partially fill the model params for the 2 x 10ms 502 /* this will partially fill the model params for the 2 x 10ms
594 frames */ 503 frames */
595 504
596 model[0].voiced = unpack(bits, &nbit, 1); 505 model[0].voiced = unpack(bits, &nbit, 1);
597 model[1].voiced = unpack(bits, &nbit, 1); 506 model[1].voiced = unpack(bits, &nbit, 1);
598 507
599 Wo_index = unpack(bits, &nbit, WO_BITS); 508 Wo_index = unpack(bits, &nbit, WO_BITS);
600 model[1].Wo = decode_Wo(&c2->c2const, Wo_index, WO_BITS); 509 model[1].Wo = decode_Wo(&c2->c2const, Wo_index, WO_BITS);
601 model[1].L = PI/model[1].Wo; 510 model[1].L = PI / model[1].Wo;
602 511
603 e_index = unpack(bits, &nbit, E_BITS); 512 e_index = unpack(bits, &nbit, E_BITS);
604 e[1] = decode_energy(e_index, E_BITS); 513 e[1] = decode_energy(e_index, E_BITS);
605 514
606 for(i=0; i<LSPD_SCALAR_INDEXES; i++) { 515 for (i = 0; i < LSPD_SCALAR_INDEXES; i++) {
607 lspd_indexes[i] = unpack(bits, &nbit, lspd_bits(i)); 516 lspd_indexes[i] = unpack(bits, &nbit, lspd_bits(i));
608 } 517 }
609 decode_lspds_scalar(&lsps[1][0], lspd_indexes, LPC_ORD); 518 decode_lspds_scalar(&lsps[1][0], lspd_indexes, LPC_ORD);
610 519
611 /* interpolate ------------------------------------------------*/ 520 /* interpolate ------------------------------------------------*/
612 521
613 /* Wo and energy are sampled every 20ms, so we interpolate just 1 522 /* Wo and energy are sampled every 20ms, so we interpolate just 1
614 10ms frame between 20ms samples */ 523 10ms frame between 20ms samples */
615 524
616 interp_Wo(&model[0], &c2->prev_model_dec, &model[1], c2->c2const.Wo_min); 525 interp_Wo(&model[0], &c2->prev_model_dec, &model[1], c2->c2const.Wo_min);
617 e[0] = interp_energy(c2->prev_e_dec, e[1]); 526 e[0] = interp_energy(c2->prev_e_dec, e[1]);
618 527
619 /* LSPs are sampled every 20ms so we interpolate the frame in 528 /* LSPs are sampled every 20ms so we interpolate the frame in
620 between, then recover spectral amplitudes */ 529 between, then recover spectral amplitudes */
621 530
622 interpolate_lsp_ver2(&lsps[0][0], c2->prev_lsps_dec, &lsps[1][0], 0.5, LPC_ORD); 531 interpolate_lsp_ver2(&lsps[0][0], c2->prev_lsps_dec, &lsps[1][0], 0.5,
532 LPC_ORD);
623 533
624 for(i=0; i<2; i++) { 534 for (i = 0; i < 2; i++) {
625 lsp_to_lpc(&lsps[i][0], &ak[i][0], LPC_ORD); 535 lsp_to_lpc(&lsps[i][0], &ak[i][0], LPC_ORD);
626 aks_to_M2(c2->fftr_fwd_cfg, &ak[i][0], LPC_ORD, &model[i], e[i], &snr, 0, 0, 536 aks_to_M2(c2->fftr_fwd_cfg, &ak[i][0], LPC_ORD, &model[i], e[i], &snr, 0, 0,
627 c2->lpc_pf, c2->bass_boost, c2->beta, c2->gamma, Aw); 537 c2->lpc_pf, c2->bass_boost, c2->beta, c2->gamma, Aw);
628 apply_lpc_correction(&model[i]); 538 apply_lpc_correction(&model[i]);
629 synthesise_one_frame(c2, &speech[c2->n_samp*i], &model[i], Aw, 1.0); 539 synthesise_one_frame(c2, &speech[c2->n_samp * i], &model[i], Aw, 1.0);
630 } 540 }
631 541
632 /* update memories for next frame ----------------------------*/ 542 /* update memories for next frame ----------------------------*/
633 543
634 c2->prev_model_dec = model[1]; 544 c2->prev_model_dec = model[1];
635 c2->prev_e_dec = e[1]; 545 c2->prev_e_dec = e[1];
636 for(i=0; i<LPC_ORD; i++) 546 for (i = 0; i < LPC_ORD; i++) c2->prev_lsps_dec[i] = lsps[1][i];
637 c2->prev_lsps_dec[i] = lsps[1][i];
638} 547}
639 548
640
641/*---------------------------------------------------------------------------*\ 549/*---------------------------------------------------------------------------*\
642 550
643 FUNCTION....: codec2_encode_2400 551 FUNCTION....: codec2_encode_2400
@@ -663,46 +571,45 @@ void codec2_decode_3200(struct CODEC2 *c2, short speech[], const unsigned char *
663 571
664\*---------------------------------------------------------------------------*/ 572\*---------------------------------------------------------------------------*/
665 573
666void codec2_encode_2400(struct CODEC2 *c2, unsigned char * bits, short speech[]) 574void codec2_encode_2400(struct CODEC2 *c2, unsigned char *bits,
667{ 575 short speech[]) {
668 MODEL model; 576 MODEL model;
669 float ak[LPC_ORD+1]; 577 float ak[LPC_ORD + 1];
670 float lsps[LPC_ORD]; 578 float lsps[LPC_ORD];
671 float e; 579 float e;
672 int WoE_index; 580 int WoE_index;
673 int lsp_indexes[LPC_ORD]; 581 int lsp_indexes[LPC_ORD];
674 int i; 582 int i;
675 int spare = 0; 583 int spare = 0;
676 unsigned int nbit = 0; 584 unsigned int nbit = 0;
677 585
678 assert(c2 != NULL); 586 assert(c2 != NULL);
679 587
680 memset(bits, '\0', ((codec2_bits_per_frame(c2) + 7) / 8)); 588 memset(bits, '\0', ((codec2_bits_per_frame(c2) + 7) / 8));
681 589
682 /* first 10ms analysis frame - we just want voicing */ 590 /* first 10ms analysis frame - we just want voicing */
683 591
684 analyse_one_frame(c2, &model, speech); 592 analyse_one_frame(c2, &model, speech);
685 pack(bits, &nbit, model.voiced, 1); 593 pack(bits, &nbit, model.voiced, 1);
686 594
687 /* second 10ms analysis frame */ 595 /* second 10ms analysis frame */
688 596
689 analyse_one_frame(c2, &model, &speech[c2->n_samp]); 597 analyse_one_frame(c2, &model, &speech[c2->n_samp]);
690 pack(bits, &nbit, model.voiced, 1); 598 pack(bits, &nbit, model.voiced, 1);
691 599
692 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD); 600 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD);
693 WoE_index = encode_WoE(&model, e, c2->xq_enc); 601 WoE_index = encode_WoE(&model, e, c2->xq_enc);
694 pack(bits, &nbit, WoE_index, WO_E_BITS); 602 pack(bits, &nbit, WoE_index, WO_E_BITS);
695 603
696 encode_lsps_scalar(lsp_indexes, lsps, LPC_ORD); 604 encode_lsps_scalar(lsp_indexes, lsps, LPC_ORD);
697 for(i=0; i<LSP_SCALAR_INDEXES; i++) { 605 for (i = 0; i < LSP_SCALAR_INDEXES; i++) {
698 pack(bits, &nbit, lsp_indexes[i], lsp_bits(i)); 606 pack(bits, &nbit, lsp_indexes[i], lsp_bits(i));
699 } 607 }
700 pack(bits, &nbit, spare, 2); 608 pack(bits, &nbit, spare, 2);
701 609
702 assert(nbit == (unsigned)codec2_bits_per_frame(c2)); 610 assert(nbit == (unsigned)codec2_bits_per_frame(c2));
703} 611}
704 612
705
706/*---------------------------------------------------------------------------*\ 613/*---------------------------------------------------------------------------*\
707 614
708 FUNCTION....: codec2_decode_2400 615 FUNCTION....: codec2_decode_2400
@@ -713,86 +620,84 @@ void codec2_encode_2400(struct CODEC2 *c2, unsigned char * bits, short speech[])
713 620
714\*---------------------------------------------------------------------------*/ 621\*---------------------------------------------------------------------------*/
715 622
716void codec2_decode_2400(struct CODEC2 *c2, short speech[], const unsigned char * bits) 623void codec2_decode_2400(struct CODEC2 *c2, short speech[],
717{ 624 const unsigned char *bits) {
718 MODEL model[2]; 625 MODEL model[2];
719 int lsp_indexes[LPC_ORD]; 626 int lsp_indexes[LPC_ORD];
720 float lsps[2][LPC_ORD]; 627 float lsps[2][LPC_ORD];
721 int WoE_index; 628 int WoE_index;
722 float e[2]; 629 float e[2];
723 float snr; 630 float snr;
724 float ak[2][LPC_ORD+1]; 631 float ak[2][LPC_ORD + 1];
725 int i,j; 632 int i, j;
726 unsigned int nbit = 0; 633 unsigned int nbit = 0;
727 COMP Aw[FFT_ENC]; 634 COMP Aw[FFT_ENC];
728 635
729 assert(c2 != NULL); 636 assert(c2 != NULL);
730 637
731 /* only need to zero these out due to (unused) snr calculation */ 638 /* only need to zero these out due to (unused) snr calculation */
732 639
733 for(i=0; i<2; i++) 640 for (i = 0; i < 2; i++)
734 for(j=1; j<=MAX_AMP; j++) 641 for (j = 1; j <= MAX_AMP; j++) model[i].A[j] = 0.0;
735 model[i].A[j] = 0.0;
736 642
737 /* unpack bits from channel ------------------------------------*/ 643 /* unpack bits from channel ------------------------------------*/
738 644
739 /* this will partially fill the model params for the 2 x 10ms 645 /* this will partially fill the model params for the 2 x 10ms
740 frames */ 646 frames */
741 647
742 model[0].voiced = unpack(bits, &nbit, 1); 648 model[0].voiced = unpack(bits, &nbit, 1);
743 649
744 model[1].voiced = unpack(bits, &nbit, 1); 650 model[1].voiced = unpack(bits, &nbit, 1);
745 WoE_index = unpack(bits, &nbit, WO_E_BITS); 651 WoE_index = unpack(bits, &nbit, WO_E_BITS);
746 decode_WoE(&c2->c2const, &model[1], &e[1], c2->xq_dec, WoE_index); 652 decode_WoE(&c2->c2const, &model[1], &e[1], c2->xq_dec, WoE_index);
747 653
748 for(i=0; i<LSP_SCALAR_INDEXES; i++) { 654 for (i = 0; i < LSP_SCALAR_INDEXES; i++) {
749 lsp_indexes[i] = unpack(bits, &nbit, lsp_bits(i)); 655 lsp_indexes[i] = unpack(bits, &nbit, lsp_bits(i));
750 } 656 }
751 decode_lsps_scalar(&lsps[1][0], lsp_indexes, LPC_ORD); 657 decode_lsps_scalar(&lsps[1][0], lsp_indexes, LPC_ORD);
752 check_lsp_order(&lsps[1][0], LPC_ORD); 658 check_lsp_order(&lsps[1][0], LPC_ORD);
753 bw_expand_lsps(&lsps[1][0], LPC_ORD, 50.0, 100.0); 659 bw_expand_lsps(&lsps[1][0], LPC_ORD, 50.0, 100.0);
754 660
755 /* interpolate ------------------------------------------------*/ 661 /* interpolate ------------------------------------------------*/
756 662
757 /* Wo and energy are sampled every 20ms, so we interpolate just 1 663 /* Wo and energy are sampled every 20ms, so we interpolate just 1
758 10ms frame between 20ms samples */ 664 10ms frame between 20ms samples */
759 665
760 interp_Wo(&model[0], &c2->prev_model_dec, &model[1], c2->c2const.Wo_min); 666 interp_Wo(&model[0], &c2->prev_model_dec, &model[1], c2->c2const.Wo_min);
761 e[0] = interp_energy(c2->prev_e_dec, e[1]); 667 e[0] = interp_energy(c2->prev_e_dec, e[1]);
762 668
763 /* LSPs are sampled every 20ms so we interpolate the frame in 669 /* LSPs are sampled every 20ms so we interpolate the frame in
764 between, then recover spectral amplitudes */ 670 between, then recover spectral amplitudes */
765 671
766 interpolate_lsp_ver2(&lsps[0][0], c2->prev_lsps_dec, &lsps[1][0], 0.5, LPC_ORD); 672 interpolate_lsp_ver2(&lsps[0][0], c2->prev_lsps_dec, &lsps[1][0], 0.5,
767 for(i=0; i<2; i++) { 673 LPC_ORD);
768 lsp_to_lpc(&lsps[i][0], &ak[i][0], LPC_ORD); 674 for (i = 0; i < 2; i++) {
769 aks_to_M2(c2->fftr_fwd_cfg, &ak[i][0], LPC_ORD, &model[i], e[i], &snr, 0, 0, 675 lsp_to_lpc(&lsps[i][0], &ak[i][0], LPC_ORD);
770 c2->lpc_pf, c2->bass_boost, c2->beta, c2->gamma, Aw); 676 aks_to_M2(c2->fftr_fwd_cfg, &ak[i][0], LPC_ORD, &model[i], e[i], &snr, 0, 0,
771 apply_lpc_correction(&model[i]); 677 c2->lpc_pf, c2->bass_boost, c2->beta, c2->gamma, Aw);
772 synthesise_one_frame(c2, &speech[c2->n_samp*i], &model[i], Aw, 1.0); 678 apply_lpc_correction(&model[i]);
773 679 synthesise_one_frame(c2, &speech[c2->n_samp * i], &model[i], Aw, 1.0);
774 /* dump parameters for deep learning experiments */ 680
775 681 /* dump parameters for deep learning experiments */
776 if (c2->fmlfeat != NULL) { 682
777 /* 10 LSPs - energy - Wo - voicing flag - 10 LPCs */ 683 if (c2->fmlfeat != NULL) {
778 fwrite(&lsps[i][0], LPC_ORD, sizeof(float), c2->fmlfeat); 684 /* 10 LSPs - energy - Wo - voicing flag - 10 LPCs */
779 fwrite(&e[i], 1, sizeof(float), c2->fmlfeat); 685 fwrite(&lsps[i][0], LPC_ORD, sizeof(float), c2->fmlfeat);
780 fwrite(&model[i].Wo, 1, sizeof(float), c2->fmlfeat); 686 fwrite(&e[i], 1, sizeof(float), c2->fmlfeat);
781 float voiced_float = model[i].voiced; 687 fwrite(&model[i].Wo, 1, sizeof(float), c2->fmlfeat);
782 fwrite(&voiced_float, 1, sizeof(float), c2->fmlfeat); 688 float voiced_float = model[i].voiced;
783 fwrite(&ak[i][1], LPC_ORD, sizeof(float), c2->fmlfeat); 689 fwrite(&voiced_float, 1, sizeof(float), c2->fmlfeat);
784 } 690 fwrite(&ak[i][1], LPC_ORD, sizeof(float), c2->fmlfeat);
785 } 691 }
692 }
786 693
787 /* update memories for next frame ----------------------------*/ 694 /* update memories for next frame ----------------------------*/
788 695
789 c2->prev_model_dec = model[1]; 696 c2->prev_model_dec = model[1];
790 c2->prev_e_dec = e[1]; 697 c2->prev_e_dec = e[1];
791 for(i=0; i<LPC_ORD; i++) 698 for (i = 0; i < LPC_ORD; i++) c2->prev_lsps_dec[i] = lsps[1][i];
792 c2->prev_lsps_dec[i] = lsps[1][i];
793} 699}
794 700
795
796/*---------------------------------------------------------------------------*\ 701/*---------------------------------------------------------------------------*\
797 702
798 FUNCTION....: codec2_encode_1600 703 FUNCTION....: codec2_encode_1600
@@ -821,65 +726,64 @@ void codec2_decode_2400(struct CODEC2 *c2, short speech[], const unsigned char *
821 726
822\*---------------------------------------------------------------------------*/ 727\*---------------------------------------------------------------------------*/
823 728
824void codec2_encode_1600(struct CODEC2 *c2, unsigned char * bits, short speech[]) 729void codec2_encode_1600(struct CODEC2 *c2, unsigned char *bits,
825{ 730 short speech[]) {
826 MODEL model; 731 MODEL model;
827 float lsps[LPC_ORD]; 732 float lsps[LPC_ORD];
828 float ak[LPC_ORD+1]; 733 float ak[LPC_ORD + 1];
829 float e; 734 float e;
830 int lsp_indexes[LPC_ORD]; 735 int lsp_indexes[LPC_ORD];
831 int Wo_index, e_index; 736 int Wo_index, e_index;
832 int i; 737 int i;
833 unsigned int nbit = 0; 738 unsigned int nbit = 0;
834 739
835 assert(c2 != NULL); 740 assert(c2 != NULL);
836 741
837 memset(bits, '\0', ((codec2_bits_per_frame(c2) + 7) / 8)); 742 memset(bits, '\0', ((codec2_bits_per_frame(c2) + 7) / 8));
838 743
839 /* frame 1: - voicing ---------------------------------------------*/ 744 /* frame 1: - voicing ---------------------------------------------*/
840 745
841 analyse_one_frame(c2, &model, speech); 746 analyse_one_frame(c2, &model, speech);
842 pack(bits, &nbit, model.voiced, 1); 747 pack(bits, &nbit, model.voiced, 1);
843 748
844 /* frame 2: - voicing, scalar Wo & E -------------------------------*/ 749 /* frame 2: - voicing, scalar Wo & E -------------------------------*/
845 750
846 analyse_one_frame(c2, &model, &speech[c2->n_samp]); 751 analyse_one_frame(c2, &model, &speech[c2->n_samp]);
847 pack(bits, &nbit, model.voiced, 1); 752 pack(bits, &nbit, model.voiced, 1);
848 753
849 Wo_index = encode_Wo(&c2->c2const, model.Wo, WO_BITS); 754 Wo_index = encode_Wo(&c2->c2const, model.Wo, WO_BITS);
850 pack(bits, &nbit, Wo_index, WO_BITS); 755 pack(bits, &nbit, Wo_index, WO_BITS);
851 756
852 /* need to run this just to get LPC energy */ 757 /* need to run this just to get LPC energy */
853 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD); 758 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD);
854 e_index = encode_energy(e, E_BITS); 759 e_index = encode_energy(e, E_BITS);
855 pack(bits, &nbit, e_index, E_BITS); 760 pack(bits, &nbit, e_index, E_BITS);
856 761
857 /* frame 3: - voicing ---------------------------------------------*/ 762 /* frame 3: - voicing ---------------------------------------------*/
858 763
859 analyse_one_frame(c2, &model, &speech[2*c2->n_samp]); 764 analyse_one_frame(c2, &model, &speech[2 * c2->n_samp]);
860 pack(bits, &nbit, model.voiced, 1); 765 pack(bits, &nbit, model.voiced, 1);
861 766
862 /* frame 4: - voicing, scalar Wo & E, scalar LSPs ------------------*/ 767 /* frame 4: - voicing, scalar Wo & E, scalar LSPs ------------------*/
863 768
864 analyse_one_frame(c2, &model, &speech[3*c2->n_samp]); 769 analyse_one_frame(c2, &model, &speech[3 * c2->n_samp]);
865 pack(bits, &nbit, model.voiced, 1); 770 pack(bits, &nbit, model.voiced, 1);
866 771
867 Wo_index = encode_Wo(&c2->c2const, model.Wo, WO_BITS); 772 Wo_index = encode_Wo(&c2->c2const, model.Wo, WO_BITS);
868 pack(bits, &nbit, Wo_index, WO_BITS); 773 pack(bits, &nbit, Wo_index, WO_BITS);
869 774
870 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD); 775 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD);
871 e_index = encode_energy(e, E_BITS); 776 e_index = encode_energy(e, E_BITS);
872 pack(bits, &nbit, e_index, E_BITS); 777 pack(bits, &nbit, e_index, E_BITS);
873 778
874 encode_lsps_scalar(lsp_indexes, lsps, LPC_ORD); 779 encode_lsps_scalar(lsp_indexes, lsps, LPC_ORD);
875 for(i=0; i<LSP_SCALAR_INDEXES; i++) { 780 for (i = 0; i < LSP_SCALAR_INDEXES; i++) {
876 pack(bits, &nbit, lsp_indexes[i], lsp_bits(i)); 781 pack(bits, &nbit, lsp_indexes[i], lsp_bits(i));
877 } 782 }
878 783
879 assert(nbit == (unsigned)codec2_bits_per_frame(c2)); 784 assert(nbit == (unsigned)codec2_bits_per_frame(c2));
880} 785}
881 786
882
883/*---------------------------------------------------------------------------*\ 787/*---------------------------------------------------------------------------*\
884 788
885 FUNCTION....: codec2_decode_1600 789 FUNCTION....: codec2_decode_1600
@@ -890,91 +794,89 @@ void codec2_encode_1600(struct CODEC2 *c2, unsigned char * bits, short speech[])
890 794
891\*---------------------------------------------------------------------------*/ 795\*---------------------------------------------------------------------------*/
892 796
893void codec2_decode_1600(struct CODEC2 *c2, short speech[], const unsigned char * bits) 797void codec2_decode_1600(struct CODEC2 *c2, short speech[],
894{ 798 const unsigned char *bits) {
895 MODEL model[4]; 799 MODEL model[4];
896 int lsp_indexes[LPC_ORD]; 800 int lsp_indexes[LPC_ORD];
897 float lsps[4][LPC_ORD]; 801 float lsps[4][LPC_ORD];
898 int Wo_index, e_index; 802 int Wo_index, e_index;
899 float e[4]; 803 float e[4];
900 float snr; 804 float snr;
901 float ak[4][LPC_ORD+1]; 805 float ak[4][LPC_ORD + 1];
902 int i,j; 806 int i, j;
903 unsigned int nbit = 0; 807 unsigned int nbit = 0;
904 float weight; 808 float weight;
905 COMP Aw[FFT_ENC]; 809 COMP Aw[FFT_ENC];
906 810
907 assert(c2 != NULL); 811 assert(c2 != NULL);
908 812
909 /* only need to zero these out due to (unused) snr calculation */ 813 /* only need to zero these out due to (unused) snr calculation */
910 814
911 for(i=0; i<4; i++) 815 for (i = 0; i < 4; i++)
912 for(j=1; j<=MAX_AMP; j++) 816 for (j = 1; j <= MAX_AMP; j++) model[i].A[j] = 0.0;
913 model[i].A[j] = 0.0; 817
914 818 /* unpack bits from channel ------------------------------------*/
915 /* unpack bits from channel ------------------------------------*/ 819
916 820 /* this will partially fill the model params for the 4 x 10ms
917 /* this will partially fill the model params for the 4 x 10ms 821 frames */
918 frames */ 822
919 823 model[0].voiced = unpack(bits, &nbit, 1);
920 model[0].voiced = unpack(bits, &nbit, 1); 824
921 825 model[1].voiced = unpack(bits, &nbit, 1);
922 model[1].voiced = unpack(bits, &nbit, 1); 826 Wo_index = unpack(bits, &nbit, WO_BITS);
923 Wo_index = unpack(bits, &nbit, WO_BITS); 827 model[1].Wo = decode_Wo(&c2->c2const, Wo_index, WO_BITS);
924 model[1].Wo = decode_Wo(&c2->c2const, Wo_index, WO_BITS); 828 model[1].L = PI / model[1].Wo;
925 model[1].L = PI/model[1].Wo; 829
926 830 e_index = unpack(bits, &nbit, E_BITS);
927 e_index = unpack(bits, &nbit, E_BITS); 831 e[1] = decode_energy(e_index, E_BITS);
928 e[1] = decode_energy(e_index, E_BITS); 832
929 833 model[2].voiced = unpack(bits, &nbit, 1);
930 model[2].voiced = unpack(bits, &nbit, 1); 834
931 835 model[3].voiced = unpack(bits, &nbit, 1);
932 model[3].voiced = unpack(bits, &nbit, 1); 836 Wo_index = unpack(bits, &nbit, WO_BITS);
933 Wo_index = unpack(bits, &nbit, WO_BITS); 837 model[3].Wo = decode_Wo(&c2->c2const, Wo_index, WO_BITS);
934 model[3].Wo = decode_Wo(&c2->c2const, Wo_index, WO_BITS); 838 model[3].L = PI / model[3].Wo;
935 model[3].L = PI/model[3].Wo; 839
936 840 e_index = unpack(bits, &nbit, E_BITS);
937 e_index = unpack(bits, &nbit, E_BITS); 841 e[3] = decode_energy(e_index, E_BITS);
938 e[3] = decode_energy(e_index, E_BITS); 842
939 843 for (i = 0; i < LSP_SCALAR_INDEXES; i++) {
940 for(i=0; i<LSP_SCALAR_INDEXES; i++) { 844 lsp_indexes[i] = unpack(bits, &nbit, lsp_bits(i));
941 lsp_indexes[i] = unpack(bits, &nbit, lsp_bits(i)); 845 }
942 } 846 decode_lsps_scalar(&lsps[3][0], lsp_indexes, LPC_ORD);
943 decode_lsps_scalar(&lsps[3][0], lsp_indexes, LPC_ORD); 847 check_lsp_order(&lsps[3][0], LPC_ORD);
944 check_lsp_order(&lsps[3][0], LPC_ORD); 848 bw_expand_lsps(&lsps[3][0], LPC_ORD, 50.0, 100.0);
945 bw_expand_lsps(&lsps[3][0], LPC_ORD, 50.0, 100.0); 849
946 850 /* interpolate ------------------------------------------------*/
947 /* interpolate ------------------------------------------------*/ 851
948 852 /* Wo and energy are sampled every 20ms, so we interpolate just 1
949 /* Wo and energy are sampled every 20ms, so we interpolate just 1 853 10ms frame between 20ms samples */
950 10ms frame between 20ms samples */ 854
951 855 interp_Wo(&model[0], &c2->prev_model_dec, &model[1], c2->c2const.Wo_min);
952 interp_Wo(&model[0], &c2->prev_model_dec, &model[1], c2->c2const.Wo_min); 856 e[0] = interp_energy(c2->prev_e_dec, e[1]);
953 e[0] = interp_energy(c2->prev_e_dec, e[1]); 857 interp_Wo(&model[2], &model[1], &model[3], c2->c2const.Wo_min);
954 interp_Wo(&model[2], &model[1], &model[3], c2->c2const.Wo_min); 858 e[2] = interp_energy(e[1], e[3]);
955 e[2] = interp_energy(e[1], e[3]); 859
956 860 /* LSPs are sampled every 40ms so we interpolate the 3 frames in
957 /* LSPs are sampled every 40ms so we interpolate the 3 frames in 861 between, then recover spectral amplitudes */
958 between, then recover spectral amplitudes */ 862
959 863 for (i = 0, weight = 0.25; i < 3; i++, weight += 0.25) {
960 for(i=0, weight=0.25; i<3; i++, weight += 0.25) { 864 interpolate_lsp_ver2(&lsps[i][0], c2->prev_lsps_dec, &lsps[3][0], weight,
961 interpolate_lsp_ver2(&lsps[i][0], c2->prev_lsps_dec, &lsps[3][0], weight, LPC_ORD); 865 LPC_ORD);
962 } 866 }
963 for(i=0; i<4; i++) { 867 for (i = 0; i < 4; i++) {
964 lsp_to_lpc(&lsps[i][0], &ak[i][0], LPC_ORD); 868 lsp_to_lpc(&lsps[i][0], &ak[i][0], LPC_ORD);
965 aks_to_M2(c2->fftr_fwd_cfg, &ak[i][0], LPC_ORD, &model[i], e[i], &snr, 0, 0, 869 aks_to_M2(c2->fftr_fwd_cfg, &ak[i][0], LPC_ORD, &model[i], e[i], &snr, 0, 0,
966 c2->lpc_pf, c2->bass_boost, c2->beta, c2->gamma, Aw); 870 c2->lpc_pf, c2->bass_boost, c2->beta, c2->gamma, Aw);
967 apply_lpc_correction(&model[i]); 871 apply_lpc_correction(&model[i]);
968 synthesise_one_frame(c2, &speech[c2->n_samp*i], &model[i], Aw, 1.0); 872 synthesise_one_frame(c2, &speech[c2->n_samp * i], &model[i], Aw, 1.0);
969 } 873 }
970 874
971 /* update memories for next frame ----------------------------*/ 875 /* update memories for next frame ----------------------------*/
972 876
973 c2->prev_model_dec = model[3]; 877 c2->prev_model_dec = model[3];
974 c2->prev_e_dec = e[3]; 878 c2->prev_e_dec = e[3];
975 for(i=0; i<LPC_ORD; i++) 879 for (i = 0; i < LPC_ORD; i++) c2->prev_lsps_dec[i] = lsps[3][i];
976 c2->prev_lsps_dec[i] = lsps[3][i];
977
978} 880}
979 881
980/*---------------------------------------------------------------------------*\ 882/*---------------------------------------------------------------------------*\
@@ -1004,60 +906,59 @@ void codec2_decode_1600(struct CODEC2 *c2, short speech[], const unsigned char *
1004 906
1005\*---------------------------------------------------------------------------*/ 907\*---------------------------------------------------------------------------*/
1006 908
1007void codec2_encode_1400(struct CODEC2 *c2, unsigned char * bits, short speech[]) 909void codec2_encode_1400(struct CODEC2 *c2, unsigned char *bits,
1008{ 910 short speech[]) {
1009 MODEL model; 911 MODEL model;
1010 float lsps[LPC_ORD]; 912 float lsps[LPC_ORD];
1011 float ak[LPC_ORD+1]; 913 float ak[LPC_ORD + 1];
1012 float e; 914 float e;
1013 int lsp_indexes[LPC_ORD]; 915 int lsp_indexes[LPC_ORD];
1014 int WoE_index; 916 int WoE_index;
1015 int i; 917 int i;
1016 unsigned int nbit = 0; 918 unsigned int nbit = 0;
1017 919
1018 assert(c2 != NULL); 920 assert(c2 != NULL);
1019 921
1020 memset(bits, '\0', ((codec2_bits_per_frame(c2) + 7) / 8)); 922 memset(bits, '\0', ((codec2_bits_per_frame(c2) + 7) / 8));
1021 923
1022 /* frame 1: - voicing ---------------------------------------------*/ 924 /* frame 1: - voicing ---------------------------------------------*/
1023 925
1024 analyse_one_frame(c2, &model, speech); 926 analyse_one_frame(c2, &model, speech);
1025 pack(bits, &nbit, model.voiced, 1); 927 pack(bits, &nbit, model.voiced, 1);
1026 928
1027 /* frame 2: - voicing, joint Wo & E -------------------------------*/ 929 /* frame 2: - voicing, joint Wo & E -------------------------------*/
1028 930
1029 analyse_one_frame(c2, &model, &speech[c2->n_samp]); 931 analyse_one_frame(c2, &model, &speech[c2->n_samp]);
1030 pack(bits, &nbit, model.voiced, 1); 932 pack(bits, &nbit, model.voiced, 1);
1031 933
1032 /* need to run this just to get LPC energy */ 934 /* need to run this just to get LPC energy */
1033 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD); 935 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD);
1034 936
1035 WoE_index = encode_WoE(&model, e, c2->xq_enc); 937 WoE_index = encode_WoE(&model, e, c2->xq_enc);
1036 pack(bits, &nbit, WoE_index, WO_E_BITS); 938 pack(bits, &nbit, WoE_index, WO_E_BITS);
1037 939
1038 /* frame 3: - voicing ---------------------------------------------*/ 940 /* frame 3: - voicing ---------------------------------------------*/
1039 941
1040 analyse_one_frame(c2, &model, &speech[2*c2->n_samp]); 942 analyse_one_frame(c2, &model, &speech[2 * c2->n_samp]);
1041 pack(bits, &nbit, model.voiced, 1); 943 pack(bits, &nbit, model.voiced, 1);
1042 944
1043 /* frame 4: - voicing, joint Wo & E, scalar LSPs ------------------*/ 945 /* frame 4: - voicing, joint Wo & E, scalar LSPs ------------------*/
1044 946
1045 analyse_one_frame(c2, &model, &speech[3*c2->n_samp]); 947 analyse_one_frame(c2, &model, &speech[3 * c2->n_samp]);
1046 pack(bits, &nbit, model.voiced, 1); 948 pack(bits, &nbit, model.voiced, 1);
1047 949
1048 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD); 950 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD);
1049 WoE_index = encode_WoE(&model, e, c2->xq_enc); 951 WoE_index = encode_WoE(&model, e, c2->xq_enc);
1050 pack(bits, &nbit, WoE_index, WO_E_BITS); 952 pack(bits, &nbit, WoE_index, WO_E_BITS);
1051 953
1052 encode_lsps_scalar(lsp_indexes, lsps, LPC_ORD); 954 encode_lsps_scalar(lsp_indexes, lsps, LPC_ORD);
1053 for(i=0; i<LSP_SCALAR_INDEXES; i++) { 955 for (i = 0; i < LSP_SCALAR_INDEXES; i++) {
1054 pack(bits, &nbit, lsp_indexes[i], lsp_bits(i)); 956 pack(bits, &nbit, lsp_indexes[i], lsp_bits(i));
1055 } 957 }
1056 958
1057 assert(nbit == (unsigned)codec2_bits_per_frame(c2)); 959 assert(nbit == (unsigned)codec2_bits_per_frame(c2));
1058} 960}
1059 961
1060
1061/*---------------------------------------------------------------------------*\ 962/*---------------------------------------------------------------------------*\
1062 963
1063 FUNCTION....: codec2_decode_1400 964 FUNCTION....: codec2_decode_1400
@@ -1068,83 +969,81 @@ void codec2_encode_1400(struct CODEC2 *c2, unsigned char * bits, short speech[])
1068 969
1069\*---------------------------------------------------------------------------*/ 970\*---------------------------------------------------------------------------*/
1070 971
1071void codec2_decode_1400(struct CODEC2 *c2, short speech[], const unsigned char * bits) 972void codec2_decode_1400(struct CODEC2 *c2, short speech[],
1072{ 973 const unsigned char *bits) {
1073 MODEL model[4]; 974 MODEL model[4];
1074 int lsp_indexes[LPC_ORD]; 975 int lsp_indexes[LPC_ORD];
1075 float lsps[4][LPC_ORD]; 976 float lsps[4][LPC_ORD];
1076 int WoE_index; 977 int WoE_index;
1077 float e[4]; 978 float e[4];
1078 float snr; 979 float snr;
1079 float ak[4][LPC_ORD+1]; 980 float ak[4][LPC_ORD + 1];
1080 int i,j; 981 int i, j;
1081 unsigned int nbit = 0; 982 unsigned int nbit = 0;
1082 float weight; 983 float weight;
1083 COMP Aw[FFT_ENC]; 984 COMP Aw[FFT_ENC];
1084 985
1085 assert(c2 != NULL); 986 assert(c2 != NULL);
1086 987
1087 /* only need to zero these out due to (unused) snr calculation */ 988 /* only need to zero these out due to (unused) snr calculation */
1088 989
1089 for(i=0; i<4; i++) 990 for (i = 0; i < 4; i++)
1090 for(j=1; j<=MAX_AMP; j++) 991 for (j = 1; j <= MAX_AMP; j++) model[i].A[j] = 0.0;
1091 model[i].A[j] = 0.0; 992
1092 993 /* unpack bits from channel ------------------------------------*/
1093 /* unpack bits from channel ------------------------------------*/ 994
1094 995 /* this will partially fill the model params for the 4 x 10ms
1095 /* this will partially fill the model params for the 4 x 10ms 996 frames */
1096 frames */ 997
1097 998 model[0].voiced = unpack(bits, &nbit, 1);
1098 model[0].voiced = unpack(bits, &nbit, 1); 999
1099 1000 model[1].voiced = unpack(bits, &nbit, 1);
1100 model[1].voiced = unpack(bits, &nbit, 1); 1001 WoE_index = unpack(bits, &nbit, WO_E_BITS);
1101 WoE_index = unpack(bits, &nbit, WO_E_BITS); 1002 decode_WoE(&c2->c2const, &model[1], &e[1], c2->xq_dec, WoE_index);
1102 decode_WoE(&c2->c2const, &model[1], &e[1], c2->xq_dec, WoE_index); 1003
1103 1004 model[2].voiced = unpack(bits, &nbit, 1);
1104 model[2].voiced = unpack(bits, &nbit, 1); 1005
1105 1006 model[3].voiced = unpack(bits, &nbit, 1);
1106 model[3].voiced = unpack(bits, &nbit, 1); 1007 WoE_index = unpack(bits, &nbit, WO_E_BITS);
1107 WoE_index = unpack(bits, &nbit, WO_E_BITS); 1008 decode_WoE(&c2->c2const, &model[3], &e[3], c2->xq_dec, WoE_index);
1108 decode_WoE(&c2->c2const, &model[3], &e[3], c2->xq_dec, WoE_index); 1009
1109 1010 for (i = 0; i < LSP_SCALAR_INDEXES; i++) {
1110 for(i=0; i<LSP_SCALAR_INDEXES; i++) { 1011 lsp_indexes[i] = unpack(bits, &nbit, lsp_bits(i));
1111 lsp_indexes[i] = unpack(bits, &nbit, lsp_bits(i)); 1012 }
1112 } 1013 decode_lsps_scalar(&lsps[3][0], lsp_indexes, LPC_ORD);
1113 decode_lsps_scalar(&lsps[3][0], lsp_indexes, LPC_ORD); 1014 check_lsp_order(&lsps[3][0], LPC_ORD);
1114 check_lsp_order(&lsps[3][0], LPC_ORD); 1015 bw_expand_lsps(&lsps[3][0], LPC_ORD, 50.0, 100.0);
1115 bw_expand_lsps(&lsps[3][0], LPC_ORD, 50.0, 100.0); 1016
1116 1017 /* interpolate ------------------------------------------------*/
1117 /* interpolate ------------------------------------------------*/ 1018
1118 1019 /* Wo and energy are sampled every 20ms, so we interpolate just 1
1119 /* Wo and energy are sampled every 20ms, so we interpolate just 1 1020 10ms frame between 20ms samples */
1120 10ms frame between 20ms samples */ 1021
1121 1022 interp_Wo(&model[0], &c2->prev_model_dec, &model[1], c2->c2const.Wo_min);
1122 interp_Wo(&model[0], &c2->prev_model_dec, &model[1], c2->c2const.Wo_min); 1023 e[0] = interp_energy(c2->prev_e_dec, e[1]);
1123 e[0] = interp_energy(c2->prev_e_dec, e[1]); 1024 interp_Wo(&model[2], &model[1], &model[3], c2->c2const.Wo_min);
1124 interp_Wo(&model[2], &model[1], &model[3], c2->c2const.Wo_min); 1025 e[2] = interp_energy(e[1], e[3]);
1125 e[2] = interp_energy(e[1], e[3]); 1026
1126 1027 /* LSPs are sampled every 40ms so we interpolate the 3 frames in
1127 /* LSPs are sampled every 40ms so we interpolate the 3 frames in 1028 between, then recover spectral amplitudes */
1128 between, then recover spectral amplitudes */ 1029
1129 1030 for (i = 0, weight = 0.25; i < 3; i++, weight += 0.25) {
1130 for(i=0, weight=0.25; i<3; i++, weight += 0.25) { 1031 interpolate_lsp_ver2(&lsps[i][0], c2->prev_lsps_dec, &lsps[3][0], weight,
1131 interpolate_lsp_ver2(&lsps[i][0], c2->prev_lsps_dec, &lsps[3][0], weight, LPC_ORD); 1032 LPC_ORD);
1132 } 1033 }
1133 for(i=0; i<4; i++) { 1034 for (i = 0; i < 4; i++) {
1134 lsp_to_lpc(&lsps[i][0], &ak[i][0], LPC_ORD); 1035 lsp_to_lpc(&lsps[i][0], &ak[i][0], LPC_ORD);
1135 aks_to_M2(c2->fftr_fwd_cfg, &ak[i][0], LPC_ORD, &model[i], e[i], &snr, 0, 0, 1036 aks_to_M2(c2->fftr_fwd_cfg, &ak[i][0], LPC_ORD, &model[i], e[i], &snr, 0, 0,
1136 c2->lpc_pf, c2->bass_boost, c2->beta, c2->gamma, Aw); 1037 c2->lpc_pf, c2->bass_boost, c2->beta, c2->gamma, Aw);
1137 apply_lpc_correction(&model[i]); 1038 apply_lpc_correction(&model[i]);
1138 synthesise_one_frame(c2, &speech[c2->n_samp*i], &model[i], Aw, 1.0); 1039 synthesise_one_frame(c2, &speech[c2->n_samp * i], &model[i], Aw, 1.0);
1139 } 1040 }
1140 1041
1141 /* update memories for next frame ----------------------------*/ 1042 /* update memories for next frame ----------------------------*/
1142 1043
1143 c2->prev_model_dec = model[3]; 1044 c2->prev_model_dec = model[3];
1144 c2->prev_e_dec = e[3]; 1045 c2->prev_e_dec = e[3];
1145 for(i=0; i<LPC_ORD; i++) 1046 for (i = 0; i < LPC_ORD; i++) c2->prev_lsps_dec[i] = lsps[3][i];
1146 c2->prev_lsps_dec[i] = lsps[3][i];
1147
1148} 1047}
1149 1048
1150/*---------------------------------------------------------------------------*\ 1049/*---------------------------------------------------------------------------*\
@@ -1175,66 +1074,56 @@ void codec2_decode_1400(struct CODEC2 *c2, short speech[], const unsigned char *
1175 1074
1176\*---------------------------------------------------------------------------*/ 1075\*---------------------------------------------------------------------------*/
1177 1076
1178void codec2_encode_1300(struct CODEC2 *c2, unsigned char * bits, short speech[]) 1077void codec2_encode_1300(struct CODEC2 *c2, unsigned char *bits,
1179{ 1078 short speech[]) {
1180 MODEL model; 1079 MODEL model;
1181 float lsps[LPC_ORD]; 1080 float lsps[LPC_ORD];
1182 float ak[LPC_ORD+1]; 1081 float ak[LPC_ORD + 1];
1183 float e; 1082 float e;
1184 int lsp_indexes[LPC_ORD]; 1083 int lsp_indexes[LPC_ORD];
1185 int Wo_index, e_index; 1084 int Wo_index, e_index;
1186 int i; 1085 int i;
1187 unsigned int nbit = 0; 1086 unsigned int nbit = 0;
1188 //#ifdef PROFILE
1189 //unsigned int quant_start;
1190 //#endif
1191 1087
1192 assert(c2 != NULL); 1088 assert(c2 != NULL);
1193 1089
1194 memset(bits, '\0', ((codec2_bits_per_frame(c2) + 7) / 8)); 1090 memset(bits, '\0', ((codec2_bits_per_frame(c2) + 7) / 8));
1195 1091
1196 /* frame 1: - voicing ---------------------------------------------*/ 1092 /* frame 1: - voicing ---------------------------------------------*/
1197 1093
1198 analyse_one_frame(c2, &model, speech); 1094 analyse_one_frame(c2, &model, speech);
1199 pack_natural_or_gray(bits, &nbit, model.voiced, 1, c2->gray); 1095 pack_natural_or_gray(bits, &nbit, model.voiced, 1, c2->gray);
1200 1096
1201 /* frame 2: - voicing ---------------------------------------------*/ 1097 /* frame 2: - voicing ---------------------------------------------*/
1202 1098
1203 analyse_one_frame(c2, &model, &speech[c2->n_samp]); 1099 analyse_one_frame(c2, &model, &speech[c2->n_samp]);
1204 pack_natural_or_gray(bits, &nbit, model.voiced, 1, c2->gray); 1100 pack_natural_or_gray(bits, &nbit, model.voiced, 1, c2->gray);
1205 1101
1206 /* frame 3: - voicing ---------------------------------------------*/ 1102 /* frame 3: - voicing ---------------------------------------------*/
1207 1103
1208 analyse_one_frame(c2, &model, &speech[2*c2->n_samp]); 1104 analyse_one_frame(c2, &model, &speech[2 * c2->n_samp]);
1209 pack_natural_or_gray(bits, &nbit, model.voiced, 1, c2->gray); 1105 pack_natural_or_gray(bits, &nbit, model.voiced, 1, c2->gray);
1210 1106
1211 /* frame 4: - voicing, scalar Wo & E, scalar LSPs ------------------*/ 1107 /* frame 4: - voicing, scalar Wo & E, scalar LSPs ------------------*/
1212 1108
1213 analyse_one_frame(c2, &model, &speech[3*c2->n_samp]); 1109 analyse_one_frame(c2, &model, &speech[3 * c2->n_samp]);
1214 pack_natural_or_gray(bits, &nbit, model.voiced, 1, c2->gray); 1110 pack_natural_or_gray(bits, &nbit, model.voiced, 1, c2->gray);
1215 1111
1216 Wo_index = encode_Wo(&c2->c2const, model.Wo, WO_BITS); 1112 Wo_index = encode_Wo(&c2->c2const, model.Wo, WO_BITS);
1217 pack_natural_or_gray(bits, &nbit, Wo_index, WO_BITS, c2->gray); 1113 pack_natural_or_gray(bits, &nbit, Wo_index, WO_BITS, c2->gray);
1218 1114
1219 //#ifdef PROFILE 1115 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD);
1220 //quant_start = machdep_profile_sample(); 1116 e_index = encode_energy(e, E_BITS);
1221 //#endif 1117 pack_natural_or_gray(bits, &nbit, e_index, E_BITS, c2->gray);
1222 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD);
1223 e_index = encode_energy(e, E_BITS);
1224 pack_natural_or_gray(bits, &nbit, e_index, E_BITS, c2->gray);
1225 1118
1226 encode_lsps_scalar(lsp_indexes, lsps, LPC_ORD); 1119 encode_lsps_scalar(lsp_indexes, lsps, LPC_ORD);
1227 for(i=0; i<LSP_SCALAR_INDEXES; i++) { 1120 for (i = 0; i < LSP_SCALAR_INDEXES; i++) {
1228 pack_natural_or_gray(bits, &nbit, lsp_indexes[i], lsp_bits(i), c2->gray); 1121 pack_natural_or_gray(bits, &nbit, lsp_indexes[i], lsp_bits(i), c2->gray);
1229 } 1122 }
1230 //#ifdef PROFILE
1231 //machdep_profile_sample_and_log(quant_start, " quant/packing");
1232 //#endif
1233 1123
1234 assert(nbit == (unsigned)codec2_bits_per_frame(c2)); 1124 assert(nbit == (unsigned)codec2_bits_per_frame(c2));
1235} 1125}
1236 1126
1237
1238/*---------------------------------------------------------------------------*\ 1127/*---------------------------------------------------------------------------*\
1239 1128
1240 FUNCTION....: codec2_decode_1300 1129 FUNCTION....: codec2_decode_1300
@@ -1244,118 +1133,106 @@ void codec2_encode_1300(struct CODEC2 *c2, unsigned char * bits, short speech[])
1244 Decodes frames of 52 bits into 320 samples (40ms) of speech. 1133 Decodes frames of 52 bits into 320 samples (40ms) of speech.
1245 1134
1246\*---------------------------------------------------------------------------*/ 1135\*---------------------------------------------------------------------------*/
1247static int frames;
1248void codec2_decode_1300(struct CODEC2 *c2, short speech[], const unsigned char * bits, float ber_est)
1249{
1250 MODEL model[4];
1251 int lsp_indexes[LPC_ORD];
1252 float lsps[4][LPC_ORD];
1253 int Wo_index, e_index;
1254 float e[4];
1255 float snr;
1256 float ak[4][LPC_ORD+1];
1257 int i,j;
1258 unsigned int nbit = 0;
1259 float weight;
1260 COMP Aw[FFT_ENC];
1261 //PROFILE_VAR(recover_start);
1262
1263 assert(c2 != NULL);
1264 frames+= 4;
1265 /* only need to zero these out due to (unused) snr calculation */
1266
1267 for(i=0; i<4; i++)
1268 for(j=1; j<=MAX_AMP; j++)
1269 model[i].A[j] = 0.0;
1270
1271 /* unpack bits from channel ------------------------------------*/
1272
1273 /* this will partially fill the model params for the 4 x 10ms
1274 frames */
1275
1276 model[0].voiced = unpack_natural_or_gray(bits, &nbit, 1, c2->gray);
1277 model[1].voiced = unpack_natural_or_gray(bits, &nbit, 1, c2->gray);
1278 model[2].voiced = unpack_natural_or_gray(bits, &nbit, 1, c2->gray);
1279 model[3].voiced = unpack_natural_or_gray(bits, &nbit, 1, c2->gray);
1280
1281 Wo_index = unpack_natural_or_gray(bits, &nbit, WO_BITS, c2->gray);
1282 model[3].Wo = decode_Wo(&c2->c2const, Wo_index, WO_BITS);
1283 model[3].L = PI/model[3].Wo;
1284 1136
1285 e_index = unpack_natural_or_gray(bits, &nbit, E_BITS, c2->gray); 1137void codec2_decode_1300(struct CODEC2 *c2, short speech[],
1286 e[3] = decode_energy(e_index, E_BITS); 1138 const unsigned char *bits, float ber_est) {
1287 //fprintf(stderr, "%d %f\n", e_index, e[3]); 1139 MODEL model[4];
1140 int lsp_indexes[LPC_ORD];
1141 float lsps[4][LPC_ORD];
1142 int Wo_index, e_index;
1143 float e[4];
1144 float snr;
1145 float ak[4][LPC_ORD + 1];
1146 int i, j;
1147 unsigned int nbit = 0;
1148 float weight;
1149 COMP Aw[FFT_ENC];
1150
1151 assert(c2 != NULL);
1152
1153 /* only need to zero these out due to (unused) snr calculation */
1154
1155 for (i = 0; i < 4; i++)
1156 for (j = 1; j <= MAX_AMP; j++) model[i].A[j] = 0.0;
1157
1158 /* unpack bits from channel ------------------------------------*/
1159
1160 /* this will partially fill the model params for the 4 x 10ms
1161 frames */
1162
1163 model[0].voiced = unpack_natural_or_gray(bits, &nbit, 1, c2->gray);
1164 model[1].voiced = unpack_natural_or_gray(bits, &nbit, 1, c2->gray);
1165 model[2].voiced = unpack_natural_or_gray(bits, &nbit, 1, c2->gray);
1166 model[3].voiced = unpack_natural_or_gray(bits, &nbit, 1, c2->gray);
1167
1168 Wo_index = unpack_natural_or_gray(bits, &nbit, WO_BITS, c2->gray);
1169 model[3].Wo = decode_Wo(&c2->c2const, Wo_index, WO_BITS);
1170 model[3].L = PI / model[3].Wo;
1171
1172 e_index = unpack_natural_or_gray(bits, &nbit, E_BITS, c2->gray);
1173 e[3] = decode_energy(e_index, E_BITS);
1174
1175 for (i = 0; i < LSP_SCALAR_INDEXES; i++) {
1176 lsp_indexes[i] = unpack_natural_or_gray(bits, &nbit, lsp_bits(i), c2->gray);
1177 }
1178 decode_lsps_scalar(&lsps[3][0], lsp_indexes, LPC_ORD);
1179 check_lsp_order(&lsps[3][0], LPC_ORD);
1180 bw_expand_lsps(&lsps[3][0], LPC_ORD, 50.0, 100.0);
1181
1182 if (ber_est > 0.15) {
1183 model[0].voiced = model[1].voiced = model[2].voiced = model[3].voiced = 0;
1184 e[3] = decode_energy(10, E_BITS);
1185 bw_expand_lsps(&lsps[3][0], LPC_ORD, 200.0, 200.0);
1186 // fprintf(stderr, "soft mute\n");
1187 }
1188
1189 /* interpolate ------------------------------------------------*/
1190
1191 /* Wo, energy, and LSPs are sampled every 40ms so we interpolate
1192 the 3 frames in between */
1193
1194 for (i = 0, weight = 0.25; i < 3; i++, weight += 0.25) {
1195 interpolate_lsp_ver2(&lsps[i][0], c2->prev_lsps_dec, &lsps[3][0], weight,
1196 LPC_ORD);
1197 interp_Wo2(&model[i], &c2->prev_model_dec, &model[3], weight,
1198 c2->c2const.Wo_min);
1199 e[i] = interp_energy2(c2->prev_e_dec, e[3], weight);
1200 }
1201
1202 /* then recover spectral amplitudes */
1203
1204 for (i = 0; i < 4; i++) {
1205 lsp_to_lpc(&lsps[i][0], &ak[i][0], LPC_ORD);
1206 aks_to_M2(c2->fftr_fwd_cfg, &ak[i][0], LPC_ORD, &model[i], e[i], &snr, 0, 0,
1207 c2->lpc_pf, c2->bass_boost, c2->beta, c2->gamma, Aw);
1208 apply_lpc_correction(&model[i]);
1209 synthesise_one_frame(c2, &speech[c2->n_samp * i], &model[i], Aw, 1.0);
1210
1211 /* dump parameters for deep learning experiments */
1288 1212
1289 for(i=0; i<LSP_SCALAR_INDEXES; i++) { 1213 if (c2->fmlfeat != NULL) {
1290 lsp_indexes[i] = unpack_natural_or_gray(bits, &nbit, lsp_bits(i), c2->gray); 1214 /* 10 LSPs - energy - Wo - voicing flag - 10 LPCs */
1291 } 1215 fwrite(&lsps[i][0], LPC_ORD, sizeof(float), c2->fmlfeat);
1292 decode_lsps_scalar(&lsps[3][0], lsp_indexes, LPC_ORD); 1216 fwrite(&e[i], 1, sizeof(float), c2->fmlfeat);
1293 check_lsp_order(&lsps[3][0], LPC_ORD); 1217 fwrite(&model[i].Wo, 1, sizeof(float), c2->fmlfeat);
1294 bw_expand_lsps(&lsps[3][0], LPC_ORD, 50.0, 100.0); 1218 float voiced_float = model[i].voiced;
1295 1219 fwrite(&voiced_float, 1, sizeof(float), c2->fmlfeat);
1296 if (ber_est > 0.15) { 1220 fwrite(&ak[i][1], LPC_ORD, sizeof(float), c2->fmlfeat);
1297 model[0].voiced = model[1].voiced = model[2].voiced = model[3].voiced = 0; 1221 }
1298 e[3] = decode_energy(10, E_BITS); 1222 }
1299 bw_expand_lsps(&lsps[3][0], LPC_ORD, 200.0, 200.0); 1223
1300 //fprintf(stderr, "soft mute\n"); 1224#ifdef DUMP
1301 } 1225 dump_lsp_(&lsps[3][0]);
1302 1226 dump_ak_(&ak[3][0], LPC_ORD);
1303 /* interpolate ------------------------------------------------*/ 1227#endif
1304
1305 /* Wo, energy, and LSPs are sampled every 40ms so we interpolate
1306 the 3 frames in between */
1307
1308 //PROFILE_SAMPLE(recover_start);
1309 for(i=0, weight=0.25; i<3; i++, weight += 0.25) {
1310 interpolate_lsp_ver2(&lsps[i][0], c2->prev_lsps_dec, &lsps[3][0], weight, LPC_ORD);
1311 interp_Wo2(&model[i], &c2->prev_model_dec, &model[3], weight, c2->c2const.Wo_min);
1312 e[i] = interp_energy2(c2->prev_e_dec, e[3],weight);
1313 }
1314 1228
1315 /* then recover spectral amplitudes */ 1229 /* update memories for next frame ----------------------------*/
1316
1317 for(i=0; i<4; i++) {
1318 lsp_to_lpc(&lsps[i][0], &ak[i][0], LPC_ORD);
1319 aks_to_M2(c2->fftr_fwd_cfg, &ak[i][0], LPC_ORD, &model[i], e[i], &snr, 0, 0,
1320 c2->lpc_pf, c2->bass_boost, c2->beta, c2->gamma, Aw);
1321 apply_lpc_correction(&model[i]);
1322 synthesise_one_frame(c2, &speech[c2->n_samp*i], &model[i], Aw, 1.0);
1323
1324 /* dump parameters for deep learning experiments */
1325
1326 if (c2->fmlfeat != NULL) {
1327 /* 10 LSPs - energy - Wo - voicing flag - 10 LPCs */
1328 fwrite(&lsps[i][0], LPC_ORD, sizeof(float), c2->fmlfeat);
1329 fwrite(&e[i], 1, sizeof(float), c2->fmlfeat);
1330 fwrite(&model[i].Wo, 1, sizeof(float), c2->fmlfeat);
1331 float voiced_float = model[i].voiced;
1332 fwrite(&voiced_float, 1, sizeof(float), c2->fmlfeat);
1333 fwrite(&ak[i][1], LPC_ORD, sizeof(float), c2->fmlfeat);
1334 }
1335 }
1336 /*
1337 for(i=0; i<4; i++) {
1338 printf("%d Wo: %f L: %d v: %d\n", frames, model[i].Wo, model[i].L, model[i].voiced);
1339 }
1340 if (frames == 4*50)
1341 exit(0);
1342 */
1343 //PROFILE_SAMPLE_AND_LOG2(recover_start, " recover");
1344 #ifdef DUMP
1345 dump_lsp_(&lsps[3][0]);
1346 dump_ak_(&ak[3][0], LPC_ORD);
1347 #endif
1348
1349 /* update memories for next frame ----------------------------*/
1350
1351 c2->prev_model_dec = model[3];
1352 c2->prev_e_dec = e[3];
1353 for(i=0; i<LPC_ORD; i++)
1354 c2->prev_lsps_dec[i] = lsps[3][i];
1355 1230
1231 c2->prev_model_dec = model[3];
1232 c2->prev_e_dec = e[3];
1233 for (i = 0; i < LPC_ORD; i++) c2->prev_lsps_dec[i] = lsps[3][i];
1356} 1234}
1357 1235
1358
1359/*---------------------------------------------------------------------------*\ 1236/*---------------------------------------------------------------------------*\
1360 1237
1361 FUNCTION....: codec2_encode_1200 1238 FUNCTION....: codec2_encode_1200
@@ -1384,63 +1261,62 @@ void codec2_decode_1300(struct CODEC2 *c2, short speech[], const unsigned char *
1384 1261
1385\*---------------------------------------------------------------------------*/ 1262\*---------------------------------------------------------------------------*/
1386 1263
1387void codec2_encode_1200(struct CODEC2 *c2, unsigned char * bits, short speech[]) 1264void codec2_encode_1200(struct CODEC2 *c2, unsigned char *bits,
1388{ 1265 short speech[]) {
1389 MODEL model; 1266 MODEL model;
1390 float lsps[LPC_ORD]; 1267 float lsps[LPC_ORD];
1391 float lsps_[LPC_ORD]; 1268 float lsps_[LPC_ORD];
1392 float ak[LPC_ORD+1]; 1269 float ak[LPC_ORD + 1];
1393 float e; 1270 float e;
1394 int lsp_indexes[LPC_ORD]; 1271 int lsp_indexes[LPC_ORD];
1395 int WoE_index; 1272 int WoE_index;
1396 int i; 1273 int i;
1397 int spare = 0; 1274 int spare = 0;
1398 unsigned int nbit = 0; 1275 unsigned int nbit = 0;
1399 1276
1400 assert(c2 != NULL); 1277 assert(c2 != NULL);
1401 1278
1402 memset(bits, '\0', ((codec2_bits_per_frame(c2) + 7) / 8)); 1279 memset(bits, '\0', ((codec2_bits_per_frame(c2) + 7) / 8));
1403 1280
1404 /* frame 1: - voicing ---------------------------------------------*/ 1281 /* frame 1: - voicing ---------------------------------------------*/
1405 1282
1406 analyse_one_frame(c2, &model, speech); 1283 analyse_one_frame(c2, &model, speech);
1407 pack(bits, &nbit, model.voiced, 1); 1284 pack(bits, &nbit, model.voiced, 1);
1408 1285
1409 /* frame 2: - voicing, joint Wo & E -------------------------------*/ 1286 /* frame 2: - voicing, joint Wo & E -------------------------------*/
1410 1287
1411 analyse_one_frame(c2, &model, &speech[c2->n_samp]); 1288 analyse_one_frame(c2, &model, &speech[c2->n_samp]);
1412 pack(bits, &nbit, model.voiced, 1); 1289 pack(bits, &nbit, model.voiced, 1);
1413 1290
1414 /* need to run this just to get LPC energy */ 1291 /* need to run this just to get LPC energy */
1415 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD); 1292 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD);
1416 1293
1417 WoE_index = encode_WoE(&model, e, c2->xq_enc); 1294 WoE_index = encode_WoE(&model, e, c2->xq_enc);
1418 pack(bits, &nbit, WoE_index, WO_E_BITS); 1295 pack(bits, &nbit, WoE_index, WO_E_BITS);
1419 1296
1420 /* frame 3: - voicing ---------------------------------------------*/ 1297 /* frame 3: - voicing ---------------------------------------------*/
1421 1298
1422 analyse_one_frame(c2, &model, &speech[2*c2->n_samp]); 1299 analyse_one_frame(c2, &model, &speech[2 * c2->n_samp]);
1423 pack(bits, &nbit, model.voiced, 1); 1300 pack(bits, &nbit, model.voiced, 1);
1424 1301
1425 /* frame 4: - voicing, joint Wo & E, scalar LSPs ------------------*/ 1302 /* frame 4: - voicing, joint Wo & E, scalar LSPs ------------------*/
1426 1303
1427 analyse_one_frame(c2, &model, &speech[3*c2->n_samp]); 1304 analyse_one_frame(c2, &model, &speech[3 * c2->n_samp]);
1428 pack(bits, &nbit, model.voiced, 1); 1305 pack(bits, &nbit, model.voiced, 1);
1429 1306
1430 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD); 1307 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD);
1431 WoE_index = encode_WoE(&model, e, c2->xq_enc); 1308 WoE_index = encode_WoE(&model, e, c2->xq_enc);
1432 pack(bits, &nbit, WoE_index, WO_E_BITS); 1309 pack(bits, &nbit, WoE_index, WO_E_BITS);
1433 1310
1434 encode_lsps_vq(lsp_indexes, lsps, lsps_, LPC_ORD); 1311 encode_lsps_vq(lsp_indexes, lsps, lsps_, LPC_ORD);
1435 for(i=0; i<LSP_PRED_VQ_INDEXES; i++) { 1312 for (i = 0; i < LSP_PRED_VQ_INDEXES; i++) {
1436 pack(bits, &nbit, lsp_indexes[i], lsp_pred_vq_bits(i)); 1313 pack(bits, &nbit, lsp_indexes[i], lsp_pred_vq_bits(i));
1437 } 1314 }
1438 pack(bits, &nbit, spare, 1); 1315 pack(bits, &nbit, spare, 1);
1439 1316
1440 assert(nbit == (unsigned)codec2_bits_per_frame(c2)); 1317 assert(nbit == (unsigned)codec2_bits_per_frame(c2));
1441} 1318}
1442 1319
1443
1444/*---------------------------------------------------------------------------*\ 1320/*---------------------------------------------------------------------------*\
1445 1321
1446 FUNCTION....: codec2_decode_1200 1322 FUNCTION....: codec2_decode_1200
@@ -1451,494 +1327,83 @@ void codec2_encode_1200(struct CODEC2 *c2, unsigned char * bits, short speech[])
1451 1327
1452\*---------------------------------------------------------------------------*/ 1328\*---------------------------------------------------------------------------*/
1453 1329
1454void codec2_decode_1200(struct CODEC2 *c2, short speech[], const unsigned char * bits) 1330void codec2_decode_1200(struct CODEC2 *c2, short speech[],
1455{ 1331 const unsigned char *bits) {
1456 MODEL model[4]; 1332 MODEL model[4];
1457 int lsp_indexes[LPC_ORD]; 1333 int lsp_indexes[LPC_ORD];
1458 float lsps[4][LPC_ORD]; 1334 float lsps[4][LPC_ORD];
1459 int WoE_index; 1335 int WoE_index;
1460 float e[4]; 1336 float e[4];
1461 float snr; 1337 float snr;
1462 float ak[4][LPC_ORD+1]; 1338 float ak[4][LPC_ORD + 1];
1463 int i,j; 1339 int i, j;
1464 unsigned int nbit = 0; 1340 unsigned int nbit = 0;
1465 float weight; 1341 float weight;
1466 COMP Aw[FFT_ENC]; 1342 COMP Aw[FFT_ENC];
1467 1343
1468 assert(c2 != NULL); 1344 assert(c2 != NULL);
1469 1345
1470 /* only need to zero these out due to (unused) snr calculation */ 1346 /* only need to zero these out due to (unused) snr calculation */
1471 1347
1472 for(i=0; i<4; i++) 1348 for (i = 0; i < 4; i++)
1473 for(j=1; j<=MAX_AMP; j++) 1349 for (j = 1; j <= MAX_AMP; j++) model[i].A[j] = 0.0;
1474 model[i].A[j] = 0.0; 1350
1475 1351 /* unpack bits from channel ------------------------------------*/
1476 /* unpack bits from channel ------------------------------------*/ 1352
1477 1353 /* this will partially fill the model params for the 4 x 10ms
1478 /* this will partially fill the model params for the 4 x 10ms 1354 frames */
1479 frames */ 1355
1480 1356 model[0].voiced = unpack(bits, &nbit, 1);
1481 model[0].voiced = unpack(bits, &nbit, 1); 1357
1482 1358 model[1].voiced = unpack(bits, &nbit, 1);
1483 model[1].voiced = unpack(bits, &nbit, 1); 1359 WoE_index = unpack(bits, &nbit, WO_E_BITS);
1484 WoE_index = unpack(bits, &nbit, WO_E_BITS); 1360 decode_WoE(&c2->c2const, &model[1], &e[1], c2->xq_dec, WoE_index);
1485 decode_WoE(&c2->c2const, &model[1], &e[1], c2->xq_dec, WoE_index); 1361
1486 1362 model[2].voiced = unpack(bits, &nbit, 1);
1487 model[2].voiced = unpack(bits, &nbit, 1); 1363
1488 1364 model[3].voiced = unpack(bits, &nbit, 1);
1489 model[3].voiced = unpack(bits, &nbit, 1); 1365 WoE_index = unpack(bits, &nbit, WO_E_BITS);
1490 WoE_index = unpack(bits, &nbit, WO_E_BITS); 1366 decode_WoE(&c2->c2const, &model[3], &e[3], c2->xq_dec, WoE_index);
1491 decode_WoE(&c2->c2const, &model[3], &e[3], c2->xq_dec, WoE_index); 1367
1492 1368 for (i = 0; i < LSP_PRED_VQ_INDEXES; i++) {
1493 for(i=0; i<LSP_PRED_VQ_INDEXES; i++) { 1369 lsp_indexes[i] = unpack(bits, &nbit, lsp_pred_vq_bits(i));
1494 lsp_indexes[i] = unpack(bits, &nbit, lsp_pred_vq_bits(i)); 1370 }
1495 } 1371 decode_lsps_vq(lsp_indexes, &lsps[3][0], LPC_ORD, 0);
1496 decode_lsps_vq(lsp_indexes, &lsps[3][0], LPC_ORD , 0); 1372 check_lsp_order(&lsps[3][0], LPC_ORD);
1497 check_lsp_order(&lsps[3][0], LPC_ORD); 1373 bw_expand_lsps(&lsps[3][0], LPC_ORD, 50.0, 100.0);
1498 bw_expand_lsps(&lsps[3][0], LPC_ORD, 50.0, 100.0); 1374
1499 1375 /* interpolate ------------------------------------------------*/
1500 /* interpolate ------------------------------------------------*/ 1376
1501 1377 /* Wo and energy are sampled every 20ms, so we interpolate just 1
1502 /* Wo and energy are sampled every 20ms, so we interpolate just 1 1378 10ms frame between 20ms samples */
1503 10ms frame between 20ms samples */ 1379
1504 1380 interp_Wo(&model[0], &c2->prev_model_dec, &model[1], c2->c2const.Wo_min);
1505 interp_Wo(&model[0], &c2->prev_model_dec, &model[1], c2->c2const.Wo_min); 1381 e[0] = interp_energy(c2->prev_e_dec, e[1]);
1506 e[0] = interp_energy(c2->prev_e_dec, e[1]); 1382 interp_Wo(&model[2], &model[1], &model[3], c2->c2const.Wo_min);
1507 interp_Wo(&model[2], &model[1], &model[3], c2->c2const.Wo_min); 1383 e[2] = interp_energy(e[1], e[3]);
1508 e[2] = interp_energy(e[1], e[3]); 1384
1509 1385 /* LSPs are sampled every 40ms so we interpolate the 3 frames in
1510 /* LSPs are sampled every 40ms so we interpolate the 3 frames in 1386 between, then recover spectral amplitudes */
1511 between, then recover spectral amplitudes */ 1387
1512 1388 for (i = 0, weight = 0.25; i < 3; i++, weight += 0.25) {
1513 for(i=0, weight=0.25; i<3; i++, weight += 0.25) { 1389 interpolate_lsp_ver2(&lsps[i][0], c2->prev_lsps_dec, &lsps[3][0], weight,
1514 interpolate_lsp_ver2(&lsps[i][0], c2->prev_lsps_dec, &lsps[3][0], weight, LPC_ORD); 1390 LPC_ORD);
1515 } 1391 }
1516 for(i=0; i<4; i++) { 1392 for (i = 0; i < 4; i++) {
1517 lsp_to_lpc(&lsps[i][0], &ak[i][0], LPC_ORD); 1393 lsp_to_lpc(&lsps[i][0], &ak[i][0], LPC_ORD);
1518 aks_to_M2(c2->fftr_fwd_cfg, &ak[i][0], LPC_ORD, &model[i], e[i], &snr, 0, 0, 1394 aks_to_M2(c2->fftr_fwd_cfg, &ak[i][0], LPC_ORD, &model[i], e[i], &snr, 0, 0,
1519 c2->lpc_pf, c2->bass_boost, c2->beta, c2->gamma, Aw); 1395 c2->lpc_pf, c2->bass_boost, c2->beta, c2->gamma, Aw);
1520 apply_lpc_correction(&model[i]); 1396 apply_lpc_correction(&model[i]);
1521 synthesise_one_frame(c2, &speech[c2->n_samp*i], &model[i], Aw, 1.0); 1397 synthesise_one_frame(c2, &speech[c2->n_samp * i], &model[i], Aw, 1.0);
1522 } 1398 }
1523 1399
1524 /* update memories for next frame ----------------------------*/ 1400 /* update memories for next frame ----------------------------*/
1525 1401
1526 c2->prev_model_dec = model[3]; 1402 c2->prev_model_dec = model[3];
1527 c2->prev_e_dec = e[3]; 1403 c2->prev_e_dec = e[3];
1528 for(i=0; i<LPC_ORD; i++) 1404 for (i = 0; i < LPC_ORD; i++) c2->prev_lsps_dec[i] = lsps[3][i];
1529 c2->prev_lsps_dec[i] = lsps[3][i];
1530} 1405}
1531 1406
1532
1533/*---------------------------------------------------------------------------*\
1534
1535 FUNCTION....: codec2_encode_700
1536 AUTHOR......: David Rowe
1537 DATE CREATED: April 2015
1538
1539 Encodes 320 speech samples (40ms of speech) into 28 bits.
1540
1541 The codec2 algorithm actually operates internally on 10ms (80
1542 sample) frames, so we run the encoding algorithm four times:
1543
1544 frame 0: nothing
1545 frame 1: nothing
1546 frame 2: nothing
1547 frame 3: voicing bit, scalar Wo and E, 17 bit LSP MEL scalar, 2 spare
1548
1549 The bit allocation is:
1550
1551 Parameter frames 1-3 frame 4 Total
1552 -----------------------------------------------------------
1553 Harmonic magnitudes (LSPs) 0 17 17
1554 Energy 0 3 3
1555 log Wo 0 5 5
1556 Voicing 0 1 1
1557 spare 0 2 2
1558 TOTAL 0 28 28
1559
1560\*---------------------------------------------------------------------------*/
1561
1562void codec2_encode_700(struct CODEC2 *c2, unsigned char * bits, short speech[])
1563{
1564 MODEL model;
1565 float lsps[LPC_ORD_LOW];
1566 float mel[LPC_ORD_LOW];
1567 float ak[LPC_ORD_LOW+1];
1568 float e, f;
1569 int indexes[LPC_ORD_LOW];
1570 int Wo_index, e_index, i;
1571 unsigned int nbit = 0;
1572 float bpf_out[4*c2->n_samp];
1573 short bpf_speech[4*c2->n_samp];
1574 int spare = 0;
1575
1576 assert(c2 != NULL);
1577
1578 memset(bits, '\0', ((codec2_bits_per_frame(c2) + 7) / 8));
1579
1580 /* band pass filter */
1581
1582 for(i=0; i<BPF_N; i++)
1583 c2->bpf_buf[i] = c2->bpf_buf[4*c2->n_samp+i];
1584 for(i=0; i<4*c2->n_samp; i++)
1585 c2->bpf_buf[BPF_N+i] = speech[i];
1586 inverse_filter(&c2->bpf_buf[BPF_N], bpf, 4*c2->n_samp, bpf_out, BPF_N-1);
1587 for(i=0; i<4*c2->n_samp; i++)
1588 bpf_speech[i] = bpf_out[i];
1589
1590 /* frame 1 --------------------------------------------------------*/
1591
1592 analyse_one_frame(c2, &model, bpf_speech);
1593
1594 /* frame 2 --------------------------------------------------------*/
1595
1596 analyse_one_frame(c2, &model, &bpf_speech[c2->n_samp]);
1597
1598 /* frame 3 --------------------------------------------------------*/
1599
1600 analyse_one_frame(c2, &model, &bpf_speech[2*c2->n_samp]);
1601
1602 /* frame 4: - voicing, scalar Wo & E, scalar LSPs -----------------*/
1603
1604 analyse_one_frame(c2, &model, &bpf_speech[3*c2->n_samp]);
1605 pack(bits, &nbit, model.voiced, 1);
1606 Wo_index = encode_log_Wo(&c2->c2const, model.Wo, 5);
1607 pack_natural_or_gray(bits, &nbit, Wo_index, 5, c2->gray);
1608
1609 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD_LOW);
1610 e_index = encode_energy(e, 3);
1611 pack_natural_or_gray(bits, &nbit, e_index, 3, c2->gray);
1612
1613 for(i=0; i<LPC_ORD_LOW; i++) {
1614 f = (4000.0/PI)*lsps[i];
1615 mel[i] = floor(2595.0*log10(1.0 + f/700.0) + 0.5);
1616 }
1617 encode_mels_scalar(indexes, mel, LPC_ORD_LOW);
1618
1619 for(i=0; i<LPC_ORD_LOW; i++) {
1620 pack_natural_or_gray(bits, &nbit, indexes[i], mel_bits(i), c2->gray);
1621 }
1622
1623 pack_natural_or_gray(bits, &nbit, spare, 2, c2->gray);
1624
1625 assert(nbit == (unsigned)codec2_bits_per_frame(c2));
1626}
1627
1628
1629/*---------------------------------------------------------------------------*\
1630
1631 FUNCTION....: codec2_decode_700
1632 AUTHOR......: David Rowe
1633 DATE CREATED: April 2015
1634
1635 Decodes frames of 28 bits into 320 samples (40ms) of speech.
1636
1637\*---------------------------------------------------------------------------*/
1638
1639void codec2_decode_700(struct CODEC2 *c2, short speech[], const unsigned char * bits)
1640{
1641 MODEL model[4];
1642 int indexes[LPC_ORD_LOW];
1643 float mel[LPC_ORD_LOW];
1644 float lsps[4][LPC_ORD_LOW];
1645 int Wo_index, e_index;
1646 float e[4];
1647 float snr, f_;
1648 float ak[4][LPC_ORD_LOW+1];
1649 int i,j;
1650 unsigned int nbit = 0;
1651 float weight;
1652 COMP Aw[FFT_ENC];
1653
1654 assert(c2 != NULL);
1655
1656 /* only need to zero these out due to (unused) snr calculation */
1657
1658 for(i=0; i<4; i++)
1659 for(j=1; j<=MAX_AMP; j++)
1660 model[i].A[j] = 0.0;
1661
1662 /* unpack bits from channel ------------------------------------*/
1663
1664 model[3].voiced = unpack(bits, &nbit, 1);
1665 model[0].voiced = model[1].voiced = model[2].voiced = model[3].voiced;
1666
1667 Wo_index = unpack_natural_or_gray(bits, &nbit, 5, c2->gray);
1668 model[3].Wo = decode_log_Wo(&c2->c2const, Wo_index, 5);
1669 model[3].L = PI/model[3].Wo;
1670
1671 e_index = unpack_natural_or_gray(bits, &nbit, 3, c2->gray);
1672 e[3] = decode_energy(e_index, 3);
1673
1674 for(i=0; i<LPC_ORD_LOW; i++) {
1675 indexes[i] = unpack_natural_or_gray(bits, &nbit, mel_bits(i), c2->gray);
1676 }
1677
1678 decode_mels_scalar(mel, indexes, LPC_ORD_LOW);
1679 for(i=0; i<LPC_ORD_LOW; i++) {
1680 f_ = 700.0*( pow(10.0, (float)mel[i]/2595.0) - 1.0);
1681 lsps[3][i] = f_*(PI/4000.0);
1682 //printf("lsps[3][%d] %f\n", i, lsps[3][i]);
1683 }
1684
1685 check_lsp_order(&lsps[3][0], LPC_ORD_LOW);
1686 bw_expand_lsps(&lsps[3][0], LPC_ORD_LOW, 50.0, 100.0);
1687
1688 #ifdef MASK_NOT_FOR_NOW
1689 /* first pass at soft decn error masking, needs further work */
1690 /* If soft dec info available expand further for low power frames */
1691
1692 if (c2->softdec) {
1693 float e = 0.0;
1694 for(i=9; i<9+17; i++)
1695 e += c2->softdec[i]*c2->softdec[i];
1696 e /= 6.0;
1697 //fprintf(stderr, "e: %f\n", e);
1698 //if (e < 0.3)
1699 // bw_expand_lsps(&lsps[3][0], LPC_ORD_LOW, 150.0, 300.0);
1700 }
1701 #endif
1702
1703 /* interpolate ------------------------------------------------*/
1704
1705 /* LSPs, Wo, and energy are sampled every 40ms so we interpolate
1706 the 3 frames in between, then recover spectral amplitudes */
1707
1708 for(i=0, weight=0.25; i<3; i++, weight += 0.25) {
1709 interpolate_lsp_ver2(&lsps[i][0], c2->prev_lsps_dec, &lsps[3][0], weight, LPC_ORD_LOW);
1710 interp_Wo2(&model[i], &c2->prev_model_dec, &model[3], weight, c2->c2const.Wo_min);
1711 e[i] = interp_energy2(c2->prev_e_dec, e[3],weight);
1712 }
1713 for(i=0; i<4; i++) {
1714 lsp_to_lpc(&lsps[i][0], &ak[i][0], LPC_ORD_LOW);
1715 aks_to_M2(c2->fftr_fwd_cfg, &ak[i][0], LPC_ORD_LOW, &model[i], e[i], &snr, 0, 0,
1716 c2->lpc_pf, c2->bass_boost, c2->beta, c2->gamma, Aw);
1717 apply_lpc_correction(&model[i]);
1718 synthesise_one_frame(c2, &speech[c2->n_samp*i], &model[i], Aw, 1.0);
1719 }
1720
1721 #ifdef DUMP
1722 dump_lsp_(&lsps[3][0]);
1723 dump_ak_(&ak[3][0], LPC_ORD_LOW);
1724 dump_model(&model[3]);
1725 if (c2->softdec)
1726 dump_softdec(c2->softdec, nbit);
1727 #endif
1728
1729 /* update memories for next frame ----------------------------*/
1730
1731 c2->prev_model_dec = model[3];
1732 c2->prev_e_dec = e[3];
1733 for(i=0; i<LPC_ORD_LOW; i++)
1734 c2->prev_lsps_dec[i] = lsps[3][i];
1735}
1736
1737
1738/*---------------------------------------------------------------------------*\
1739
1740 FUNCTION....: codec2_encode_700b
1741 AUTHOR......: David Rowe
1742 DATE CREATED: August 2015
1743
1744 Version b of 700 bit/s codec. After some experiments over the air I
1745 wanted was unhappy with the rate 700 codec so spent a few weeks
1746 trying to improve the speech quality. This version uses a wider BPF
1747 and vector quantised mel-lsps.
1748
1749 Encodes 320 speech samples (40ms of speech) into 28 bits.
1750
1751 The codec2 algorithm actually operates internally on 10ms (80
1752 sample) frames, so we run the encoding algorithm four times:
1753
1754 frame 0: nothing
1755 frame 1: nothing
1756 frame 2: nothing
1757 frame 3: voicing bit, 5 bit scalar Wo and 3 bit E, 18 bit LSP MEL VQ,
1758 1 spare
1759
1760 The bit allocation is:
1761
1762 Parameter frames 1-3 frame 4 Total
1763 -----------------------------------------------------------
1764 Harmonic magnitudes (LSPs) 0 18 18
1765 Energy 0 3 3
1766 log Wo 0 5 5
1767 Voicing 0 1 1
1768 spare 0 1 1
1769 TOTAL 0 28 28
1770
1771\*---------------------------------------------------------------------------*/
1772
1773void codec2_encode_700b(struct CODEC2 *c2, unsigned char * bits, short speech[])
1774{
1775 MODEL model;
1776 float lsps[LPC_ORD_LOW];
1777 float mel[LPC_ORD_LOW];
1778 float mel_[LPC_ORD_LOW];
1779 float ak[LPC_ORD_LOW+1];
1780 float e, f;
1781 int indexes[3];
1782 int Wo_index, e_index, i;
1783 unsigned int nbit = 0;
1784 float bpf_out[4*c2->n_samp];
1785 short bpf_speech[4*c2->n_samp];
1786 int spare = 0;
1787
1788 assert(c2 != NULL);
1789
1790 memset(bits, '\0', ((codec2_bits_per_frame(c2) + 7) / 8));
1791
1792 /* band pass filter */
1793
1794 for(i=0; i<BPF_N; i++)
1795 c2->bpf_buf[i] = c2->bpf_buf[4*c2->n_samp+i];
1796 for(i=0; i<4*c2->n_samp; i++)
1797 c2->bpf_buf[BPF_N+i] = speech[i];
1798 inverse_filter(&c2->bpf_buf[BPF_N], bpfb, 4*c2->n_samp, bpf_out, BPF_N-1);
1799 for(i=0; i<4*c2->n_samp; i++)
1800 bpf_speech[i] = bpf_out[i];
1801
1802 /* frame 1 --------------------------------------------------------*/
1803
1804 analyse_one_frame(c2, &model, bpf_speech);
1805
1806 /* frame 2 --------------------------------------------------------*/
1807
1808 analyse_one_frame(c2, &model, &bpf_speech[c2->n_samp]);
1809
1810 /* frame 3 --------------------------------------------------------*/
1811
1812 analyse_one_frame(c2, &model, &bpf_speech[2*c2->n_samp]);
1813
1814 /* frame 4: - voicing, scalar Wo & E, VQ mel LSPs -----------------*/
1815
1816 analyse_one_frame(c2, &model, &bpf_speech[3*c2->n_samp]);
1817 pack(bits, &nbit, model.voiced, 1);
1818 Wo_index = encode_log_Wo(&c2->c2const, model.Wo, 5);
1819 pack_natural_or_gray(bits, &nbit, Wo_index, 5, c2->gray);
1820
1821 e = speech_to_uq_lsps(lsps, ak, c2->Sn, c2->w, c2->m_pitch, LPC_ORD_LOW);
1822 e_index = encode_energy(e, 3);
1823 pack_natural_or_gray(bits, &nbit, e_index, 3, c2->gray);
1824
1825 for(i=0; i<LPC_ORD_LOW; i++) {
1826 f = (4000.0/PI)*lsps[i];
1827 mel[i] = floor(2595.0*log10(1.0 + f/700.0) + 0.5);
1828 }
1829 lspmelvq_mbest_encode(indexes, mel, mel_, LPC_ORD_LOW, 5);
1830
1831 for(i=0; i<3; i++) {
1832 pack_natural_or_gray(bits, &nbit, indexes[i], lspmelvq_cb_bits(i), c2->gray);
1833 }
1834
1835 pack_natural_or_gray(bits, &nbit, spare, 1, c2->gray);
1836
1837 assert(nbit == (unsigned)codec2_bits_per_frame(c2));
1838}
1839
1840
1841/*---------------------------------------------------------------------------*\
1842
1843 FUNCTION....: codec2_decode_700b
1844 AUTHOR......: David Rowe
1845 DATE CREATED: August 2015
1846
1847 Decodes frames of 28 bits into 320 samples (40ms) of speech.
1848
1849\*---------------------------------------------------------------------------*/
1850
1851void codec2_decode_700b(struct CODEC2 *c2, short speech[], const unsigned char * bits)
1852{
1853 MODEL model[4];
1854 int indexes[3];
1855 float mel[LPC_ORD_LOW];
1856 float lsps[4][LPC_ORD_LOW];
1857 int Wo_index, e_index;
1858 float e[4];
1859 float snr, f_;
1860 float ak[4][LPC_ORD_LOW+1];
1861 int i,j;
1862 unsigned int nbit = 0;
1863 float weight;
1864 COMP Aw[FFT_ENC];
1865
1866 assert(c2 != NULL);
1867
1868 /* only need to zero these out due to (unused) snr calculation */
1869
1870 for(i=0; i<4; i++)
1871 for(j=1; j<=MAX_AMP; j++)
1872 model[i].A[j] = 0.0;
1873
1874 /* unpack bits from channel ------------------------------------*/
1875
1876 model[3].voiced = unpack(bits, &nbit, 1);
1877 model[0].voiced = model[1].voiced = model[2].voiced = model[3].voiced;
1878
1879 Wo_index = unpack_natural_or_gray(bits, &nbit, 5, c2->gray);
1880 model[3].Wo = decode_log_Wo(&c2->c2const, Wo_index, 5);
1881 model[3].L = PI/model[3].Wo;
1882
1883 e_index = unpack_natural_or_gray(bits, &nbit, 3, c2->gray);
1884 e[3] = decode_energy(e_index, 3);
1885
1886 for(i=0; i<3; i++) {
1887 indexes[i] = unpack_natural_or_gray(bits, &nbit, lspmelvq_cb_bits(i), c2->gray);
1888 }
1889
1890 lspmelvq_decode(indexes, mel, LPC_ORD_LOW);
1891
1892 #define MEL_ROUND 10
1893 for(i=1; i<LPC_ORD_LOW; i++) {
1894 if (mel[i] <= mel[i-1]+MEL_ROUND) {
1895 mel[i]+=MEL_ROUND/2;
1896 mel[i-1]-=MEL_ROUND/2;
1897 i = 1;
1898 }
1899 }
1900
1901 for(i=0; i<LPC_ORD_LOW; i++) {
1902 f_ = 700.0*( pow(10.0, (float)mel[i]/2595.0) - 1.0);
1903 lsps[3][i] = f_*(PI/4000.0);
1904 //printf("lsps[3][%d] %f\n", i, lsps[3][i]);
1905 }
1906
1907 /* interpolate ------------------------------------------------*/
1908
1909 /* LSPs, Wo, and energy are sampled every 40ms so we interpolate
1910 the 3 frames in between, then recover spectral amplitudes */
1911
1912 for(i=0, weight=0.25; i<3; i++, weight += 0.25) {
1913 interpolate_lsp_ver2(&lsps[i][0], c2->prev_lsps_dec, &lsps[3][0], weight, LPC_ORD_LOW);
1914 interp_Wo2(&model[i], &c2->prev_model_dec, &model[3], weight, c2->c2const.Wo_min);
1915 e[i] = interp_energy2(c2->prev_e_dec, e[3],weight);
1916 }
1917 for(i=0; i<4; i++) {
1918 lsp_to_lpc(&lsps[i][0], &ak[i][0], LPC_ORD_LOW);
1919 aks_to_M2(c2->fftr_fwd_cfg, &ak[i][0], LPC_ORD_LOW, &model[i], e[i], &snr, 0, 0,
1920 c2->lpc_pf, c2->bass_boost, c2->beta, c2->gamma, Aw);
1921 apply_lpc_correction(&model[i]);
1922 synthesise_one_frame(c2, &speech[c2->n_samp*i], &model[i], Aw, 1.0);
1923 }
1924
1925 #ifdef DUMP
1926 dump_lsp_(&lsps[3][0]);
1927 dump_ak_(&ak[3][0], LPC_ORD_LOW);
1928 dump_model(&model[3]);
1929 if (c2->softdec)
1930 dump_softdec(c2->softdec, nbit);
1931 #endif
1932
1933 /* update memories for next frame ----------------------------*/
1934
1935 c2->prev_model_dec = model[3];
1936 c2->prev_e_dec = e[3];
1937 for(i=0; i<LPC_ORD_LOW; i++)
1938 c2->prev_lsps_dec[i] = lsps[3][i];
1939}
1940
1941
1942/*---------------------------------------------------------------------------*\ 1407/*---------------------------------------------------------------------------*\
1943 1408
1944 FUNCTION....: codec2_encode_700c 1409 FUNCTION....: codec2_encode_700c
@@ -1955,7 +1420,7 @@ void codec2_decode_700b(struct CODEC2 *c2, short speech[], const unsigned char *
1955 frame 0: nothing 1420 frame 0: nothing
1956 frame 1: nothing 1421 frame 1: nothing
1957 frame 2: nothing 1422 frame 2: nothing
1958 frame 3: 18 bit 2 stage VQ (9 bits/stage), 4 bits energy, 1423 frame 3: 18 bit 2 stage VQ (9 bits/stage), 4 bits energy,
1959 6 bit scalar Wo/voicing. No spare bits. 1424 6 bit scalar Wo/voicing. No spare bits.
1960 1425
1961 Voicing is encoded using the 0 index of the Wo quantiser. 1426 Voicing is encoded using the 0 index of the Wo quantiser.
@@ -1971,52 +1436,54 @@ void codec2_decode_700b(struct CODEC2 *c2, short speech[], const unsigned char *
1971 1436
1972\*---------------------------------------------------------------------------*/ 1437\*---------------------------------------------------------------------------*/
1973 1438
1974void codec2_encode_700c(struct CODEC2 *c2, unsigned char * bits, short speech[]) 1439void codec2_encode_700c(struct CODEC2 *c2, unsigned char *bits,
1975{ 1440 short speech[]) {
1976 MODEL model; 1441 MODEL model;
1977 int indexes[4], i, M=4; 1442 int indexes[4], i, M = 4;
1978 unsigned int nbit = 0; 1443 unsigned int nbit = 0;
1979 1444
1980 assert(c2 != NULL); 1445 assert(c2 != NULL);
1981 1446
1982 memset(bits, '\0', ((codec2_bits_per_frame(c2) + 7) / 8)); 1447 memset(bits, '\0', ((codec2_bits_per_frame(c2) + 7) / 8));
1983 1448
1984 for(i=0; i<M; i++) { 1449 for (i = 0; i < M; i++) {
1985 analyse_one_frame(c2, &model, &speech[i*c2->n_samp]); 1450 analyse_one_frame(c2, &model, &speech[i * c2->n_samp]);
1986 } 1451 }
1987 1452
1988 int K = 20; 1453 int K = 20;
1989 float rate_K_vec[K], mean; 1454 float rate_K_vec[K], mean;
1990 float rate_K_vec_no_mean[K], rate_K_vec_no_mean_[K]; 1455 float rate_K_vec_no_mean[K], rate_K_vec_no_mean_[K];
1991 1456
1992 newamp1_model_to_indexes(&c2->c2const, 1457 newamp1_model_to_indexes(&c2->c2const, indexes, &model, rate_K_vec,
1993 indexes, 1458 c2->rate_K_sample_freqs_kHz, K, &mean,
1994 &model, 1459 rate_K_vec_no_mean, rate_K_vec_no_mean_, &c2->se,
1995 rate_K_vec, 1460 c2->eq, c2->eq_en);
1996 c2->rate_K_sample_freqs_kHz, 1461 c2->nse += K;
1997 K,
1998 &mean,
1999 rate_K_vec_no_mean,
2000 rate_K_vec_no_mean_, &c2->se, c2->eq, c2->eq_en);
2001 c2->nse += K;
2002 1462
2003#ifndef CORTEX_M4 1463#ifndef CORTEX_M4
2004 /* dump features for deep learning experiments */ 1464 /* dump features for deep learning experiments */
2005 if (c2->fmlfeat != NULL) { 1465 if (c2->fmlfeat != NULL) {
2006 fwrite(&mean, 1, sizeof(float), c2->fmlfeat); 1466 fwrite(&mean, 1, sizeof(float), c2->fmlfeat);
2007 fwrite(rate_K_vec_no_mean, K, sizeof(float), c2->fmlfeat); 1467 fwrite(rate_K_vec_no_mean, K, sizeof(float), c2->fmlfeat);
2008 fwrite(rate_K_vec_no_mean_, K, sizeof(float), c2->fmlfeat); 1468 fwrite(rate_K_vec_no_mean_, K, sizeof(float), c2->fmlfeat);
2009 } 1469 MODEL model_;
1470 memcpy(&model_, &model, sizeof(model));
1471 float rate_K_vec_[K];
1472 for (int k = 0; k < K; k++) rate_K_vec_[k] = rate_K_vec_no_mean_[k] + mean;
1473 resample_rate_L(&c2->c2const, &model_, rate_K_vec_,
1474 c2->rate_K_sample_freqs_kHz, K);
1475 fwrite(&model_.A, MAX_AMP, sizeof(float), c2->fmlfeat);
1476 }
1477 if (c2->fmlmodel != NULL) fwrite(&model, sizeof(MODEL), 1, c2->fmlmodel);
2010#endif 1478#endif
2011
2012 pack_natural_or_gray(bits, &nbit, indexes[0], 9, 0);
2013 pack_natural_or_gray(bits, &nbit, indexes[1], 9, 0);
2014 pack_natural_or_gray(bits, &nbit, indexes[2], 4, 0);
2015 pack_natural_or_gray(bits, &nbit, indexes[3], 6, 0);
2016 1479
2017 assert(nbit == (unsigned)codec2_bits_per_frame(c2)); 1480 pack_natural_or_gray(bits, &nbit, indexes[0], 9, 0);
2018} 1481 pack_natural_or_gray(bits, &nbit, indexes[1], 9, 0);
1482 pack_natural_or_gray(bits, &nbit, indexes[2], 4, 0);
1483 pack_natural_or_gray(bits, &nbit, indexes[3], 6, 0);
2019 1484
1485 assert(nbit == (unsigned)codec2_bits_per_frame(c2));
1486}
2020 1487
2021/*---------------------------------------------------------------------------*\ 1488/*---------------------------------------------------------------------------*\
2022 1489
@@ -2028,46 +1495,53 @@ void codec2_encode_700c(struct CODEC2 *c2, unsigned char * bits, short speech[])
2028 1495
2029\*---------------------------------------------------------------------------*/ 1496\*---------------------------------------------------------------------------*/
2030 1497
2031void codec2_decode_700c(struct CODEC2 *c2, short speech[], const unsigned char * bits) 1498void codec2_decode_700c(struct CODEC2 *c2, short speech[],
2032{ 1499 const unsigned char *bits) {
2033 MODEL model[4]; 1500 MODEL model[4];
2034 int indexes[4]; 1501 int indexes[4];
2035 int i; 1502 int i;
2036 unsigned int nbit = 0; 1503 unsigned int nbit = 0;
2037 1504
2038 assert(c2 != NULL); 1505 assert(c2 != NULL);
2039 1506
2040 /* unpack bits from channel ------------------------------------*/ 1507 /* unpack bits from channel ------------------------------------*/
2041 1508
2042 indexes[0] = unpack_natural_or_gray(bits, &nbit, 9, 0); 1509 indexes[0] = unpack_natural_or_gray(bits, &nbit, 9, 0);
2043 indexes[1] = unpack_natural_or_gray(bits, &nbit, 9, 0); 1510 indexes[1] = unpack_natural_or_gray(bits, &nbit, 9, 0);
2044 indexes[2] = unpack_natural_or_gray(bits, &nbit, 4, 0); 1511 indexes[2] = unpack_natural_or_gray(bits, &nbit, 4, 0);
2045 indexes[3] = unpack_natural_or_gray(bits, &nbit, 6, 0); 1512 indexes[3] = unpack_natural_or_gray(bits, &nbit, 6, 0);
2046 1513
2047 int M = 4; 1514 int M = 4;
2048 COMP HH[M][MAX_AMP+1]; 1515 COMP HH[M][MAX_AMP + 1];
2049 float interpolated_surface_[M][NEWAMP1_K]; 1516 float interpolated_surface_[M][NEWAMP1_K];
2050 1517
2051 newamp1_indexes_to_model(&c2->c2const, 1518 newamp1_indexes_to_model(
2052 model, 1519 &c2->c2const, model, (COMP *)HH, (float *)interpolated_surface_,
2053 (COMP*)HH, 1520 c2->prev_rate_K_vec_, &c2->Wo_left, &c2->voicing_left,
2054 (float*)interpolated_surface_, 1521 c2->rate_K_sample_freqs_kHz, NEWAMP1_K, c2->phase_fft_fwd_cfg,
2055 c2->prev_rate_K_vec_, 1522 c2->phase_fft_inv_cfg, indexes, c2->user_rate_K_vec_no_mean_,
2056 &c2->Wo_left, 1523 c2->post_filter_en);
2057 &c2->voicing_left, 1524
2058 c2->rate_K_sample_freqs_kHz, 1525 for (i = 0; i < M; i++) {
2059 NEWAMP1_K, 1526 if (c2->fmlfeat != NULL) {
2060 c2->phase_fft_fwd_cfg, 1527 /* We use standard nb_features=55 feature records for compatibility with
2061 c2->phase_fft_inv_cfg, 1528 * train_lpcnet.py */
2062 indexes, 1529 float features[55] = {0};
2063 c2->user_rate_K_vec_no_mean_, 1530 /* just using 18/20 for compatibility with LPCNet, coarse scaling for NN
2064 c2->post_filter_en); 1531 * input */
2065 1532 for (int j = 0; j < 18; j++)
2066 1533 features[j] = (interpolated_surface_[i][j] - 30) / 40;
2067 for(i=0; i<M; i++) { 1534 int pitch_index = 21 + 2.0 * M_PI / model[i].Wo;
2068 /* 700C is a little quiter so lets apply some experimentally derived audio gain */ 1535 features[36] = 0.02 * (pitch_index - 100);
2069 synthesise_one_frame(c2, &speech[c2->n_samp*i], &model[i], &HH[i][0], 1.5); 1536 features[37] = model[i].voiced;
2070 } 1537 fwrite(features, 55, sizeof(float), c2->fmlfeat);
1538 }
1539
1540 /* 700C is a little quieter so lets apply some experimentally derived audio
1541 * gain */
1542 synthesise_one_frame(c2, &speech[c2->n_samp * i], &model[i], &HH[i][0],
1543 1.5);
1544 }
2071} 1545}
2072 1546
2073/*---------------------------------------------------------------------------*\ 1547/*---------------------------------------------------------------------------*\
@@ -2080,48 +1554,24 @@ void codec2_decode_700c(struct CODEC2 *c2, short speech[], const unsigned char *
2080 1554
2081\*---------------------------------------------------------------------------*/ 1555\*---------------------------------------------------------------------------*/
2082 1556
2083float codec2_energy_700c(struct CODEC2 *c2, const unsigned char * bits) 1557float codec2_energy_700c(struct CODEC2 *c2, const unsigned char *bits) {
2084{ 1558 int indexes[4];
2085 int indexes[4]; 1559 unsigned int nbit = 0;
2086 unsigned int nbit = 0;
2087 1560
2088 assert(c2 != NULL); 1561 assert(c2 != NULL);
2089 1562
2090 /* unpack bits from channel ------------------------------------*/ 1563 /* unpack bits from channel ------------------------------------*/
2091 1564
2092 indexes[0] = unpack_natural_or_gray(bits, &nbit, 9, 0); 1565 indexes[0] = unpack_natural_or_gray(bits, &nbit, 9, 0);
2093 indexes[1] = unpack_natural_or_gray(bits, &nbit, 9, 0); 1566 indexes[1] = unpack_natural_or_gray(bits, &nbit, 9, 0);
2094 indexes[2] = unpack_natural_or_gray(bits, &nbit, 4, 0); 1567 indexes[2] = unpack_natural_or_gray(bits, &nbit, 4, 0);
2095 indexes[3] = unpack_natural_or_gray(bits, &nbit, 6, 0); 1568 indexes[3] = unpack_natural_or_gray(bits, &nbit, 6, 0);
2096 1569
2097 float mean = newamp1_energy_cb[0].cb[indexes[2]]; 1570 float mean = newamp1_energy_cb[0].cb[indexes[2]];
2098 mean -= 10; 1571 mean -= 10;
2099 if (indexes[3] == 0) 1572 if (indexes[3] == 0) mean -= 10;
2100 mean -= 10;
2101 1573
2102 return POW10F(mean/10.0); 1574 return POW10F(mean / 10.0);
2103}
2104
2105float codec2_energy_450(struct CODEC2 *c2, const unsigned char * bits)
2106{
2107 int indexes[4];
2108 unsigned int nbit = 0;
2109
2110 assert(c2 != NULL);
2111
2112 /* unpack bits from channel ------------------------------------*/
2113
2114 indexes[0] = unpack_natural_or_gray(bits, &nbit, 9, 0);
2115 //indexes[1] = unpack_natural_or_gray(bits, &nbit, 9, 0);
2116 indexes[2] = unpack_natural_or_gray(bits, &nbit, 3, 0);
2117 indexes[3] = unpack_natural_or_gray(bits, &nbit, 6, 0);
2118
2119 float mean = newamp2_energy_cb[0].cb[indexes[2]];
2120 mean -= 10;
2121 if (indexes[3] == 0)
2122 mean -= 10;
2123
2124 return POW10F(mean/10.0);
2125} 1575}
2126 1576
2127/*---------------------------------------------------------------------------*\ 1577/*---------------------------------------------------------------------------*\
@@ -2134,300 +1584,58 @@ float codec2_energy_450(struct CODEC2 *c2, const unsigned char * bits)
2134 1584
2135\*---------------------------------------------------------------------------*/ 1585\*---------------------------------------------------------------------------*/
2136 1586
2137float codec2_get_energy(struct CODEC2 *c2, const unsigned char *bits) 1587float codec2_get_energy(struct CODEC2 *c2, const unsigned char *bits) {
2138{ 1588 assert(c2 != NULL);
2139 assert(c2 != NULL); 1589 assert((CODEC2_MODE_ACTIVE(CODEC2_MODE_3200, c2->mode)) ||
2140 assert( 1590 (CODEC2_MODE_ACTIVE(CODEC2_MODE_2400, c2->mode)) ||
2141 ( CODEC2_MODE_ACTIVE(CODEC2_MODE_3200, c2->mode)) || 1591 (CODEC2_MODE_ACTIVE(CODEC2_MODE_1600, c2->mode)) ||
2142 ( CODEC2_MODE_ACTIVE(CODEC2_MODE_2400, c2->mode)) || 1592 (CODEC2_MODE_ACTIVE(CODEC2_MODE_1400, c2->mode)) ||
2143 ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1600, c2->mode)) || 1593 (CODEC2_MODE_ACTIVE(CODEC2_MODE_1300, c2->mode)) ||
2144 ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1400, c2->mode)) || 1594 (CODEC2_MODE_ACTIVE(CODEC2_MODE_1200, c2->mode)) ||
2145 ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1300, c2->mode)) || 1595 (CODEC2_MODE_ACTIVE(CODEC2_MODE_700C, c2->mode)));
2146 ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1200, c2->mode)) || 1596 MODEL model;
2147 ( CODEC2_MODE_ACTIVE(CODEC2_MODE_700, c2->mode)) || 1597 float xq_dec[2] = {};
2148 ( CODEC2_MODE_ACTIVE(CODEC2_MODE_700B, c2->mode)) || 1598 int e_index, WoE_index;
2149 ( CODEC2_MODE_ACTIVE(CODEC2_MODE_700C, c2->mode)) || 1599 float e = 0.0f;
2150 ( CODEC2_MODE_ACTIVE(CODEC2_MODE_450, c2->mode)) || 1600 unsigned int nbit;
2151 ( CODEC2_MODE_ACTIVE(CODEC2_MODE_450PWB, c2->mode)) 1601
2152 ); 1602 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_3200, c2->mode)) {
2153 MODEL model; 1603 nbit = 1 + 1 + WO_BITS;
2154 float xq_dec[2] = {}; 1604 e_index = unpack(bits, &nbit, E_BITS);
2155 int e_index, WoE_index; 1605 e = decode_energy(e_index, E_BITS);
2156 float e; 1606 }
2157 unsigned int nbit; 1607 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_2400, c2->mode)) {
2158 1608 nbit = 1 + 1;
2159 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_3200, c2->mode)) { 1609 WoE_index = unpack(bits, &nbit, WO_E_BITS);
2160 nbit = 1 + 1 + WO_BITS; 1610 decode_WoE(&c2->c2const, &model, &e, xq_dec, WoE_index);
2161 e_index = unpack(bits, &nbit, E_BITS); 1611 }
2162 e = decode_energy(e_index, E_BITS); 1612 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_1600, c2->mode)) {
2163 } 1613 nbit = 1 + 1 + WO_BITS;
2164 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_2400, c2->mode)) { 1614 e_index = unpack(bits, &nbit, E_BITS);
2165 nbit = 1 + 1; 1615 e = decode_energy(e_index, E_BITS);
2166 WoE_index = unpack(bits, &nbit, WO_E_BITS); 1616 }
2167 decode_WoE(&c2->c2const, &model, &e, xq_dec, WoE_index); 1617 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_1400, c2->mode)) {
2168 } 1618 nbit = 1 + 1;
2169 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1600, c2->mode)) { 1619 WoE_index = unpack(bits, &nbit, WO_E_BITS);
2170 nbit = 1 + 1 + WO_BITS; 1620 decode_WoE(&c2->c2const, &model, &e, xq_dec, WoE_index);
2171 e_index = unpack(bits, &nbit, E_BITS); 1621 }
2172 e = decode_energy(e_index, E_BITS); 1622 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_1300, c2->mode)) {
2173 } 1623 nbit = 1 + 1 + 1 + 1 + WO_BITS;
2174 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1400, c2->mode)) { 1624 e_index = unpack_natural_or_gray(bits, &nbit, E_BITS, c2->gray);
2175 nbit = 1 + 1; 1625 e = decode_energy(e_index, E_BITS);
2176 WoE_index = unpack(bits, &nbit, WO_E_BITS); 1626 }
2177 decode_WoE(&c2->c2const, &model, &e, xq_dec, WoE_index); 1627 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_1200, c2->mode)) {
2178 } 1628 nbit = 1 + 1;
2179 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1300, c2->mode)) { 1629 WoE_index = unpack(bits, &nbit, WO_E_BITS);
2180 nbit = 1 + 1 + 1 + 1 + WO_BITS; 1630 decode_WoE(&c2->c2const, &model, &e, xq_dec, WoE_index);
2181 e_index = unpack_natural_or_gray(bits, &nbit, E_BITS, c2->gray); 1631 }
2182 e = decode_energy(e_index, E_BITS); 1632 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_700C, c2->mode)) {
2183 } 1633 e = codec2_energy_700c(c2, bits);
2184 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_1200, c2->mode)) { 1634 }
2185 nbit = 1 + 1;
2186 WoE_index = unpack(bits, &nbit, WO_E_BITS);
2187 decode_WoE(&c2->c2const, &model, &e, xq_dec, WoE_index);
2188 }
2189 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_700, c2->mode)) {
2190 nbit = 1 + 5;
2191 e_index = unpack_natural_or_gray(bits, &nbit, 3, c2->gray);
2192 e = decode_energy(e_index, 3);
2193 }
2194 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_700B, c2->mode)) {
2195 nbit = 1 + 5;
2196 e_index = unpack_natural_or_gray(bits, &nbit, 3, c2->gray);
2197 e = decode_energy(e_index, 3);
2198 }
2199 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_700C, c2->mode)) {
2200 e = codec2_energy_700c(c2, bits);
2201 }
2202 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_450, c2->mode) || CODEC2_MODE_ACTIVE(CODEC2_MODE_450PWB, c2->mode)) {
2203 e = codec2_energy_450(c2, bits);
2204 }
2205
2206 return e;
2207}
2208
2209
2210/*---------------------------------------------------------------------------*\
2211
2212 FUNCTION....: codec2_encode_450
2213 AUTHOR......: Thomas Kurin and Stefan Erhardt
2214 INSTITUTE...: Institute for Electronics Engineering, University of Erlangen-Nuremberg
2215 DATE CREATED: July 2018
2216
2217 450 bit/s codec that uses newamp2 fixed rate VQ of amplitudes.
2218
2219 Encodes 320 speech samples (40ms of speech) into 28 bits.
2220
2221 The codec2 algorithm actually operates internally on 10ms (80
2222 sample) frames, so we run the encoding algorithm four times:
2223
2224 frame 0: nothing
2225 frame 1: nothing
2226 frame 2: nothing
2227 frame 3: 9 bit 1 stage VQ, 3 bits energy,
2228 6 bit scalar Wo/voicing/plosive. No spare bits.
2229
2230 If a plosive is detected the frame at the energy-step is encoded.
2231
2232 Voicing is encoded using the 000000 index of the Wo quantiser.
2233 Plosive is encoded using the 111111 index of the Wo quantiser.
2234
2235 The bit allocation is:
2236
2237 Parameter frames 1-3 frame 4 Total
2238 -----------------------------------------------------------
2239 Harmonic magnitudes (rate k VQ) 0 9 9
2240 Energy 0 3 3
2241 log Wo/voicing/plosive 0 6 6
2242 TOTAL 0 18 18
2243
2244
2245\*---------------------------------------------------------------------------*/
2246
2247void codec2_encode_450(struct CODEC2 *c2, unsigned char * bits, short speech[])
2248{
2249 MODEL model;
2250 int indexes[4], i,h, M=4;
2251 unsigned int nbit = 0;
2252 int plosiv = 0;
2253 float energydelta[M];
2254 int spectralCounter;
2255
2256 assert(c2 != NULL);
2257
2258 memset(bits, '\0', ((codec2_bits_per_frame(c2) + 7) / 8));
2259 for(i=0; i<M; i++){
2260 analyse_one_frame(c2, &model, &speech[i*c2->n_samp]);
2261 energydelta[i] = 0;
2262 spectralCounter = 0;
2263 for(h = 0;h<(model.L);h++){
2264 //only detect above 300 Hz
2265 if(h*model.Wo*(c2->c2const.Fs/2000.0)/M_PI > 0.3){
2266 energydelta[i] = energydelta[i] + 20.0*log10(model.A[10]+1E-16);
2267 spectralCounter = spectralCounter+1;
2268 }
2269
2270 }
2271 energydelta[i] = energydelta[i] / spectralCounter ;
2272 }
2273 //Constants for plosive Detection tdB = threshold; minPwr = from below this level plosives have to rise
2274 float tdB = 15; //not fixed can be changed
2275 float minPwr = 15; //not fixed can be changed
2276 if((c2->energy_prev)<minPwr && energydelta[0]>((c2->energy_prev)+tdB)){
2277
2278 plosiv = 1;
2279 }
2280 if(energydelta[0]<minPwr && energydelta[1]>(energydelta[0]+tdB)){
2281
2282 plosiv = 2;
2283 }
2284 if(energydelta[1]<minPwr &&energydelta[2]>(energydelta[1]+tdB)){
2285
2286 plosiv = 3;
2287 }
2288 if(energydelta[2]<minPwr &&energydelta[3]>(energydelta[2]+tdB)){
2289
2290 plosiv = 4;
2291 }
2292 if(plosiv != 0 && plosiv != 4){
2293 analyse_one_frame(c2, &model, &speech[(plosiv-1)*c2->n_samp]);
2294 }
2295
2296 c2->energy_prev = energydelta[3];
2297
2298
2299 int K = 29;
2300 float rate_K_vec[K], mean;
2301 float rate_K_vec_no_mean[K], rate_K_vec_no_mean_[K];
2302 if(plosiv > 0){
2303 plosiv = 1;
2304 }
2305 newamp2_model_to_indexes(&c2->c2const,
2306 indexes,
2307 &model,
2308 rate_K_vec,
2309 c2->n2_rate_K_sample_freqs_kHz,
2310 K,
2311 &mean,
2312 rate_K_vec_no_mean,
2313 rate_K_vec_no_mean_,
2314 plosiv);
2315
2316
2317 pack_natural_or_gray(bits, &nbit, indexes[0], 9, 0);
2318 //pack_natural_or_gray(bits, &nbit, indexes[1], 9, 0);
2319 pack_natural_or_gray(bits, &nbit, indexes[2], 3, 0);
2320 pack_natural_or_gray(bits, &nbit, indexes[3], 6, 0);
2321
2322 assert(nbit == (unsigned)codec2_bits_per_frame(c2));
2323}
2324
2325
2326/*---------------------------------------------------------------------------*\
2327
2328 FUNCTION....: codec2_decode_450
2329 AUTHOR......: Thomas Kurin and Stefan Erhardt
2330 INSTITUTE...: Institute for Electronics Engineering, University of Erlangen-Nuremberg
2331 DATE CREATED: July 2018
2332
2333\*---------------------------------------------------------------------------*/
2334
2335void codec2_decode_450(struct CODEC2 *c2, short speech[], const unsigned char * bits)
2336{
2337 MODEL model[4];
2338 int indexes[4];
2339 int i;
2340 unsigned int nbit = 0;
2341
2342 assert(c2 != NULL);
2343
2344 /* unpack bits from channel ------------------------------------*/
2345
2346 indexes[0] = unpack_natural_or_gray(bits, &nbit, 9, 0);
2347 //indexes[1] = unpack_natural_or_gray(bits, &nbit, 9, 0);
2348 indexes[2] = unpack_natural_or_gray(bits, &nbit, 3, 0);
2349 indexes[3] = unpack_natural_or_gray(bits, &nbit, 6, 0);
2350
2351 int M = 4;
2352 COMP HH[M][MAX_AMP+1];
2353 float interpolated_surface_[M][NEWAMP2_K];
2354 int pwbFlag = 0;
2355
2356 newamp2_indexes_to_model(&c2->c2const,
2357 model,
2358 (COMP*)HH,
2359 (float*)interpolated_surface_,
2360 c2->n2_prev_rate_K_vec_,
2361 &c2->Wo_left,
2362 &c2->voicing_left,
2363 c2->n2_rate_K_sample_freqs_kHz,
2364 NEWAMP2_K,
2365 c2->phase_fft_fwd_cfg,
2366 c2->phase_fft_inv_cfg,
2367 indexes,
2368 1.5,
2369 pwbFlag);
2370
2371
2372 for(i=0; i<M; i++) {
2373 synthesise_one_frame(c2, &speech[c2->n_samp*i], &model[i], &HH[i][0], 1.5);
2374 }
2375}
2376
2377/*---------------------------------------------------------------------------*\
2378
2379 FUNCTION....: codec2_decode_450pwb
2380 AUTHOR......: Thomas Kurin and Stefan Erhardt
2381 INSTITUTE...: Institute for Electronics Engineering, University of Erlangen-Nuremberg
2382 DATE CREATED: July 2018
2383
2384 Decodes the 450 codec data in pseudo wideband at 16kHz samplerate.
2385
2386\*---------------------------------------------------------------------------*/
2387 1635
2388void codec2_decode_450pwb(struct CODEC2 *c2, short speech[], const unsigned char * bits) 1636 return e;
2389{
2390 MODEL model[4];
2391 int indexes[4];
2392 int i;
2393 unsigned int nbit = 0;
2394
2395 assert(c2 != NULL);
2396
2397 /* unpack bits from channel ------------------------------------*/
2398
2399 indexes[0] = unpack_natural_or_gray(bits, &nbit, 9, 0);
2400 //indexes[1] = unpack_natural_or_gray(bits, &nbit, 9, 0);
2401 indexes[2] = unpack_natural_or_gray(bits, &nbit, 3, 0);
2402 indexes[3] = unpack_natural_or_gray(bits, &nbit, 6, 0);
2403
2404 int M = 4;
2405 COMP HH[M][MAX_AMP+1];
2406 float interpolated_surface_[M][NEWAMP2_16K_K];
2407 int pwbFlag = 1;
2408
2409 newamp2_indexes_to_model(&c2->c2const,
2410 model,
2411 (COMP*)HH,
2412 (float*)interpolated_surface_,
2413 c2->n2_pwb_prev_rate_K_vec_,
2414 &c2->Wo_left,
2415 &c2->voicing_left,
2416 c2->n2_pwb_rate_K_sample_freqs_kHz,
2417 NEWAMP2_16K_K,
2418 c2->phase_fft_fwd_cfg,
2419 c2->phase_fft_inv_cfg,
2420 indexes,
2421 1.5,
2422 pwbFlag);
2423
2424
2425 for(i=0; i<M; i++) {
2426 synthesise_one_frame(c2, &speech[c2->n_samp*i], &model[i], &HH[i][0], 1.5);
2427 }
2428} 1637}
2429 1638
2430
2431/*---------------------------------------------------------------------------* \ 1639/*---------------------------------------------------------------------------* \
2432 1640
2433 FUNCTION....: synthesise_one_frame() 1641 FUNCTION....: synthesise_one_frame()
@@ -2438,56 +1646,41 @@ void codec2_decode_450pwb(struct CODEC2 *c2, short speech[], const unsigned char
2438 1646
2439\*---------------------------------------------------------------------------*/ 1647\*---------------------------------------------------------------------------*/
2440 1648
2441void synthesise_one_frame(struct CODEC2 *c2, short speech[], MODEL *model, COMP Aw[], float gain) 1649void synthesise_one_frame(struct CODEC2 *c2, short speech[], MODEL *model,
2442{ 1650 COMP Aw[], float gain) {
2443 int i; 1651 int i;
2444 //PROFILE_VAR(phase_start, pf_start, synth_start); 1652
2445 1653 if (CODEC2_MODE_ACTIVE(CODEC2_MODE_700C, c2->mode)) {
2446 //#ifdef DUMP 1654 /* newamp1, we've already worked out rate L phase */
2447 //dump_quantised_model(model); 1655 COMP *H = Aw;
2448 //#endif 1656 phase_synth_zero_order(c2->n_samp, model, &c2->ex_phase, H);
2449 1657 } else {
2450 //PROFILE_SAMPLE(phase_start); 1658 /* LPC based phase synthesis */
2451 1659 COMP H[MAX_AMP + 1];
2452 if ( CODEC2_MODE_ACTIVE(CODEC2_MODE_700C, c2->mode) || CODEC2_MODE_ACTIVE(CODEC2_MODE_450, c2->mode) || CODEC2_MODE_ACTIVE(CODEC2_MODE_450PWB, c2->mode) ) { 1660 sample_phase(model, H, Aw);
2453 /* newamp1/2, we've already worked out rate L phase */ 1661 phase_synth_zero_order(c2->n_samp, model, &c2->ex_phase, H);
2454 COMP *H = Aw; 1662 }
2455 phase_synth_zero_order(c2->n_samp, model, &c2->ex_phase, H); 1663
2456 } else { 1664 postfilter(model, &c2->bg_est);
2457 /* LPC based phase synthesis */ 1665 synthesise(c2->n_samp, c2->fftr_inv_cfg, c2->Sn_, model, c2->Pn, 1);
2458 COMP H[MAX_AMP+1]; 1666
2459 sample_phase(model, H, Aw); 1667 for (i = 0; i < c2->n_samp; i++) {
2460 phase_synth_zero_order(c2->n_samp, model, &c2->ex_phase, H); 1668 c2->Sn_[i] *= gain;
2461 } 1669 }
2462 1670
2463 //PROFILE_SAMPLE_AND_LOG(pf_start, phase_start, " phase_synth"); 1671 ear_protection(c2->Sn_, c2->n_samp);
2464 1672
2465 postfilter(model, &c2->bg_est); 1673 for (i = 0; i < c2->n_samp; i++) {
2466 1674 if (c2->Sn_[i] > 32767.0)
2467 //PROFILE_SAMPLE_AND_LOG(synth_start, pf_start, " postfilter"); 1675 speech[i] = 32767;
2468 1676 else if (c2->Sn_[i] < -32767.0)
2469 synthesise(c2->n_samp, c2->fftr_inv_cfg, c2->Sn_, model, c2->Pn, 1); 1677 speech[i] = -32767;
2470 1678 else
2471 for(i=0; i<c2->n_samp; i++) { 1679 speech[i] = c2->Sn_[i];
2472 c2->Sn_[i] *= gain; 1680 }
2473 }
2474
2475 //PROFILE_SAMPLE_AND_LOG2(synth_start, " synth");
2476
2477 ear_protection(c2->Sn_, c2->n_samp);
2478
2479 for(i=0; i<c2->n_samp; i++) {
2480 if (c2->Sn_[i] > 32767.0)
2481 speech[i] = 32767;
2482 else if (c2->Sn_[i] < -32767.0)
2483 speech[i] = -32767;
2484 else
2485 speech[i] = c2->Sn_[i];
2486 }
2487
2488} 1681}
2489 1682
2490/*---------------------------------------------------------------------------*\ 1683/*---------------------------------------------------------------------------* \
2491 1684
2492 FUNCTION....: analyse_one_frame() 1685 FUNCTION....: analyse_one_frame()
2493 AUTHOR......: David Rowe 1686 AUTHOR......: David Rowe
@@ -2498,48 +1691,40 @@ void synthesise_one_frame(struct CODEC2 *c2, short speech[], MODEL *model, COMP
2498 1691
2499\*---------------------------------------------------------------------------*/ 1692\*---------------------------------------------------------------------------*/
2500 1693
2501void analyse_one_frame(struct CODEC2 *c2, MODEL *model, short speech[]) 1694void analyse_one_frame(struct CODEC2 *c2, MODEL *model, short speech[]) {
2502{ 1695 COMP Sw[FFT_ENC];
2503 COMP Sw[FFT_ENC]; 1696 float pitch;
2504 float pitch; 1697 int i;
2505 int i; 1698 int n_samp = c2->n_samp;
2506 //PROFILE_VAR(dft_start, nlp_start, model_start, two_stage, estamps); 1699 int m_pitch = c2->m_pitch;
2507 int n_samp = c2->n_samp;
2508 int m_pitch = c2->m_pitch;
2509 1700
2510 /* Read input speech */ 1701 /* Read input speech */
2511 1702
2512 for(i=0; i<m_pitch-n_samp; i++) 1703 for (i = 0; i < m_pitch - n_samp; i++) c2->Sn[i] = c2->Sn[i + n_samp];
2513 c2->Sn[i] = c2->Sn[i+n_samp]; 1704 for (i = 0; i < n_samp; i++) c2->Sn[i + m_pitch - n_samp] = speech[i];
2514 for(i=0; i<n_samp; i++)
2515 c2->Sn[i+m_pitch-n_samp] = speech[i];
2516 1705
2517 //PROFILE_SAMPLE(dft_start); 1706 dft_speech(&c2->c2const, c2->fft_fwd_cfg, Sw, c2->Sn, c2->w);
2518 dft_speech(&c2->c2const, c2->fft_fwd_cfg, Sw, c2->Sn, c2->w);
2519 //PROFILE_SAMPLE_AND_LOG(nlp_start, dft_start, " dft_speech");
2520 1707
2521 /* Estimate pitch */ 1708 /* Estimate pitch */
1709 nlp(c2->nlp, c2->Sn, n_samp, &pitch, Sw, c2->W, &c2->prev_f0_enc);
1710 model->Wo = TWO_PI / pitch;
1711 model->L = PI / model->Wo;
2522 1712
2523 nlp(c2->nlp, c2->Sn, n_samp, &pitch, Sw, c2->W, &c2->prev_f0_enc); 1713 /* estimate model parameters */
2524 //PROFILE_SAMPLE_AND_LOG(model_start, nlp_start, " nlp"); 1714 two_stage_pitch_refinement(&c2->c2const, model, Sw);
2525 1715
2526 model->Wo = TWO_PI/pitch; 1716 /* estimate phases when doing ML experiments */
2527 model->L = PI/model->Wo; 1717 if (c2->fmlfeat != NULL)
2528 1718 estimate_amplitudes(model, Sw, c2->W, 1);
2529 /* estimate model parameters */ 1719 else
2530
2531 two_stage_pitch_refinement(&c2->c2const, model, Sw);
2532 //PROFILE_SAMPLE_AND_LOG(two_stage, model_start, " two_stage");
2533 estimate_amplitudes(model, Sw, c2->W, 0); 1720 estimate_amplitudes(model, Sw, c2->W, 0);
2534 //PROFILE_SAMPLE_AND_LOG(estamps, two_stage, " est_amps"); 1721 est_voicing_mbe(&c2->c2const, model, Sw, c2->W);
2535 est_voicing_mbe(&c2->c2const, model, Sw, c2->W); 1722#ifdef DUMP
2536 //PROFILE_SAMPLE_AND_LOG2(estamps, " est_voicing"); 1723 dump_model(model);
2537 #ifdef DUMP 1724#endif
2538 dump_model(model);
2539 #endif
2540} 1725}
2541 1726
2542/*---------------------------------------------------------------------------*\ 1727/*---------------------------------------------------------------------------* \
2543 1728
2544 FUNCTION....: ear_protection() 1729 FUNCTION....: ear_protection()
2545 AUTHOR......: David Rowe 1730 AUTHOR......: David Rowe
@@ -2552,40 +1737,37 @@ void analyse_one_frame(struct CODEC2 *c2, MODEL *model, short speech[])
2552\*---------------------------------------------------------------------------*/ 1737\*---------------------------------------------------------------------------*/
2553 1738
2554static void ear_protection(float in_out[], int n) { 1739static void ear_protection(float in_out[], int n) {
2555 float max_sample, over, gain; 1740 float max_sample, over, gain;
2556 int i; 1741 int i;
2557 1742
2558 /* find maximum sample in frame */ 1743 /* find maximum sample in frame */
2559 1744
2560 max_sample = 0.0; 1745 max_sample = 0.0;
2561 for(i=0; i<n; i++) 1746 for (i = 0; i < n; i++)
2562 if (in_out[i] > max_sample) 1747 if (in_out[i] > max_sample) max_sample = in_out[i];
2563 max_sample = in_out[i];
2564 1748
2565 /* determine how far above set point */ 1749 /* determine how far above set point */
2566 1750
2567 over = max_sample/30000.0; 1751 over = max_sample / 30000.0;
2568 1752
2569 /* If we are x dB over set point we reduce level by 2x dB, this 1753 /* If we are x dB over set point we reduce level by 2x dB, this
2570 attenuates major excursions in amplitude (likely to be caused 1754 attenuates major excursions in amplitude (likely to be caused
2571 by bit errors) more than smaller ones */ 1755 by bit errors) more than smaller ones */
2572 1756
2573 if (over > 1.0) { 1757 if (over > 1.0) {
2574 gain = 1.0/(over*over); 1758 gain = 1.0 / (over * over);
2575 //fprintf(stderr, "gain: %f\n", gain); 1759 for (i = 0; i < n; i++) in_out[i] *= gain;
2576 for(i=0; i<n; i++) 1760 }
2577 in_out[i] *= gain;
2578 }
2579} 1761}
2580 1762
2581void codec2_set_lpc_post_filter(struct CODEC2 *c2, int enable, int bass_boost, float beta, float gamma) 1763void codec2_set_lpc_post_filter(struct CODEC2 *c2, int enable, int bass_boost,
2582{ 1764 float beta, float gamma) {
2583 assert((beta >= 0.0) && (beta <= 1.0)); 1765 assert((beta >= 0.0) && (beta <= 1.0));
2584 assert((gamma >= 0.0) && (gamma <= 1.0)); 1766 assert((gamma >= 0.0) && (gamma <= 1.0));
2585 c2->lpc_pf = enable; 1767 c2->lpc_pf = enable;
2586 c2->bass_boost = bass_boost; 1768 c2->bass_boost = bass_boost;
2587 c2->beta = beta; 1769 c2->beta = beta;
2588 c2->gamma = gamma; 1770 c2->gamma = gamma;
2589} 1771}
2590 1772
2591/* 1773/*
@@ -2594,29 +1776,22 @@ void codec2_set_lpc_post_filter(struct CODEC2 *c2, int enable, int bass_boost, f
2594 Experimental method of sending voice/data frames for FreeDV. 1776 Experimental method of sending voice/data frames for FreeDV.
2595*/ 1777*/
2596 1778
2597int codec2_get_spare_bit_index(struct CODEC2 *c2) 1779int codec2_get_spare_bit_index(struct CODEC2 *c2) {
2598{ 1780 assert(c2 != NULL);
2599 assert(c2 != NULL);
2600 1781
2601 switch(c2->mode) { 1782 switch (c2->mode) {
2602 case CODEC2_MODE_1300: 1783 case CODEC2_MODE_1300:
2603 return 2; // bit 2 (3th bit) is v2 (third voicing bit) 1784 return 2; // bit 2 (3th bit) is v2 (third voicing bit)
2604 break; 1785 break;
2605 case CODEC2_MODE_1400: 1786 case CODEC2_MODE_1400:
2606 return 10; // bit 10 (11th bit) is v2 (third voicing bit) 1787 return 10; // bit 10 (11th bit) is v2 (third voicing bit)
2607 break; 1788 break;
2608 case CODEC2_MODE_1600: 1789 case CODEC2_MODE_1600:
2609 return 15; // bit 15 (16th bit) is v2 (third voicing bit) 1790 return 15; // bit 15 (16th bit) is v2 (third voicing bit)
2610 break; 1791 break;
2611 case CODEC2_MODE_700: 1792 }
2612 return 26; // bits 26 and 27 are spare
2613 break;
2614 case CODEC2_MODE_700B:
2615 return 27; // bit 27 is spare
2616 break;
2617 }
2618 1793
2619 return -1; 1794 return -1;
2620} 1795}
2621 1796
2622/* 1797/*
@@ -2624,111 +1799,123 @@ int codec2_get_spare_bit_index(struct CODEC2 *c2)
2624 for convenience. 1799 for convenience.
2625*/ 1800*/
2626 1801
2627int codec2_rebuild_spare_bit(struct CODEC2 *c2, int unpacked_bits[]) 1802int codec2_rebuild_spare_bit(struct CODEC2 *c2, char unpacked_bits[]) {
2628{ 1803 int v1, v3;
2629 int v1,v3;
2630 1804
2631 assert(c2 != NULL); 1805 assert(c2 != NULL);
2632 1806
2633 v1 = unpacked_bits[1]; 1807 v1 = unpacked_bits[1];
2634 1808
2635 switch(c2->mode) { 1809 switch (c2->mode) {
2636 case CODEC2_MODE_1300: 1810 case CODEC2_MODE_1300:
2637 1811
2638 v3 = unpacked_bits[1+1+1]; 1812 v3 = unpacked_bits[1 + 1 + 1];
2639 1813
2640 /* if either adjacent frame is voiced, make this one voiced */ 1814 /* if either adjacent frame is voiced, make this one voiced */
2641 1815
2642 unpacked_bits[2] = (v1 || v3); 1816 unpacked_bits[2] = (v1 || v3);
2643 1817
2644 return 0; 1818 return 0;
2645 1819
2646 break; 1820 break;
2647 1821
2648 case CODEC2_MODE_1400: 1822 case CODEC2_MODE_1400:
2649 1823
2650 v3 = unpacked_bits[1+1+8+1]; 1824 v3 = unpacked_bits[1 + 1 + 8 + 1];
2651 1825
2652 /* if either adjacent frame is voiced, make this one voiced */ 1826 /* if either adjacent frame is voiced, make this one voiced */
2653 1827
2654 unpacked_bits[10] = (v1 || v3); 1828 unpacked_bits[10] = (v1 || v3);
2655 1829
2656 return 0; 1830 return 0;
2657 1831
2658 break; 1832 break;
2659 1833
2660 case CODEC2_MODE_1600: 1834 case CODEC2_MODE_1600:
2661 v3 = unpacked_bits[1+1+8+5+1]; 1835 v3 = unpacked_bits[1 + 1 + 8 + 5 + 1];
2662 1836
2663 /* if either adjacent frame is voiced, make this one voiced */ 1837 /* if either adjacent frame is voiced, make this one voiced */
2664 1838
2665 unpacked_bits[15] = (v1 || v3); 1839 unpacked_bits[15] = (v1 || v3);
2666 1840
2667 return 0; 1841 return 0;
2668 1842
2669 break; 1843 break;
2670 } 1844 }
2671 1845
2672 return -1; 1846 return -1;
2673} 1847}
2674 1848
2675void codec2_set_natural_or_gray(struct CODEC2 *c2, int gray) 1849void codec2_set_natural_or_gray(struct CODEC2 *c2, int gray) {
2676{ 1850 assert(c2 != NULL);
2677 assert(c2 != NULL); 1851 c2->gray = gray;
2678 c2->gray = gray;
2679} 1852}
2680 1853
2681void codec2_set_softdec(struct CODEC2 *c2, float *softdec) 1854void codec2_set_softdec(struct CODEC2 *c2, float *softdec) {
2682{ 1855 assert(c2 != NULL);
2683 assert(c2 != NULL); 1856 c2->softdec = softdec;
2684 c2->softdec = softdec;
2685} 1857}
2686 1858
2687void codec2_open_mlfeat(struct CODEC2 *codec2_state, char *filename) { 1859void codec2_open_mlfeat(struct CODEC2 *codec2_state, char *feat_fn,
2688 if ((codec2_state->fmlfeat = fopen(filename, "wb")) == NULL) { 1860 char *model_fn) {
2689 fprintf(stderr, "error opening machine learning feature file: %s\n", filename); 1861 if ((codec2_state->fmlfeat = fopen(feat_fn, "wb")) == NULL) {
2690 exit(1); 1862 fprintf(stderr, "error opening machine learning feature file: %s\n",
2691 } 1863 feat_fn);
1864 exit(1);
1865 }
1866 if (model_fn) {
1867 if ((codec2_state->fmlmodel = fopen(model_fn, "wb")) == NULL) {
1868 fprintf(stderr, "error opening machine learning Codec 2 model file: %s\n",
1869 feat_fn);
1870 exit(1);
1871 }
1872 }
2692} 1873}
2693 1874
2694#ifndef __EMBEDDED__ 1875#ifndef __EMBEDDED__
2695void codec2_load_codebook(struct CODEC2 *codec2_state, int num, char *filename) { 1876void codec2_load_codebook(struct CODEC2 *codec2_state, int num,
2696 FILE *f; 1877 char *filename) {
2697 1878 FILE *f;
2698 if ((f = fopen(filename, "rb")) == NULL) { 1879
2699 fprintf(stderr, "error opening codebook file: %s\n", filename); 1880 if ((f = fopen(filename, "rb")) == NULL) {
2700 exit(1); 1881 fprintf(stderr, "error opening codebook file: %s\n", filename);
2701 } 1882 exit(1);
2702 //fprintf(stderr, "reading newamp1vq_cb[%d] k=%d m=%d\n", num, newamp1vq_cb[num].k, newamp1vq_cb[num].m); 1883 }
2703 float tmp[newamp1vq_cb[num].k*newamp1vq_cb[num].m]; 1884 // fprintf(stderr, "reading newamp1vq_cb[%d] k=%d m=%d\n", num,
2704 int nread = fread(tmp, sizeof(float), newamp1vq_cb[num].k*newamp1vq_cb[num].m, f); 1885 // newamp1vq_cb[num].k, newamp1vq_cb[num].m);
2705 float *p = (float*)newamp1vq_cb[num].cb; 1886 float tmp[newamp1vq_cb[num].k * newamp1vq_cb[num].m];
2706 for(int i=0; i<newamp1vq_cb[num].k*newamp1vq_cb[num].m; i++) 1887 int nread =
2707 p[i] = tmp[i]; 1888 fread(tmp, sizeof(float), newamp1vq_cb[num].k * newamp1vq_cb[num].m, f);
2708 // fprintf(stderr, "nread = %d %f %f\n", nread, newamp1vq_cb[num].cb[0], newamp1vq_cb[num].cb[1]); 1889 float *p = (float *)newamp1vq_cb[num].cb;
2709 assert(nread == newamp1vq_cb[num].k*newamp1vq_cb[num].m); 1890 for (int i = 0; i < newamp1vq_cb[num].k * newamp1vq_cb[num].m; i++)
2710 fclose(f); 1891 p[i] = tmp[i];
1892 // fprintf(stderr, "nread = %d %f %f\n", nread, newamp1vq_cb[num].cb[0],
1893 // newamp1vq_cb[num].cb[1]);
1894 assert(nread == newamp1vq_cb[num].k * newamp1vq_cb[num].m);
1895 fclose(f);
2711} 1896}
2712#endif 1897#endif
2713 1898
2714float codec2_get_var(struct CODEC2 *codec2_state) { 1899float codec2_get_var(struct CODEC2 *codec2_state) {
2715 if (codec2_state->nse) 1900 if (codec2_state->nse)
2716 return codec2_state->se/codec2_state->nse; 1901 return codec2_state->se / codec2_state->nse;
2717 else 1902 else
2718 return 0; 1903 return 0;
2719} 1904}
2720 1905
2721float *codec2_enable_user_ratek(struct CODEC2 *codec2_state, int *K) { 1906float *codec2_enable_user_ratek(struct CODEC2 *codec2_state, int *K) {
2722 codec2_state->user_rate_K_vec_no_mean_ = (float*)malloc(sizeof(float)*NEWAMP1_K); 1907 codec2_state->user_rate_K_vec_no_mean_ =
2723 *K = NEWAMP1_K; 1908 (float *)malloc(sizeof(float) * NEWAMP1_K);
2724 return codec2_state->user_rate_K_vec_no_mean_; 1909 *K = NEWAMP1_K;
1910 return codec2_state->user_rate_K_vec_no_mean_;
2725} 1911}
2726 1912
2727void codec2_700c_post_filter(struct CODEC2 *codec2_state, int en) { 1913void codec2_700c_post_filter(struct CODEC2 *codec2_state, bool en) {
2728 codec2_state->post_filter_en = en; 1914 codec2_state->post_filter_en = en;
2729} 1915}
2730 1916
2731void codec2_700c_eq(struct CODEC2 *codec2_state, int en) { 1917void codec2_700c_eq(struct CODEC2 *codec2_state, bool en) {
2732 codec2_state->eq_en = en; 1918 codec2_state->eq_en = en;
2733 codec2_state->se = 0.0; codec2_state->nse = 0; 1919 codec2_state->se = 0.0;
1920 codec2_state->nse = 0;
2734} 1921}
diff --git a/codec2.h b/codec2.h
index 709b462..97e9262 100644
--- a/codec2.h
+++ b/codec2.h
@@ -26,26 +26,22 @@
26 along with this program; if not, see <http://www.gnu.org/licenses/>. 26 along with this program; if not, see <http://www.gnu.org/licenses/>.
27*/ 27*/
28 28
29#ifndef __CODEC2__
30#define __CODEC2__
31#include "version.h"
32#include <stdbool.h>
33
29#ifdef __cplusplus 34#ifdef __cplusplus
30 extern "C" { 35extern "C" {
31#endif 36#endif
32 37
33#ifndef __CODEC2__ 38#define CODEC2_MODE_3200 0
34#define __CODEC2__ 39#define CODEC2_MODE_2400 1
35 40#define CODEC2_MODE_1600 2
36#include <version.h> 41#define CODEC2_MODE_1400 3
37 42#define CODEC2_MODE_1300 4
38#define CODEC2_MODE_3200 0 43#define CODEC2_MODE_1200 5
39#define CODEC2_MODE_2400 1 44#define CODEC2_MODE_700C 8
40#define CODEC2_MODE_1600 2
41#define CODEC2_MODE_1400 3
42#define CODEC2_MODE_1300 4
43#define CODEC2_MODE_1200 5
44#define CODEC2_MODE_700 6
45#define CODEC2_MODE_700B 7
46#define CODEC2_MODE_700C 8
47#define CODEC2_MODE_450 10
48#define CODEC2_MODE_450PWB 11
49 45
50#ifndef CODEC2_MODE_EN_DEFAULT 46#ifndef CODEC2_MODE_EN_DEFAULT
51#define CODEC2_MODE_EN_DEFAULT 1 47#define CODEC2_MODE_EN_DEFAULT 1
@@ -55,76 +51,70 @@
55// disable during compile time with -DCODEC2_MODE_1600_EN=0 51// disable during compile time with -DCODEC2_MODE_1600_EN=0
56// all but CODEC2 1600 are enabled then 52// all but CODEC2 1600 are enabled then
57 53
58//or the other way round 54// or the other way round
59// -DCODEC2_MODE_EN_DEFAULT=0 -DCODEC2_MODE_1600_EN=1 55// -DCODEC2_MODE_EN_DEFAULT=0 -DCODEC2_MODE_1600_EN=1
60// only CODEC2 Mode 1600 56// only CODEC2 Mode 1600
61 57
62#if !defined(CODEC2_MODE_3200_EN) 58#if !defined(CODEC2_MODE_3200_EN)
63 #define CODEC2_MODE_3200_EN CODEC2_MODE_EN_DEFAULT 59#define CODEC2_MODE_3200_EN CODEC2_MODE_EN_DEFAULT
64#endif 60#endif
65#if !defined(CODEC2_MODE_2400_EN) 61#if !defined(CODEC2_MODE_2400_EN)
66 #define CODEC2_MODE_2400_EN CODEC2_MODE_EN_DEFAULT 62#define CODEC2_MODE_2400_EN CODEC2_MODE_EN_DEFAULT
67#endif 63#endif
68#if !defined(CODEC2_MODE_1600_EN) 64#if !defined(CODEC2_MODE_1600_EN)
69 #define CODEC2_MODE_1600_EN CODEC2_MODE_EN_DEFAULT 65#define CODEC2_MODE_1600_EN CODEC2_MODE_EN_DEFAULT
70#endif 66#endif
71#if !defined(CODEC2_MODE_1400_EN) 67#if !defined(CODEC2_MODE_1400_EN)
72 #define CODEC2_MODE_1400_EN CODEC2_MODE_EN_DEFAULT 68#define CODEC2_MODE_1400_EN CODEC2_MODE_EN_DEFAULT
73#endif 69#endif
74#if !defined(CODEC2_MODE_1300_EN) 70#if !defined(CODEC2_MODE_1300_EN)
75 #define CODEC2_MODE_1300_EN CODEC2_MODE_EN_DEFAULT 71#define CODEC2_MODE_1300_EN CODEC2_MODE_EN_DEFAULT
76#endif 72#endif
77#if !defined(CODEC2_MODE_1200_EN) 73#if !defined(CODEC2_MODE_1200_EN)
78 #define CODEC2_MODE_1200_EN CODEC2_MODE_EN_DEFAULT 74#define CODEC2_MODE_1200_EN CODEC2_MODE_EN_DEFAULT
79#endif
80#if !defined(CODEC2_MODE_700_EN)
81 #define CODEC2_MODE_700_EN CODEC2_MODE_EN_DEFAULT
82#endif
83#if !defined(CODEC2_MODE_700B_EN)
84 #define CODEC2_MODE_700B_EN CODEC2_MODE_EN_DEFAULT
85#endif 75#endif
86#if !defined(CODEC2_MODE_700C_EN) 76#if !defined(CODEC2_MODE_700C_EN)
87 #define CODEC2_MODE_700C_EN CODEC2_MODE_EN_DEFAULT 77#define CODEC2_MODE_700C_EN CODEC2_MODE_EN_DEFAULT
88#endif
89#if !defined(CODEC2_MODE_450_EN)
90 #define CODEC2_MODE_450_EN CODEC2_MODE_EN_DEFAULT
91#endif
92#if !defined(CODEC2_MODE_450PWB_EN)
93 #define CODEC2_MODE_450PWB_EN CODEC2_MODE_EN_DEFAULT
94#endif 78#endif
95 79
96#define CODEC2_MODE_ACTIVE(mode_name, var) ((mode_name##_EN) == 0 ? 0: (var) == mode_name) 80#define CODEC2_MODE_ACTIVE(mode_name, var) \
81 ((mode_name##_EN) == 0 ? 0 : (var) == mode_name)
97 82
98struct CODEC2; 83struct CODEC2;
99 84
100struct CODEC2 * codec2_create(int mode); 85struct CODEC2 *codec2_create(int mode);
101void codec2_destroy(struct CODEC2 *codec2_state); 86void codec2_destroy(struct CODEC2 *codec2_state);
102void codec2_encode(struct CODEC2 *codec2_state, unsigned char * bits, short speech_in[]); 87void codec2_encode(struct CODEC2 *codec2_state, unsigned char bytes[],
103void codec2_decode(struct CODEC2 *codec2_state, short speech_out[], const unsigned char *bits); 88 short speech_in[]);
104void codec2_decode_ber(struct CODEC2 *codec2_state, short speech_out[], const unsigned char *bits, float ber_est); 89void codec2_decode(struct CODEC2 *codec2_state, short speech_out[],
105int codec2_samples_per_frame(struct CODEC2 *codec2_state); 90 const unsigned char bytes[]);
106int codec2_bits_per_frame(struct CODEC2 *codec2_state); 91void codec2_decode_ber(struct CODEC2 *codec2_state, short speech_out[],
107 92 const unsigned char *bytes, float ber_est);
108void codec2_set_lpc_post_filter(struct CODEC2 *codec2_state, int enable, int bass_boost, float beta, float gamma); 93int codec2_samples_per_frame(struct CODEC2 *codec2_state);
109int codec2_get_spare_bit_index(struct CODEC2 *codec2_state); 94int codec2_bits_per_frame(struct CODEC2 *codec2_state);
110int codec2_rebuild_spare_bit(struct CODEC2 *codec2_state, int unpacked_bits[]); 95int codec2_bytes_per_frame(struct CODEC2 *codec2_state);
96
97void codec2_set_lpc_post_filter(struct CODEC2 *codec2_state, int enable,
98 int bass_boost, float beta, float gamma);
99int codec2_get_spare_bit_index(struct CODEC2 *codec2_state);
100int codec2_rebuild_spare_bit(struct CODEC2 *codec2_state, char unpacked_bits[]);
111void codec2_set_natural_or_gray(struct CODEC2 *codec2_state, int gray); 101void codec2_set_natural_or_gray(struct CODEC2 *codec2_state, int gray);
112void codec2_set_softdec(struct CODEC2 *c2, float *softdec); 102void codec2_set_softdec(struct CODEC2 *c2, float *softdec);
113float codec2_get_energy(struct CODEC2 *codec2_state, const unsigned char *bits); 103float codec2_get_energy(struct CODEC2 *codec2_state, const unsigned char *bits);
114 104
115// support for ML and VQ experiments 105// support for ML and VQ experiments
116void codec2_open_mlfeat(struct CODEC2 *codec2_state, char *filename); 106void codec2_open_mlfeat(struct CODEC2 *codec2_state, char *feat_filename,
107 char *model_filename);
117void codec2_load_codebook(struct CODEC2 *codec2_state, int num, char *filename); 108void codec2_load_codebook(struct CODEC2 *codec2_state, int num, char *filename);
118float codec2_get_var(struct CODEC2 *codec2_state); 109float codec2_get_var(struct CODEC2 *codec2_state);
119float *codec2_enable_user_ratek(struct CODEC2 *codec2_state, int *K); 110float *codec2_enable_user_ratek(struct CODEC2 *codec2_state, int *K);
120 111
121// 700C post filter and equaliser 112// 700C post filter and equaliser
122void codec2_700c_post_filter(struct CODEC2 *codec2_state, int en); 113void codec2_700c_post_filter(struct CODEC2 *codec2_state, bool en);
123void codec2_700c_eq(struct CODEC2 *codec2_state, int en); 114void codec2_700c_eq(struct CODEC2 *codec2_state, bool en);
124
125#endif
126 115
127#ifdef __cplusplus 116#ifdef __cplusplus
128} 117}
129#endif 118#endif
130 119
120#endif
diff --git a/codec2_fft.c b/codec2_fft.c
index 14b8670..7a75062 100644
--- a/codec2_fft.c
+++ b/codec2_fft.c
@@ -61,66 +61,63 @@ static const arm_cfft_instance_f32* arm_fft_cache_get(const arm_cfft_instance_f3
61#endif 61#endif
62#endif 62#endif
63 63
64void codec2_fft_free(codec2_fft_cfg cfg) 64void codec2_fft_free(codec2_fft_cfg cfg) {
65{
66#ifdef USE_KISS_FFT 65#ifdef USE_KISS_FFT
67 KISS_FFT_FREE(cfg); 66 KISS_FFT_FREE(cfg);
68#else 67#else
69 FREE(cfg); 68 FREE(cfg);
70#endif 69#endif
71} 70}
72 71
73codec2_fft_cfg codec2_fft_alloc(int nfft, int inverse_fft, void* mem, size_t* lenmem) 72codec2_fft_cfg codec2_fft_alloc(int nfft, int inverse_fft, void* mem,
74{ 73 size_t* lenmem) {
75 codec2_fft_cfg retval; 74 codec2_fft_cfg retval;
76#ifdef USE_KISS_FFT 75#ifdef USE_KISS_FFT
77 retval = kiss_fft_alloc(nfft, inverse_fft, mem, lenmem); 76 retval = kiss_fft_alloc(nfft, inverse_fft, mem, lenmem);
78#else 77#else
79 retval = MALLOC(sizeof(codec2_fft_struct)); 78 retval = MALLOC(sizeof(codec2_fft_struct));
80 retval->inverse = inverse_fft; 79 retval->inverse = inverse_fft;
81 switch(nfft) 80 switch (nfft) {
82 {
83 case 128: 81 case 128:
84 retval->instance = &arm_cfft_sR_f32_len128; 82 retval->instance = &arm_cfft_sR_f32_len128;
85 break; 83 break;
86 case 256: 84 case 256:
87 retval->instance = &arm_cfft_sR_f32_len256; 85 retval->instance = &arm_cfft_sR_f32_len256;
88 break; 86 break;
89 case 512: 87 case 512:
90 retval->instance = &arm_cfft_sR_f32_len512; 88 retval->instance = &arm_cfft_sR_f32_len512;
91 break; 89 break;
92// case 1024: 90 // case 1024:
93// retval->instance = &arm_cfft_sR_f32_len1024; 91 // retval->instance = &arm_cfft_sR_f32_len1024;
94// break; 92 // break;
95 default: 93 default:
96 abort(); 94 abort();
97 } 95 }
98 // retval->instance = arm_fft_cache_get(retval->instance); 96 // retval->instance = arm_fft_cache_get(retval->instance);
99#endif 97#endif
100 return retval; 98 return retval;
101} 99}
102 100
103codec2_fftr_cfg codec2_fftr_alloc(int nfft, int inverse_fft, void* mem, size_t* lenmem) 101codec2_fftr_cfg codec2_fftr_alloc(int nfft, int inverse_fft, void* mem,
104{ 102 size_t* lenmem) {
105 codec2_fftr_cfg retval; 103 codec2_fftr_cfg retval;
106#ifdef USE_KISS_FFT 104#ifdef USE_KISS_FFT
107 retval = kiss_fftr_alloc(nfft, inverse_fft, mem, lenmem); 105 retval = kiss_fftr_alloc(nfft, inverse_fft, mem, lenmem);
108#else 106#else
109 retval = MALLOC(sizeof(codec2_fftr_struct)); 107 retval = MALLOC(sizeof(codec2_fftr_struct));
110 retval->inverse = inverse_fft; 108 retval->inverse = inverse_fft;
111 retval->instance = MALLOC(sizeof(arm_rfft_fast_instance_f32)); 109 retval->instance = MALLOC(sizeof(arm_rfft_fast_instance_f32));
112 arm_rfft_fast_init_f32(retval->instance,nfft); 110 arm_rfft_fast_init_f32(retval->instance, nfft);
113 // memcpy(&retval->instance->Sint,arm_fft_cache_get(&retval->instance->Sint),sizeof(arm_cfft_instance_f32)); 111 // memcpy(&retval->instance->Sint,arm_fft_cache_get(&retval->instance->Sint),sizeof(arm_cfft_instance_f32));
114#endif 112#endif
115 return retval; 113 return retval;
116} 114}
117void codec2_fftr_free(codec2_fftr_cfg cfg) 115void codec2_fftr_free(codec2_fftr_cfg cfg) {
118{
119#ifdef USE_KISS_FFT 116#ifdef USE_KISS_FFT
120 KISS_FFT_FREE(cfg); 117 KISS_FFT_FREE(cfg);
121#else 118#else
122 FREE(cfg->instance); 119 FREE(cfg->instance);
123 FREE(cfg); 120 FREE(cfg);
124#endif 121#endif
125} 122}
126 123
@@ -129,30 +126,25 @@ void codec2_fftr_free(codec2_fftr_cfg cfg)
129// not noticeable 126// not noticeable
130// the reduced usage of RAM and increased performance on STM32 platforms 127// the reduced usage of RAM and increased performance on STM32 platforms
131// should be worth it. 128// should be worth it.
132void codec2_fft_inplace(codec2_fft_cfg cfg, codec2_fft_cpx* inout) 129void codec2_fft_inplace(codec2_fft_cfg cfg, codec2_fft_cpx* inout) {
133{
134
135#ifdef USE_KISS_FFT 130#ifdef USE_KISS_FFT
131 // decide whether to use the local stack based buffer for in
132 // or to allow kiss_fft to allocate RAM
133 // second part is just to play safe since first method
134 // is much faster and uses less RAM
135 if (cfg->nfft <= 512) {
136 kiss_fft_cpx in[512]; 136 kiss_fft_cpx in[512];
137 // decide whether to use the local stack based buffer for in 137 memcpy(in, inout, cfg->nfft * sizeof(kiss_fft_cpx));
138 // or to allow kiss_fft to allocate RAM 138 kiss_fft(cfg, in, (kiss_fft_cpx*)inout);
139 // second part is just to play safe since first method 139 } else {
140 // is much faster and uses less RAM 140 kiss_fft(cfg, (kiss_fft_cpx*)inout, (kiss_fft_cpx*)inout);
141 if (cfg->nfft <= 512) 141 }
142 {
143 memcpy(in,inout,cfg->nfft*sizeof(kiss_fft_cpx));
144 kiss_fft(cfg, in, (kiss_fft_cpx*)inout);
145 }
146 else
147 {
148 kiss_fft(cfg, (kiss_fft_cpx*)inout, (kiss_fft_cpx*)inout);
149 }
150#else 142#else
151 arm_cfft_f32(cfg->instance,(float*)inout,cfg->inverse,1); 143 arm_cfft_f32(cfg->instance, (float*)inout, cfg->inverse, 1);
152 if (cfg->inverse) 144 if (cfg->inverse) {
153 { 145 arm_scale_f32((float*)inout, cfg->instance->fftLen, (float*)inout,
154 arm_scale_f32((float*)inout,cfg->instance->fftLen,(float*)inout,cfg->instance->fftLen*2); 146 cfg->instance->fftLen * 2);
155 } 147 }
156 148
157#endif 149#endif
158} 150}
diff --git a/codec2_fft.h b/codec2_fft.h
index c741202..1952864 100644
--- a/codec2_fft.h
+++ b/codec2_fft.h
@@ -9,96 +9,91 @@
9#define DRIVERS_FREEDV_CODEC2_FFT_H_ 9#define DRIVERS_FREEDV_CODEC2_FFT_H_
10 10
11#include <assert.h> 11#include <assert.h>
12#include <stdlib.h> 12#include <math.h>
13#include <stdio.h> 13#include <stdio.h>
14#include <stdlib.h>
14#include <string.h> 15#include <string.h>
15#include <math.h>
16 16
17#ifdef FDV_ARM_MATH 17#ifndef FDV_ARM_MATH
18 #include "fdv_arm_math.h" 18#define USE_KISS_FFT
19#else 19#else
20 #define USE_KISS_FFT 20#include "arm_const_structs.h"
21#include "arm_math.h"
21#endif 22#endif
22 23
23#include "defines.h"
24#include "comp.h" 24#include "comp.h"
25#include "defines.h"
25 26
26 27typedef COMP codec2_fft_cpx;
27typedef COMP codec2_fft_cpx;
28#include "kiss_fftr.h" 28#include "kiss_fftr.h"
29 29
30#ifdef USE_KISS_FFT 30#ifdef USE_KISS_FFT
31 #include "kiss_fft.h" 31#include "kiss_fft.h"
32 typedef kiss_fftr_cfg codec2_fftr_cfg; 32typedef kiss_fftr_cfg codec2_fftr_cfg;
33 typedef kiss_fft_cfg codec2_fft_cfg; 33typedef kiss_fft_cfg codec2_fft_cfg;
34 typedef kiss_fft_scalar codec2_fft_scalar; 34typedef kiss_fft_scalar codec2_fft_scalar;
35#else 35#else
36 typedef float32_t codec2_fft_scalar; 36typedef float32_t codec2_fft_scalar;
37 typedef struct { 37typedef struct {
38 arm_rfft_fast_instance_f32* instance; 38 arm_rfft_fast_instance_f32* instance;
39 int inverse; 39 int inverse;
40 } codec2_fftr_struct; 40} codec2_fftr_struct;
41 41
42 typedef codec2_fftr_struct* codec2_fftr_cfg; 42typedef codec2_fftr_struct* codec2_fftr_cfg;
43 43
44 typedef struct { 44typedef struct {
45 const arm_cfft_instance_f32* instance; 45 const arm_cfft_instance_f32* instance;
46 int inverse; 46 int inverse;
47 } codec2_fft_struct; 47} codec2_fft_struct;
48 typedef codec2_fft_struct* codec2_fft_cfg; 48typedef codec2_fft_struct* codec2_fft_cfg;
49#endif 49#endif
50 50
51 51static inline void codec2_fftr(codec2_fftr_cfg cfg, codec2_fft_scalar* in,
52 52 codec2_fft_cpx* out) {
53static inline void codec2_fftr(codec2_fftr_cfg cfg, codec2_fft_scalar* in, codec2_fft_cpx* out)
54{
55
56#ifdef USE_KISS_FFT 53#ifdef USE_KISS_FFT
57 kiss_fftr(cfg, in, (kiss_fft_cpx*)out); 54 kiss_fftr(cfg, in, (kiss_fft_cpx*)out);
58#else 55#else
59 arm_rfft_fast_f32(cfg->instance,in,(float*)out,cfg->inverse); 56 arm_rfft_fast_f32(cfg->instance, in, (float*)out, cfg->inverse);
60 out->imag = 0; // remove out[FFT_ENC/2]->real stored in out[0].imag 57 out->imag = 0; // remove out[FFT_ENC/2]->real stored in out[0].imag
61#endif 58#endif
62} 59}
63 60
64static inline void codec2_fftri(codec2_fftr_cfg cfg, codec2_fft_cpx* in, codec2_fft_scalar* out) 61static inline void codec2_fftri(codec2_fftr_cfg cfg, codec2_fft_cpx* in,
65{ 62 codec2_fft_scalar* out) {
66#ifdef USE_KISS_FFT 63#ifdef USE_KISS_FFT
67 kiss_fftri(cfg, (kiss_fft_cpx*)in, out); 64 kiss_fftri(cfg, (kiss_fft_cpx*)in, out);
68#else 65#else
69 arm_rfft_fast_f32(cfg->instance,(float*)in,out,cfg->inverse); 66 arm_rfft_fast_f32(cfg->instance, (float*)in, out, cfg->inverse);
70 // arm_scale_f32(out,cfg->instance->fftLenRFFT,out,cfg->instance->fftLenRFFT); 67 // arm_scale_f32(out,cfg->instance->fftLenRFFT,out,cfg->instance->fftLenRFFT);
71#endif 68#endif
72
73} 69}
74 70
75codec2_fft_cfg codec2_fft_alloc(int nfft, int inverse_fft, void* mem, size_t* lenmem); 71codec2_fft_cfg codec2_fft_alloc(int nfft, int inverse_fft, void* mem,
76codec2_fftr_cfg codec2_fftr_alloc(int nfft, int inverse_fft, void* mem, size_t* lenmem); 72 size_t* lenmem);
73codec2_fftr_cfg codec2_fftr_alloc(int nfft, int inverse_fft, void* mem,
74 size_t* lenmem);
77void codec2_fft_free(codec2_fft_cfg cfg); 75void codec2_fft_free(codec2_fft_cfg cfg);
78void codec2_fftr_free(codec2_fftr_cfg cfg); 76void codec2_fftr_free(codec2_fftr_cfg cfg);
79 77
80 78static inline void codec2_fft(codec2_fft_cfg cfg, codec2_fft_cpx* in,
81static inline void codec2_fft(codec2_fft_cfg cfg, codec2_fft_cpx* in, codec2_fft_cpx* out) 79 codec2_fft_cpx* out) {
82{
83
84#ifdef USE_KISS_FFT 80#ifdef USE_KISS_FFT
85 kiss_fft(cfg, (kiss_fft_cpx*)in, (kiss_fft_cpx*)out); 81 kiss_fft(cfg, (kiss_fft_cpx*)in, (kiss_fft_cpx*)out);
86#else 82#else
87 memcpy(out,in,cfg->instance->fftLen*2*sizeof(float)); 83 memcpy(out, in, cfg->instance->fftLen * 2 * sizeof(float));
88 arm_cfft_f32(cfg->instance,(float*)out,cfg->inverse, 1); 84 arm_cfft_f32(cfg->instance, (float*)out, cfg->inverse, 1);
89 // TODO: this is not nice, but for now required to keep changes minimal 85 // TODO: this is not nice, but for now required to keep changes minimal
90 // however, since main goal is to reduce the memory usage 86 // however, since main goal is to reduce the memory usage
91 // we should convert to an in place interface 87 // we should convert to an in place interface
92 // on PC like platforms the overhead of using the "inplace" kiss_fft calls 88 // on PC like platforms the overhead of using the "inplace" kiss_fft calls
93 // is neglectable compared to the gain in memory usage on STM32 platforms 89 // is neglectable compared to the gain in memory usage on STM32 platforms
94 if (cfg->inverse) 90 if (cfg->inverse) {
95 { 91 arm_scale_f32((float*)out, cfg->instance->fftLen, (float*)out,
96 arm_scale_f32((float*)out,cfg->instance->fftLen,(float*)out,cfg->instance->fftLen*2); 92 cfg->instance->fftLen * 2);
97 } 93 }
98#endif 94#endif
99} 95}
100 96
101void codec2_fft_inplace(codec2_fft_cfg cfg, codec2_fft_cpx* inout); 97void codec2_fft_inplace(codec2_fft_cfg cfg, codec2_fft_cpx* inout);
102 98
103
104#endif 99#endif
diff --git a/codec2_internal.h b/codec2_internal.h
index b46e358..32cd7eb 100644
--- a/codec2_internal.h
+++ b/codec2_internal.h
@@ -28,81 +28,76 @@
28 28
29#ifndef __CODEC2_INTERNAL__ 29#ifndef __CODEC2_INTERNAL__
30#define __CODEC2_INTERNAL__ 30#define __CODEC2_INTERNAL__
31#include <stdbool.h>
31 32
32#include "codec2_fft.h" 33#include "codec2_fft.h"
33#include "newamp1.h" 34#include "newamp1.h"
34#include "newamp2.h"
35 35
36struct CODEC2 { 36struct CODEC2 {
37 int mode; 37 int mode;
38 C2CONST c2const; 38 C2CONST c2const;
39 int Fs; 39 int Fs;
40 int n_samp; 40 int n_samp;
41 int m_pitch; 41 int m_pitch;
42 codec2_fft_cfg fft_fwd_cfg; /* forward FFT config */ 42 codec2_fft_cfg fft_fwd_cfg; /* forward FFT config */
43 codec2_fftr_cfg fftr_fwd_cfg; /* forward real FFT config */ 43 codec2_fftr_cfg fftr_fwd_cfg; /* forward real FFT config */
44 float *w; /* [m_pitch] time domain hamming window */ 44 float *w; /* [m_pitch] time domain hamming window */
45 float W[FFT_ENC]; /* DFT of w[] */ 45 float W[FFT_ENC]; /* DFT of w[] */
46 float *Pn; /* [2*n_samp] trapezoidal synthesis window */ 46 float *Pn; /* [2*n_samp] trapezoidal synthesis window */
47 float *bpf_buf; /* buffer for band pass filter */ 47 float *bpf_buf; /* buffer for band pass filter */
48 float *Sn; /* [m_pitch] input speech */ 48 float *Sn; /* [m_pitch] input speech */
49 float hpf_states[2]; /* high pass filter states */ 49 float hpf_states[2]; /* high pass filter states */
50 void *nlp; /* pitch predictor states */ 50 void *nlp; /* pitch predictor states */
51 int gray; /* non-zero for gray encoding */ 51 int gray; /* non-zero for gray encoding */
52 52
53 codec2_fftr_cfg fftr_inv_cfg; /* inverse FFT config */ 53 codec2_fftr_cfg fftr_inv_cfg; /* inverse FFT config */
54 float *Sn_; /* [2*n_samp] synthesised output speech */ 54 float *Sn_; /* [2*n_samp] synthesised output speech */
55 float ex_phase; /* excitation model phase track */ 55 float ex_phase; /* excitation model phase track */
56 float bg_est; /* background noise estimate for post filter */ 56 float bg_est; /* background noise estimate for post filter */
57 float prev_f0_enc; /* previous frame's f0 estimate */ 57 float prev_f0_enc; /* previous frame's f0 estimate */
58 MODEL prev_model_dec; /* previous frame's model parameters */ 58 MODEL prev_model_dec; /* previous frame's model parameters */
59 float prev_lsps_dec[LPC_ORD]; /* previous frame's LSPs */ 59 float prev_lsps_dec[LPC_ORD]; /* previous frame's LSPs */
60 float prev_e_dec; /* previous frame's LPC energy */ 60 float prev_e_dec; /* previous frame's LPC energy */
61 61
62 int lpc_pf; /* LPC post filter on */ 62 int lpc_pf; /* LPC post filter on */
63 int bass_boost; /* LPC post filter bass boost */ 63 int bass_boost; /* LPC post filter bass boost */
64 float beta; /* LPC post filter parameters */ 64 float beta; /* LPC post filter parameters */
65 float gamma; 65 float gamma;
66 66
67 float xq_enc[2]; /* joint pitch and energy VQ states */ 67 float xq_enc[2]; /* joint pitch and energy VQ states */
68 float xq_dec[2]; 68 float xq_dec[2];
69 69
70 int smoothing; /* enable smoothing for channels with errors */ 70 int smoothing; /* enable smoothing for channels with errors */
71 float *softdec; /* optional soft decn bits from demod */ 71 float *softdec; /* optional soft decn bits from demod */
72 72
73 /* newamp1 states */ 73 /* newamp1 states */
74 74
75 float rate_K_sample_freqs_kHz[NEWAMP1_K]; 75 float rate_K_sample_freqs_kHz[NEWAMP1_K];
76 float prev_rate_K_vec_[NEWAMP1_K]; 76 float prev_rate_K_vec_[NEWAMP1_K];
77 float Wo_left; 77 float Wo_left;
78 int voicing_left; 78 int voicing_left;
79 codec2_fft_cfg phase_fft_fwd_cfg; 79 codec2_fft_cfg phase_fft_fwd_cfg;
80 codec2_fft_cfg phase_fft_inv_cfg; 80 codec2_fft_cfg phase_fft_inv_cfg;
81 float se; /* running sum of squared error */ 81 float se; /* running sum of squared error */
82 unsigned int nse; /* number of terms in sum */ 82 unsigned int nse; /* number of terms in sum */
83 float *user_rate_K_vec_no_mean_; /* optional, user supplied vector for quantisation experiments */ 83 float *user_rate_K_vec_no_mean_; /* optional, user supplied vector for
84 int post_filter_en; 84 quantisation experiments */
85 float eq[NEWAMP1_K]; /* optional equaliser */ 85 bool post_filter_en;
86 int eq_en; 86 float eq[NEWAMP1_K]; /* optional equaliser */
87 87 bool eq_en;
88 /*newamp2 states (also uses newamp1 states )*/ 88
89 float energy_prev; 89 /* used to dump features for deep learning experiments */
90 float n2_rate_K_sample_freqs_kHz[NEWAMP2_K]; 90 FILE *fmlfeat, *fmlmodel;
91 float n2_prev_rate_K_vec_[NEWAMP2_K]; 91
92 float n2_pwb_rate_K_sample_freqs_kHz[NEWAMP2_16K_K]; 92 /* encode/decode function pointers for the selected mode */
93 float n2_pwb_prev_rate_K_vec_[NEWAMP2_16K_K]; 93 void (*encode)(struct CODEC2 *c2, unsigned char *bits, short speech[]);
94 94 void (*decode)(struct CODEC2 *c2, short speech[], const unsigned char *bits);
95 /* used to dump features for deep learning experiments */ 95 void (*decode_ber)(struct CODEC2 *c2, short speech[],
96 FILE *fmlfeat; 96 const unsigned char *bits, float ber_est);
97
98 /* encode/decode function pointers for the selected mode */
99 void (*encode)(struct CODEC2 *c2, unsigned char * bits, short speech[]);
100 void (*decode)(struct CODEC2 *c2, short speech[], const unsigned char * bits);
101 void (*decode_ber)(struct CODEC2 *c2, short speech[], const unsigned char * bits, float ber_est);
102}; 97};
103 98
104// test and debug 99// test and debug
105void analyse_one_frame(struct CODEC2 *c2, MODEL *model, short speech[]); 100void analyse_one_frame(struct CODEC2 *c2, MODEL *model, short speech[]);
106void synthesise_one_frame(struct CODEC2 *c2, short speech[], MODEL *model, 101void synthesise_one_frame(struct CODEC2 *c2, short speech[], MODEL *model,
107 COMP Aw[], float gain); 102 COMP Aw[], float gain);
108#endif 103#endif
diff --git a/comp_prim.h b/comp_prim.h
index d0f070a..e0166d3 100644
--- a/comp_prim.h
+++ b/comp_prim.h
@@ -28,114 +28,111 @@
28#ifndef __COMP_PRIM__ 28#ifndef __COMP_PRIM__
29#define __COMP_PRIM__ 29#define __COMP_PRIM__
30 30
31#include <math.h>
32
31/*---------------------------------------------------------------------------*\ 33/*---------------------------------------------------------------------------*\
32 34
33 FUNCTIONS 35 FUNCTIONS
34 36
35\*---------------------------------------------------------------------------*/ 37\*---------------------------------------------------------------------------*/
36 38
37inline static COMP cneg(COMP a) 39inline static COMP cneg(COMP a) {
38{ 40 COMP res;
39 COMP res;
40 41
41 res.real = -a.real; 42 res.real = -a.real;
42 res.imag = -a.imag; 43 res.imag = -a.imag;
43 44
44 return res; 45 return res;
45} 46}
46 47
47inline static COMP cconj(COMP a) 48inline static COMP cconj(COMP a) {
48{ 49 COMP res;
49 COMP res;
50 50
51 res.real = a.real; 51 res.real = a.real;
52 res.imag = -a.imag; 52 res.imag = -a.imag;
53 53
54 return res; 54 return res;
55} 55}
56 56
57inline static COMP cmult(COMP a, COMP b) 57inline static COMP cmult(COMP a, COMP b) {
58{ 58 COMP res;
59 COMP res;
60 59
61 res.real = a.real*b.real - a.imag*b.imag; 60 res.real = a.real * b.real - a.imag * b.imag;
62 res.imag = a.real*b.imag + a.imag*b.real; 61 res.imag = a.real * b.imag + a.imag * b.real;
63 62
64 return res; 63 return res;
65} 64}
66 65
67inline static COMP fcmult(float a, COMP b) 66inline static COMP fcmult(float a, COMP b) {
68{ 67 COMP res;
69 COMP res;
70 68
71 res.real = a*b.real; 69 res.real = a * b.real;
72 res.imag = a*b.imag; 70 res.imag = a * b.imag;
73 71
74 return res; 72 return res;
75} 73}
76 74
77inline static COMP cadd(COMP a, COMP b) 75inline static COMP cadd(COMP a, COMP b) {
78{ 76 COMP res;
79 COMP res;
80 77
81 res.real = a.real + b.real; 78 res.real = a.real + b.real;
82 res.imag = a.imag + b.imag; 79 res.imag = a.imag + b.imag;
83 80
84 return res; 81 return res;
85} 82}
86 83
87inline static float cabsolute(COMP a) 84inline static float cabsolute(COMP a) {
88{ 85 return sqrtf((a.real * a.real) + (a.imag * a.imag));
89 return sqrtf((a.real * a.real) + (a.imag * a.imag) );
90} 86}
91 87
92/* 88/*
93 * Euler's formula in a new convenient function 89 * Euler's formula in a new convenient function
94 */ 90 */
95inline static COMP comp_exp_j(float phi){ 91inline static COMP comp_exp_j(float phi) {
96 COMP res; 92 COMP res;
97 res.real = cosf(phi); 93 res.real = cosf(phi);
98 res.imag = sinf(phi); 94 res.imag = sinf(phi);
99 return res; 95 return res;
100} 96}
101 97
102/* 98/*
103 * Quick and easy complex 0 99 * Quick and easy complex 0
104 */ 100 */
105inline static COMP comp0(){ 101inline static COMP comp0() {
106 COMP res; 102 COMP res;
107 res.real = 0; 103 res.real = 0;
108 res.imag = 0; 104 res.imag = 0;
109 return res; 105 return res;
110} 106}
111 107
112/* 108/*
113 * Quick and easy complex subtract 109 * Quick and easy complex subtract
114 */ 110 */
115inline static COMP csub(COMP a, COMP b){ 111inline static COMP csub(COMP a, COMP b) {
116 COMP res; 112 COMP res;
117 res.real = a.real-b.real; 113 res.real = a.real - b.real;
118 res.imag = a.imag-b.imag; 114 res.imag = a.imag - b.imag;
119 return res; 115 return res;
120} 116}
121 117
122/* 118/*
123 * Compare the magnitude of a and b. if |a|>|b|, return true, otw false. 119 * Compare the magnitude of a and b. if |a|>|b|, return true, otw false.
124 * This needs no square roots 120 * This needs no square roots
125 */ 121 */
126inline static int comp_mag_gt(COMP a,COMP b){ 122inline static int comp_mag_gt(COMP a, COMP b) {
127 return ((a.real*a.real)+(a.imag*a.imag)) > ((b.real*b.real)+(b.imag*b.imag)); 123 return ((a.real * a.real) + (a.imag * a.imag)) >
124 ((b.real * b.real) + (b.imag * b.imag));
128} 125}
129 126
130/* 127/*
131 * Normalize a complex number's magnitude to 1 128 * Normalize a complex number's magnitude to 1
132 */ 129 */
133inline static COMP comp_normalize(COMP a){ 130inline static COMP comp_normalize(COMP a) {
134 COMP b; 131 COMP b;
135 float av = cabsolute(a); 132 float av = cabsolute(a);
136 b.real = a.real/av; 133 b.real = a.real / av;
137 b.imag = a.imag/av; 134 b.imag = a.imag / av;
138 return b; 135 return b;
139} 136}
140 137
141#endif 138#endif
diff --git a/defines.h b/defines.h
index dcd1841..a2fd9ec 100644
--- a/defines.h
+++ b/defines.h
@@ -30,96 +30,93 @@
30 30
31/*---------------------------------------------------------------------------*\ 31/*---------------------------------------------------------------------------*\
32 32
33 DEFINES 33 DEFINES
34 34
35\*---------------------------------------------------------------------------*/ 35\*---------------------------------------------------------------------------*/
36 36
37/* General defines */ 37/* General defines */
38 38
39#define N_S 0.01 /* internal proc frame length in secs */ 39#define N_S 0.01 /* internal proc frame length in secs */
40#define TW_S 0.005 /* trapezoidal synth window overlap */ 40#define TW_S 0.005 /* trapezoidal synth window overlap */
41#define MAX_AMP 160 /* maximum number of harmonics */ 41#define MAX_AMP 160 /* maximum number of harmonics */
42#ifndef PI 42#ifndef PI
43#define PI 3.141592654 /* mathematical constant */ 43#define PI 3.141592654 /* mathematical constant */
44#endif 44#endif
45#define TWO_PI 6.283185307 /* mathematical constant */ 45#ifndef M_PI
46#define MAX_STR 2048 /* maximum string size */ 46#define M_PI 3.14159265358979323846f
47#endif
48#define TWO_PI 6.283185307 /* mathematical constant */
49#define MAX_STR 2048 /* maximum string size */
47 50
48#define FFT_ENC 512 /* size of FFT used for encoder */ 51#define FFT_ENC 512 /* size of FFT used for encoder */
49#define FFT_DEC 512 /* size of FFT used in decoder */ 52#define FFT_DEC 512 /* size of FFT used in decoder */
50#define V_THRESH 6.0 /* voicing threshold in dB */ 53#define V_THRESH 6.0 /* voicing threshold in dB */
51#define LPC_ORD 10 /* LPC order */ 54#define LPC_ORD 10 /* LPC order */
52#define LPC_ORD_LOW 6 /* LPC order for lower rates */ 55#define LPC_ORD_LOW 6 /* LPC order for lower rates */
53 56
54/* Pitch estimation defines */ 57/* Pitch estimation defines */
55 58
56#define M_PITCH_S 0.0400 /* pitch analysis window in s */ 59#define M_PITCH_S 0.0400 /* pitch analysis window in s */
57#define P_MIN_S 0.0025 /* minimum pitch period in s */ 60#define P_MIN_S 0.0025 /* minimum pitch period in s */
58#define P_MAX_S 0.0200 /* maximum pitch period in s */ 61#define P_MAX_S 0.0200 /* maximum pitch period in s */
59 62
60/*---------------------------------------------------------------------------*\ 63/*---------------------------------------------------------------------------*\
61 64
62 TYPEDEFS 65 TYPEDEFS
63 66
64\*---------------------------------------------------------------------------*/ 67\*---------------------------------------------------------------------------*/
65 68
66/* Structure to hold constants calculated at run time based on sample rate */ 69/* Structure to hold constants calculated at run time based on sample rate */
67 70
68typedef struct { 71typedef struct {
69 int Fs; /* sample rate of this instance */ 72 int Fs; /* sample rate of this instance */
70 int n_samp; /* number of samples per 10ms frame at Fs */ 73 int n_samp; /* number of samples per 10ms frame at Fs */
71 int max_amp; /* maximum number of harmonics */ 74 int max_amp; /* maximum number of harmonics */
72 int m_pitch; /* pitch estimation window size in samples */ 75 int m_pitch; /* pitch estimation window size in samples */
73 int p_min; /* minimum pitch period in samples */ 76 int p_min; /* minimum pitch period in samples */
74 int p_max; /* maximum pitch period in samples */ 77 int p_max; /* maximum pitch period in samples */
75 float Wo_min; 78 float Wo_min;
76 float Wo_max; 79 float Wo_max;
77 int nw; /* analysis window size in samples */ 80 int nw; /* analysis window size in samples */
78 int tw; /* trapezoidal synthesis window overlap */ 81 int tw; /* trapezoidal synthesis window overlap */
79} C2CONST; 82} C2CONST;
80 83
81/* Structure to hold model parameters for one frame */ 84/* Structure to hold model parameters for one frame */
82 85
83typedef struct { 86typedef struct {
84 float Wo; /* fundamental frequency estimate in radians */ 87 float Wo; /* fundamental frequency estimate in radians */
85 int L; /* number of harmonics */ 88 int L; /* number of harmonics */
86 float A[MAX_AMP+1]; /* amplitiude of each harmonic */ 89 float A[MAX_AMP + 1]; /* amplitiude of each harmonic */
87 float phi[MAX_AMP+1]; /* phase of each harmonic */ 90 float phi[MAX_AMP + 1]; /* phase of each harmonic */
88 int voiced; /* non-zero if this frame is voiced */ 91 int voiced; /* non-zero if this frame is voiced */
89} MODEL; 92} MODEL;
90 93
91/* describes each codebook */ 94/* describes each codebook */
92 95
93struct lsp_codebook { 96struct lsp_codebook {
94 int k; /* dimension of vector */ 97 int k; /* dimension of vector */
95 int log2m; /* number of bits in m */ 98 int log2m; /* number of bits in m */
96 int m; /* elements in codebook */ 99 int m; /* elements in codebook */
97#ifdef __EMBEDDED /* make sure stored in flash */ 100#ifdef __EMBEDDED__ /* make sure stored in flash */
98 const float *cb; /* The elements */ 101 const float *cb; /* The elements */
99#else 102#else
100 float *cb; /* The elements */ 103 float *cb; /* The elements */
101#endif 104#endif
102}; 105};
103 106
104extern const struct lsp_codebook lsp_cb[]; 107extern const struct lsp_codebook lsp_cb[];
105extern const struct lsp_codebook lsp_cbd[]; 108extern const struct lsp_codebook lsp_cbd[];
106extern const struct lsp_codebook lsp_cbvq[]; 109extern const struct lsp_codebook lsp_cbjmv[];
107extern const struct lsp_codebook lsp_cbjnd[];
108extern const struct lsp_codebook lsp_cbdt[];
109extern const struct lsp_codebook lsp_cbjvm[];
110extern const struct lsp_codebook lsp_cbvqanssi[];
111extern const struct lsp_codebook mel_cb[];
112extern const struct lsp_codebook ge_cb[]; 110extern const struct lsp_codebook ge_cb[];
113extern const struct lsp_codebook lspmelvq_cb[];
114extern const struct lsp_codebook newamp1vq_cb[]; 111extern const struct lsp_codebook newamp1vq_cb[];
115extern const struct lsp_codebook newamp1_energy_cb[]; 112extern const struct lsp_codebook newamp1_energy_cb[];
116extern const struct lsp_codebook newamp2vq_cb[]; 113extern const struct lsp_codebook newamp2vq_cb[];
117extern const struct lsp_codebook newamp2_energy_cb[]; 114extern const struct lsp_codebook newamp2_energy_cb[];
118 115
119#ifdef _GNU_SOURCE 116#ifdef _GNU_SOURCE
120 #define POW10F(x) exp10f((x)) 117#define POW10F(x) exp10f((x))
121#else 118#else
122 #define POW10F(x) expf(2.302585092994046f*(x)) 119#define POW10F(x) expf(2.302585092994046f * (x))
123#endif 120#endif
124 121
125#endif 122#endif
diff --git a/interp.c b/interp.c
index 2109acb..08a9518 100644
--- a/interp.c
+++ b/interp.c
@@ -25,13 +25,14 @@
25 along with this program; if not, see <http://www.gnu.org/licenses/>. 25 along with this program; if not, see <http://www.gnu.org/licenses/>.
26*/ 26*/
27 27
28#include "interp.h"
29
28#include <assert.h> 30#include <assert.h>
29#include <math.h> 31#include <math.h>
30#include <string.h>
31#include <stdio.h> 32#include <stdio.h>
33#include <string.h>
32 34
33#include "defines.h" 35#include "defines.h"
34#include "interp.h"
35#include "lsp.h" 36#include "lsp.h"
36#include "quantise.h" 37#include "quantise.h"
37 38
@@ -44,7 +45,7 @@ float sample_log_amp(MODEL *model, float w);
44 AUTHOR......: David Rowe 45 AUTHOR......: David Rowe
45 DATE CREATED: 22/8/10 46 DATE CREATED: 22/8/10
46 47
47 Given two frames decribed by model parameters 20ms apart, determines 48 Given two frames described by model parameters 20ms apart, determines
48 the model parameters of the 10ms frame between them. Assumes 49 the model parameters of the 10ms frame between them. Assumes
49 voicing is available for middle (interpolated) frame. Outputs are 50 voicing is available for middle (interpolated) frame. Outputs are
50 amplitudes and Wo for the interpolated frame. 51 amplitudes and Wo for the interpolated frame.
@@ -107,29 +108,27 @@ void interpolate(
107 108
108\*---------------------------------------------------------------------------*/ 109\*---------------------------------------------------------------------------*/
109 110
110float sample_log_amp(MODEL *model, float w) 111float sample_log_amp(MODEL *model, float w) {
111{ 112 int m;
112 int m; 113 float f, log_amp;
113 float f, log_amp;
114 114
115 assert(w > 0.0); assert (w <= PI); 115 assert(w > 0.0);
116 assert(w <= PI);
116 117
117 m = floorf(w/model->Wo + 0.5); 118 m = floorf(w / model->Wo + 0.5);
118 f = (w - m*model->Wo)/w; 119 f = (w - m * model->Wo) / w;
119 assert(f <= 1.0); 120 assert(f <= 1.0);
120 121
121 if (m < 1) { 122 if (m < 1) {
122 log_amp = f*log10f(model->A[1] + 1E-6); 123 log_amp = f * log10f(model->A[1] + 1E-6);
123 } 124 } else if ((m + 1) > model->L) {
124 else if ((m+1) > model->L) { 125 log_amp = (1.0 - f) * log10f(model->A[model->L] + 1E-6);
125 log_amp = (1.0-f)*log10f(model->A[model->L] + 1E-6); 126 } else {
126 } 127 log_amp = (1.0 - f) * log10f(model->A[m] + 1E-6) +
127 else { 128 f * log10f(model->A[m + 1] + 1E-6);
128 log_amp = (1.0-f)*log10f(model->A[m] + 1E-6) + 129 }
129 f*log10f(model->A[m+1] + 1E-6);
130 }
131 130
132 return log_amp; 131 return log_amp;
133} 132}
134 133
135#ifdef NOT_NEEDED 134#ifdef NOT_NEEDED
@@ -140,7 +139,7 @@ float sample_log_amp(MODEL *model, float w)
140 AUTHOR......: David Rowe 139 AUTHOR......: David Rowe
141 DATE CREATED: 10 Nov 2010 140 DATE CREATED: 10 Nov 2010
142 141
143 Given two frames decribed by model parameters 20ms apart, determines 142 Given two frames described by model parameters 20ms apart, determines
144 the model parameters of the 10ms frame between them. Assumes 143 the model parameters of the 10ms frame between them. Assumes
145 voicing is available for middle (interpolated) frame. Outputs are 144 voicing is available for middle (interpolated) frame. Outputs are
146 amplitudes and Wo for the interpolated frame. 145 amplitudes and Wo for the interpolated frame.
@@ -150,65 +149,59 @@ float sample_log_amp(MODEL *model, float w)
150 149
151\*---------------------------------------------------------------------------*/ 150\*---------------------------------------------------------------------------*/
152 151
153void interpolate_lsp( 152void interpolate_lsp(codec2_fft_cfg fft_fwd_cfg,
154 codec2_fft_cfg fft_fwd_cfg, 153 MODEL *interp, /* interpolated model params */
155 MODEL *interp, /* interpolated model params */ 154 MODEL *prev, /* previous frames model params */
156 MODEL *prev, /* previous frames model params */ 155 MODEL *next, /* next frames model params */
157 MODEL *next, /* next frames model params */ 156 float *prev_lsps, /* previous frames LSPs */
158 float *prev_lsps, /* previous frames LSPs */ 157 float prev_e, /* previous frames LPC energy */
159 float prev_e, /* previous frames LPC energy */ 158 float *next_lsps, /* next frames LSPs */
160 float *next_lsps, /* next frames LSPs */ 159 float next_e, /* next frames LPC energy */
161 float next_e, /* next frames LPC energy */ 160 float *ak_interp, /* interpolated aks for this frame */
162 float *ak_interp, /* interpolated aks for this frame */ 161 float *lsps_interp, /* interpolated lsps for this frame */
163 float *lsps_interp, /* interpolated lsps for this frame */ 162 float Wo_min) {
164 float Wo_min 163 int i;
165) 164 float e;
166{ 165 float snr;
167 int i; 166
168 float e; 167 /* trap corner case where V est is probably wrong */
169 float snr; 168
170 169 if (interp->voiced && !prev->voiced && !next->voiced) {
171 /* trap corner case where V est is probably wrong */ 170 interp->voiced = 0;
172 171 }
173 if (interp->voiced && !prev->voiced && !next->voiced) { 172
174 interp->voiced = 0; 173 /* Wo depends on voicing of this and adjacent frames */
175 } 174
176 175 if (interp->voiced) {
177 /* Wo depends on voicing of this and adjacent frames */ 176 if (prev->voiced && next->voiced) interp->Wo = (prev->Wo + next->Wo) / 2.0;
178 177 if (!prev->voiced && next->voiced) interp->Wo = next->Wo;
179 if (interp->voiced) { 178 if (prev->voiced && !next->voiced) interp->Wo = prev->Wo;
180 if (prev->voiced && next->voiced) 179 } else {
181 interp->Wo = (prev->Wo + next->Wo)/2.0; 180 interp->Wo = Wo_min;
182 if (!prev->voiced && next->voiced) 181 }
183 interp->Wo = next->Wo; 182 interp->L = PI / interp->Wo;
184 if (prev->voiced && !next->voiced) 183
185 interp->Wo = prev->Wo; 184 // printf(" interp: prev_v: %d next_v: %d prev_Wo: %f next_Wo: %f\n",
186 } 185 // prev->voiced, next->voiced, prev->Wo, next->Wo);
187 else { 186 // printf(" interp: Wo: %1.5f L: %d\n", interp->Wo, interp->L);
188 interp->Wo = Wo_min; 187
189 } 188 /* interpolate LSPs */
190 interp->L = PI/interp->Wo; 189
191 190 for (i = 0; i < LPC_ORD; i++) {
192 //printf(" interp: prev_v: %d next_v: %d prev_Wo: %f next_Wo: %f\n", 191 lsps_interp[i] = (prev_lsps[i] + next_lsps[i]) / 2.0;
193 // prev->voiced, next->voiced, prev->Wo, next->Wo); 192 }
194 //printf(" interp: Wo: %1.5f L: %d\n", interp->Wo, interp->L); 193
195 194 /* Interpolate LPC energy in log domain */
196 /* interpolate LSPs */ 195
197 196 e = powf(10.0, (log10f(prev_e) + log10f(next_e)) / 2.0);
198 for(i=0; i<LPC_ORD; i++) { 197 // printf(" interp: e: %f\n", e);
199 lsps_interp[i] = (prev_lsps[i] + next_lsps[i])/2.0; 198
200 } 199 /* convert back to amplitudes */
201 200
202 /* Interpolate LPC energy in log domain */ 201 lsp_to_lpc(lsps_interp, ak_interp, LPC_ORD);
203 202 aks_to_M2(fft_fwd_cfg, ak_interp, LPC_ORD, interp, e, &snr, 0, 0, 1, 1,
204 e = powf(10.0, (log10f(prev_e) + log10f(next_e))/2.0); 203 LPCPF_BETA, LPCPF_GAMMA);
205 //printf(" interp: e: %f\n", e); 204 // printf(" interp: ak[1]: %f A[1] %f\n", ak_interp[1], interp->A[1]);
206
207 /* convert back to amplitudes */
208
209 lsp_to_lpc(lsps_interp, ak_interp, LPC_ORD);
210 aks_to_M2(fft_fwd_cfg, ak_interp, LPC_ORD, interp, e, &snr, 0, 0, 1, 1, LPCPF_BETA, LPCPF_GAMMA);
211 //printf(" interp: ak[1]: %f A[1] %f\n", ak_interp[1], interp->A[1]);
212} 205}
213#endif 206#endif
214 207
@@ -224,14 +217,11 @@ void interpolate_lsp(
224 217
225\*---------------------------------------------------------------------------*/ 218\*---------------------------------------------------------------------------*/
226 219
227void interp_Wo( 220void interp_Wo(MODEL *interp, /* interpolated model params */
228 MODEL *interp, /* interpolated model params */ 221 MODEL *prev, /* previous frames model params */
229 MODEL *prev, /* previous frames model params */ 222 MODEL *next, /* next frames model params */
230 MODEL *next, /* next frames model params */ 223 float Wo_min) {
231 float Wo_min 224 interp_Wo2(interp, prev, next, 0.5, Wo_min);
232 )
233{
234 interp_Wo2(interp, prev, next, 0.5, Wo_min);
235} 225}
236 226
237/*---------------------------------------------------------------------------*\ 227/*---------------------------------------------------------------------------*\
@@ -244,37 +234,29 @@ void interp_Wo(
244 234
245\*---------------------------------------------------------------------------*/ 235\*---------------------------------------------------------------------------*/
246 236
247void interp_Wo2( 237void interp_Wo2(MODEL *interp, /* interpolated model params */
248 MODEL *interp, /* interpolated model params */ 238 MODEL *prev, /* previous frames model params */
249 MODEL *prev, /* previous frames model params */ 239 MODEL *next, /* next frames model params */
250 MODEL *next, /* next frames model params */ 240 float weight, float Wo_min) {
251 float weight, 241 /* trap corner case where voicing est is probably wrong */
252 float Wo_min 242
253) 243 if (interp->voiced && !prev->voiced && !next->voiced) {
254{ 244 interp->voiced = 0;
255 /* trap corner case where voicing est is probably wrong */ 245 }
256 246
257 if (interp->voiced && !prev->voiced && !next->voiced) { 247 /* Wo depends on voicing of this and adjacent frames */
258 interp->voiced = 0; 248
259 } 249 if (interp->voiced) {
260 250 if (prev->voiced && next->voiced)
261 /* Wo depends on voicing of this and adjacent frames */ 251 interp->Wo = (1.0 - weight) * prev->Wo + weight * next->Wo;
262 252 if (!prev->voiced && next->voiced) interp->Wo = next->Wo;
263 if (interp->voiced) { 253 if (prev->voiced && !next->voiced) interp->Wo = prev->Wo;
264 if (prev->voiced && next->voiced) 254 } else {
265 interp->Wo = (1.0 - weight)*prev->Wo + weight*next->Wo; 255 interp->Wo = Wo_min;
266 if (!prev->voiced && next->voiced) 256 }
267 interp->Wo = next->Wo; 257 interp->L = PI / interp->Wo;
268 if (prev->voiced && !next->voiced)
269 interp->Wo = prev->Wo;
270 }
271 else {
272 interp->Wo = Wo_min;
273 }
274 interp->L = PI/interp->Wo;
275} 258}
276 259
277
278/*---------------------------------------------------------------------------*\ 260/*---------------------------------------------------------------------------*\
279 261
280 FUNCTION....: interp_energy() 262 FUNCTION....: interp_energy()
@@ -286,13 +268,12 @@ void interp_Wo2(
286 268
287\*---------------------------------------------------------------------------*/ 269\*---------------------------------------------------------------------------*/
288 270
289float interp_energy(float prev_e, float next_e) 271float interp_energy(float prev_e, float next_e) {
290{ 272 // return powf(10.0, (log10f(prev_e) + log10f(next_e))/2.0);
291 //return powf(10.0, (log10f(prev_e) + log10f(next_e))/2.0); 273 return sqrtf(prev_e *
292 return sqrtf(prev_e * next_e); //looks better is math. identical and faster math 274 next_e); // looks better is math. identical and faster math
293} 275}
294 276
295
296/*---------------------------------------------------------------------------*\ 277/*---------------------------------------------------------------------------*\
297 278
298 FUNCTION....: interp_energy2() 279 FUNCTION....: interp_energy2()
@@ -304,13 +285,10 @@ float interp_energy(float prev_e, float next_e)
304 285
305\*---------------------------------------------------------------------------*/ 286\*---------------------------------------------------------------------------*/
306 287
307float interp_energy2(float prev_e, float next_e, float weight) 288float interp_energy2(float prev_e, float next_e, float weight) {
308{ 289 return POW10F((1.0 - weight) * log10f(prev_e) + weight * log10f(next_e));
309 return POW10F((1.0 - weight)*log10f(prev_e) + weight*log10f(next_e));
310
311} 290}
312 291
313
314/*---------------------------------------------------------------------------*\ 292/*---------------------------------------------------------------------------*\
315 293
316 FUNCTION....: interpolate_lsp_ver2() 294 FUNCTION....: interpolate_lsp_ver2()
@@ -321,11 +299,10 @@ float interp_energy2(float prev_e, float next_e, float weight)
321 299
322\*---------------------------------------------------------------------------*/ 300\*---------------------------------------------------------------------------*/
323 301
324void interpolate_lsp_ver2(float interp[], float prev[], float next[], float weight, int order) 302void interpolate_lsp_ver2(float interp[], float prev[], float next[],
325{ 303 float weight, int order) {
326 int i; 304 int i;
327 305
328 for(i=0; i<order; i++) 306 for (i = 0; i < order; i++)
329 interp[i] = (1.0 - weight)*prev[i] + weight*next[i]; 307 interp[i] = (1.0 - weight) * prev[i] + weight * next[i];
330} 308}
331
diff --git a/interp.h b/interp.h
index 1247b7e..276edd9 100644
--- a/interp.h
+++ b/interp.h
@@ -28,18 +28,20 @@
28#ifndef __INTERP__ 28#ifndef __INTERP__
29#define __INTERP__ 29#define __INTERP__
30 30
31#include "defines.h"
31#include "kiss_fft.h" 32#include "kiss_fft.h"
32 33
33void interpolate(MODEL *interp, MODEL *prev, MODEL *next); 34void interpolate(MODEL *interp, MODEL *prev, MODEL *next);
34void interpolate_lsp(kiss_fft_cfg fft_dec_cfg, 35void interpolate_lsp(kiss_fft_cfg fft_dec_cfg, MODEL *interp, MODEL *prev,
35 MODEL *interp, MODEL *prev, MODEL *next, 36 MODEL *next, float *prev_lsps, float prev_e,
36 float *prev_lsps, float prev_e, 37 float *next_lsps, float next_e, float *ak_interp,
37 float *next_lsps, float next_e, 38 float *lsps_interp, float Wo_min);
38 float *ak_interp, float *lsps_interp, float Wo_min);
39void interp_Wo(MODEL *interp, MODEL *prev, MODEL *next, float Wo_min); 39void interp_Wo(MODEL *interp, MODEL *prev, MODEL *next, float Wo_min);
40void interp_Wo2(MODEL *interp, MODEL *prev, MODEL *next, float weight, float Wo_min); 40void interp_Wo2(MODEL *interp, MODEL *prev, MODEL *next, float weight,
41 float Wo_min);
41float interp_energy(float prev, float next); 42float interp_energy(float prev, float next);
42float interp_energy2(float prev, float next, float weight); 43float interp_energy2(float prev, float next, float weight);
43void interpolate_lsp_ver2(float interp[], float prev[], float next[], float weight, int order); 44void interpolate_lsp_ver2(float interp[], float prev[], float next[],
45 float weight, int order);
44 46
45#endif 47#endif
diff --git a/kiss_fft.c b/kiss_fft.c
index 05f7f27..e7993cf 100644
--- a/kiss_fft.c
+++ b/kiss_fft.c
@@ -3,406 +3,424 @@ Copyright (c) 2003-2010, Mark Borgerding
3 3
4All rights reserved. 4All rights reserved.
5 5
6Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: 6Redistribution and use in source and binary forms, with or without modification,
7 7are permitted provided that the following conditions are met:
8 * Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. 8
9 * Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. 9 * Redistributions of source code must retain the above copyright notice,
10 * Neither the author nor the names of any contributors may be used to endorse or promote products derived from this software without specific prior written permission. 10this list of conditions and the following disclaimer.
11 11 * Redistributions in binary form must reproduce the above copyright notice,
12THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 12this list of conditions and the following disclaimer in the documentation and/or
13other materials provided with the distribution.
14 * Neither the author nor the names of any contributors may be used to
15endorse or promote products derived from this software without specific prior
16written permission.
17
18THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
19ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
20WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
21DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
22ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
23(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
24LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
25ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
27SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
13*/ 28*/
14 29
15
16#include "_kiss_fft_guts.h" 30#include "_kiss_fft_guts.h"
17/* The guts header contains all the multiplication and addition macros that are defined for 31/* The guts header contains all the multiplication and addition macros that are
18 fixed or floating point complex numbers. It also delares the kf_ internal functions. 32 defined for fixed or floating point complex numbers. It also declares the kf_
33 internal functions.
19 */ 34 */
20 35
21static void kf_bfly2( 36static void kf_bfly2(kiss_fft_cpx *Fout, const size_t fstride,
22 kiss_fft_cpx * Fout, 37 const kiss_fft_cfg st, int m) {
23 const size_t fstride, 38 kiss_fft_cpx *Fout2;
24 const kiss_fft_cfg st, 39 kiss_fft_cpx *tw1 = st->twiddles;
25 int m 40 kiss_fft_cpx t;
26 ) 41 Fout2 = Fout + m;
27{ 42 do {
28 kiss_fft_cpx * Fout2; 43 C_FIXDIV(*Fout, 2);
29 kiss_fft_cpx * tw1 = st->twiddles; 44 C_FIXDIV(*Fout2, 2);
30 kiss_fft_cpx t; 45
31 Fout2 = Fout + m; 46 C_MUL(t, *Fout2, *tw1);
32 do{ 47 tw1 += fstride;
33 C_FIXDIV(*Fout,2); C_FIXDIV(*Fout2,2); 48 C_SUB(*Fout2, *Fout, t);
34 49 C_ADDTO(*Fout, t);
35 C_MUL (t, *Fout2 , *tw1); 50 ++Fout2;
36 tw1 += fstride; 51 ++Fout;
37 C_SUB( *Fout2 , *Fout , t ); 52 } while (--m);
38 C_ADDTO( *Fout , t );
39 ++Fout2;
40 ++Fout;
41 }while (--m);
42} 53}
43 54
44static void kf_bfly4( 55static void kf_bfly4(kiss_fft_cpx *Fout, const size_t fstride,
45 kiss_fft_cpx * Fout, 56 const kiss_fft_cfg st, const size_t m) {
46 const size_t fstride, 57 kiss_fft_cpx *tw1, *tw2, *tw3;
47 const kiss_fft_cfg st, 58 kiss_fft_cpx scratch[6];
48 const size_t m 59 size_t k = m;
49 ) 60 const size_t m2 = 2 * m;
50{ 61 const size_t m3 = 3 * m;
51 kiss_fft_cpx *tw1,*tw2,*tw3; 62
52 kiss_fft_cpx scratch[6]; 63 tw3 = tw2 = tw1 = st->twiddles;
53 size_t k=m; 64
54 const size_t m2=2*m; 65 do {
55 const size_t m3=3*m; 66 C_FIXDIV(*Fout, 4);
56 67 C_FIXDIV(Fout[m], 4);
57 68 C_FIXDIV(Fout[m2], 4);
58 tw3 = tw2 = tw1 = st->twiddles; 69 C_FIXDIV(Fout[m3], 4);
59 70
60 do { 71 C_MUL(scratch[0], Fout[m], *tw1);
61 C_FIXDIV(*Fout,4); C_FIXDIV(Fout[m],4); C_FIXDIV(Fout[m2],4); C_FIXDIV(Fout[m3],4); 72 C_MUL(scratch[1], Fout[m2], *tw2);
62 73 C_MUL(scratch[2], Fout[m3], *tw3);
63 C_MUL(scratch[0],Fout[m] , *tw1 ); 74
64 C_MUL(scratch[1],Fout[m2] , *tw2 ); 75 C_SUB(scratch[5], *Fout, scratch[1]);
65 C_MUL(scratch[2],Fout[m3] , *tw3 ); 76 C_ADDTO(*Fout, scratch[1]);
66 77 C_ADD(scratch[3], scratch[0], scratch[2]);
67 C_SUB( scratch[5] , *Fout, scratch[1] ); 78 C_SUB(scratch[4], scratch[0], scratch[2]);
68 C_ADDTO(*Fout, scratch[1]); 79 C_SUB(Fout[m2], *Fout, scratch[3]);
69 C_ADD( scratch[3] , scratch[0] , scratch[2] ); 80 tw1 += fstride;
70 C_SUB( scratch[4] , scratch[0] , scratch[2] ); 81 tw2 += fstride * 2;
71 C_SUB( Fout[m2], *Fout, scratch[3] ); 82 tw3 += fstride * 3;
72 tw1 += fstride; 83 C_ADDTO(*Fout, scratch[3]);
73 tw2 += fstride*2; 84
74 tw3 += fstride*3; 85 if (st->inverse) {
75 C_ADDTO( *Fout , scratch[3] ); 86 Fout[m].r = scratch[5].r - scratch[4].i;
76 87 Fout[m].i = scratch[5].i + scratch[4].r;
77 if(st->inverse) { 88 Fout[m3].r = scratch[5].r + scratch[4].i;
78 Fout[m].r = scratch[5].r - scratch[4].i; 89 Fout[m3].i = scratch[5].i - scratch[4].r;
79 Fout[m].i = scratch[5].i + scratch[4].r; 90 } else {
80 Fout[m3].r = scratch[5].r + scratch[4].i; 91 Fout[m].r = scratch[5].r + scratch[4].i;
81 Fout[m3].i = scratch[5].i - scratch[4].r; 92 Fout[m].i = scratch[5].i - scratch[4].r;
82 }else{ 93 Fout[m3].r = scratch[5].r - scratch[4].i;
83 Fout[m].r = scratch[5].r + scratch[4].i; 94 Fout[m3].i = scratch[5].i + scratch[4].r;
84 Fout[m].i = scratch[5].i - scratch[4].r; 95 }
85 Fout[m3].r = scratch[5].r - scratch[4].i; 96 ++Fout;
86 Fout[m3].i = scratch[5].i + scratch[4].r; 97 } while (--k);
87 }
88 ++Fout;
89 }while(--k);
90} 98}
91 99
92static void kf_bfly3( 100static void kf_bfly3(kiss_fft_cpx *Fout, const size_t fstride,
93 kiss_fft_cpx * Fout, 101 const kiss_fft_cfg st, size_t m) {
94 const size_t fstride, 102 size_t k = m;
95 const kiss_fft_cfg st, 103 const size_t m2 = 2 * m;
96 size_t m 104 kiss_fft_cpx *tw1, *tw2;
97 ) 105 kiss_fft_cpx scratch[5];
98{ 106 kiss_fft_cpx epi3;
99 size_t k=m; 107 epi3 = st->twiddles[fstride * m];
100 const size_t m2 = 2*m;
101 kiss_fft_cpx *tw1,*tw2;
102 kiss_fft_cpx scratch[5];
103 kiss_fft_cpx epi3;
104 epi3 = st->twiddles[fstride*m];
105 108
106 tw1=tw2=st->twiddles; 109 tw1 = tw2 = st->twiddles;
107 110
108 do{ 111 do {
109 C_FIXDIV(*Fout,3); C_FIXDIV(Fout[m],3); C_FIXDIV(Fout[m2],3); 112 C_FIXDIV(*Fout, 3);
113 C_FIXDIV(Fout[m], 3);
114 C_FIXDIV(Fout[m2], 3);
110 115
111 C_MUL(scratch[1],Fout[m] , *tw1); 116 C_MUL(scratch[1], Fout[m], *tw1);
112 C_MUL(scratch[2],Fout[m2] , *tw2); 117 C_MUL(scratch[2], Fout[m2], *tw2);
113 118
114 C_ADD(scratch[3],scratch[1],scratch[2]); 119 C_ADD(scratch[3], scratch[1], scratch[2]);
115 C_SUB(scratch[0],scratch[1],scratch[2]); 120 C_SUB(scratch[0], scratch[1], scratch[2]);
116 tw1 += fstride; 121 tw1 += fstride;
117 tw2 += fstride*2; 122 tw2 += fstride * 2;
118 123
119 Fout[m].r = Fout->r - HALF_OF(scratch[3].r); 124 Fout[m].r = Fout->r - HALF_OF(scratch[3].r);
120 Fout[m].i = Fout->i - HALF_OF(scratch[3].i); 125 Fout[m].i = Fout->i - HALF_OF(scratch[3].i);
121 126
122 C_MULBYSCALAR( scratch[0] , epi3.i ); 127 C_MULBYSCALAR(scratch[0], epi3.i);
123 128
124 C_ADDTO(*Fout,scratch[3]); 129 C_ADDTO(*Fout, scratch[3]);
125 130
126 Fout[m2].r = Fout[m].r + scratch[0].i; 131 Fout[m2].r = Fout[m].r + scratch[0].i;
127 Fout[m2].i = Fout[m].i - scratch[0].r; 132 Fout[m2].i = Fout[m].i - scratch[0].r;
128 133
129 Fout[m].r -= scratch[0].i; 134 Fout[m].r -= scratch[0].i;
130 Fout[m].i += scratch[0].r; 135 Fout[m].i += scratch[0].r;
131 136
132 ++Fout; 137 ++Fout;
133 }while(--k); 138 } while (--k);
134} 139}
135 140
136static void kf_bfly5( 141static void kf_bfly5(kiss_fft_cpx *Fout, const size_t fstride,
137 kiss_fft_cpx * Fout, 142 const kiss_fft_cfg st, int m) {
138 const size_t fstride, 143 kiss_fft_cpx *Fout0, *Fout1, *Fout2, *Fout3, *Fout4;
139 const kiss_fft_cfg st, 144 int u;
140 int m 145 kiss_fft_cpx scratch[13];
141 ) 146 kiss_fft_cpx *twiddles = st->twiddles;
142{ 147 kiss_fft_cpx *tw;
143 kiss_fft_cpx *Fout0,*Fout1,*Fout2,*Fout3,*Fout4; 148 kiss_fft_cpx ya, yb;
144 int u; 149 ya = twiddles[fstride * m];
145 kiss_fft_cpx scratch[13]; 150 yb = twiddles[fstride * 2 * m];
146 kiss_fft_cpx * twiddles = st->twiddles; 151
147 kiss_fft_cpx *tw; 152 Fout0 = Fout;
148 kiss_fft_cpx ya,yb; 153 Fout1 = Fout0 + m;
149 ya = twiddles[fstride*m]; 154 Fout2 = Fout0 + 2 * m;
150 yb = twiddles[fstride*2*m]; 155 Fout3 = Fout0 + 3 * m;
151 156 Fout4 = Fout0 + 4 * m;
152 Fout0=Fout; 157
153 Fout1=Fout0+m; 158 tw = st->twiddles;
154 Fout2=Fout0+2*m; 159 for (u = 0; u < m; ++u) {
155 Fout3=Fout0+3*m; 160 C_FIXDIV(*Fout0, 5);
156 Fout4=Fout0+4*m; 161 C_FIXDIV(*Fout1, 5);
157 162 C_FIXDIV(*Fout2, 5);
158 tw=st->twiddles; 163 C_FIXDIV(*Fout3, 5);
159 for ( u=0; u<m; ++u ) { 164 C_FIXDIV(*Fout4, 5);
160 C_FIXDIV( *Fout0,5); C_FIXDIV( *Fout1,5); C_FIXDIV( *Fout2,5); C_FIXDIV( *Fout3,5); C_FIXDIV( *Fout4,5); 165 scratch[0] = *Fout0;
161 scratch[0] = *Fout0; 166
162 167 C_MUL(scratch[1], *Fout1, tw[u * fstride]);
163 C_MUL(scratch[1] ,*Fout1, tw[u*fstride]); 168 C_MUL(scratch[2], *Fout2, tw[2 * u * fstride]);
164 C_MUL(scratch[2] ,*Fout2, tw[2*u*fstride]); 169 C_MUL(scratch[3], *Fout3, tw[3 * u * fstride]);
165 C_MUL(scratch[3] ,*Fout3, tw[3*u*fstride]); 170 C_MUL(scratch[4], *Fout4, tw[4 * u * fstride]);
166 C_MUL(scratch[4] ,*Fout4, tw[4*u*fstride]); 171
167 172 C_ADD(scratch[7], scratch[1], scratch[4]);
168 C_ADD( scratch[7],scratch[1],scratch[4]); 173 C_SUB(scratch[10], scratch[1], scratch[4]);
169 C_SUB( scratch[10],scratch[1],scratch[4]); 174 C_ADD(scratch[8], scratch[2], scratch[3]);
170 C_ADD( scratch[8],scratch[2],scratch[3]); 175 C_SUB(scratch[9], scratch[2], scratch[3]);
171 C_SUB( scratch[9],scratch[2],scratch[3]); 176
172 177 Fout0->r += scratch[7].r + scratch[8].r;
173 Fout0->r += scratch[7].r + scratch[8].r; 178 Fout0->i += scratch[7].i + scratch[8].i;
174 Fout0->i += scratch[7].i + scratch[8].i; 179
175 180 scratch[5].r =
176 scratch[5].r = scratch[0].r + S_MUL(scratch[7].r,ya.r) + S_MUL(scratch[8].r,yb.r); 181 scratch[0].r + S_MUL(scratch[7].r, ya.r) + S_MUL(scratch[8].r, yb.r);
177 scratch[5].i = scratch[0].i + S_MUL(scratch[7].i,ya.r) + S_MUL(scratch[8].i,yb.r); 182 scratch[5].i =
178 183 scratch[0].i + S_MUL(scratch[7].i, ya.r) + S_MUL(scratch[8].i, yb.r);
179 scratch[6].r = S_MUL(scratch[10].i,ya.i) + S_MUL(scratch[9].i,yb.i); 184
180 scratch[6].i = -S_MUL(scratch[10].r,ya.i) - S_MUL(scratch[9].r,yb.i); 185 scratch[6].r = S_MUL(scratch[10].i, ya.i) + S_MUL(scratch[9].i, yb.i);
181 186 scratch[6].i = -S_MUL(scratch[10].r, ya.i) - S_MUL(scratch[9].r, yb.i);
182 C_SUB(*Fout1,scratch[5],scratch[6]); 187
183 C_ADD(*Fout4,scratch[5],scratch[6]); 188 C_SUB(*Fout1, scratch[5], scratch[6]);
184 189 C_ADD(*Fout4, scratch[5], scratch[6]);
185 scratch[11].r = scratch[0].r + S_MUL(scratch[7].r,yb.r) + S_MUL(scratch[8].r,ya.r); 190
186 scratch[11].i = scratch[0].i + S_MUL(scratch[7].i,yb.r) + S_MUL(scratch[8].i,ya.r); 191 scratch[11].r =
187 scratch[12].r = - S_MUL(scratch[10].i,yb.i) + S_MUL(scratch[9].i,ya.i); 192 scratch[0].r + S_MUL(scratch[7].r, yb.r) + S_MUL(scratch[8].r, ya.r);
188 scratch[12].i = S_MUL(scratch[10].r,yb.i) - S_MUL(scratch[9].r,ya.i); 193 scratch[11].i =
189 194 scratch[0].i + S_MUL(scratch[7].i, yb.r) + S_MUL(scratch[8].i, ya.r);
190 C_ADD(*Fout2,scratch[11],scratch[12]); 195 scratch[12].r = -S_MUL(scratch[10].i, yb.i) + S_MUL(scratch[9].i, ya.i);
191 C_SUB(*Fout3,scratch[11],scratch[12]); 196 scratch[12].i = S_MUL(scratch[10].r, yb.i) - S_MUL(scratch[9].r, ya.i);
192 197
193 ++Fout0;++Fout1;++Fout2;++Fout3;++Fout4; 198 C_ADD(*Fout2, scratch[11], scratch[12]);
194 } 199 C_SUB(*Fout3, scratch[11], scratch[12]);
200
201 ++Fout0;
202 ++Fout1;
203 ++Fout2;
204 ++Fout3;
205 ++Fout4;
206 }
195} 207}
196 208
197/* perform the butterfly for one stage of a mixed radix FFT */ 209/* perform the butterfly for one stage of a mixed radix FFT */
198static void kf_bfly_generic( 210static void kf_bfly_generic(kiss_fft_cpx *Fout, const size_t fstride,
199 kiss_fft_cpx * Fout, 211 const kiss_fft_cfg st, int m, int p) {
200 const size_t fstride, 212 int u, k, q1, q;
201 const kiss_fft_cfg st, 213 kiss_fft_cpx *twiddles = st->twiddles;
202 int m, 214 kiss_fft_cpx t;
203 int p 215 int Norig = st->nfft;
204 ) 216
205{ 217 kiss_fft_cpx *scratch =
206 int u,k,q1,q; 218 (kiss_fft_cpx *)KISS_FFT_TMP_ALLOC(sizeof(kiss_fft_cpx) * p);
207 kiss_fft_cpx * twiddles = st->twiddles; 219
208 kiss_fft_cpx t; 220 for (u = 0; u < m; ++u) {
209 int Norig = st->nfft; 221 k = u;
210 222 for (q1 = 0; q1 < p; ++q1) {
211 kiss_fft_cpx * scratch = (kiss_fft_cpx*)KISS_FFT_TMP_ALLOC(sizeof(kiss_fft_cpx)*p); 223 scratch[q1] = Fout[k];
212 224 C_FIXDIV(scratch[q1], p);
213 for ( u=0; u<m; ++u ) { 225 k += m;
214 k=u;
215 for ( q1=0 ; q1<p ; ++q1 ) {
216 scratch[q1] = Fout[ k ];
217 C_FIXDIV(scratch[q1],p);
218 k += m;
219 }
220
221 k=u;
222 for ( q1=0 ; q1<p ; ++q1 ) {
223 int twidx=0;
224 Fout[ k ] = scratch[0];
225 for (q=1;q<p;++q ) {
226 twidx += fstride * k;
227 if (twidx>=Norig) twidx-=Norig;
228 C_MUL(t,scratch[q] , twiddles[twidx] );
229 C_ADDTO( Fout[ k ] ,t);
230 }
231 k += m;
232 }
233 } 226 }
234 KISS_FFT_TMP_FREE(scratch);
235}
236
237static
238void kf_work(
239 kiss_fft_cpx * Fout,
240 const kiss_fft_cpx * f,
241 const size_t fstride,
242 int in_stride,
243 int * factors,
244 const kiss_fft_cfg st
245 )
246{
247 kiss_fft_cpx * Fout_beg=Fout;
248 const int p=*factors++; /* the radix */
249 const int m=*factors++; /* stage's fft length/p */
250 const kiss_fft_cpx * Fout_end = Fout + p*m;
251 227
252#ifdef _OPENMP 228 k = u;
253 // use openmp extensions at the 229 for (q1 = 0; q1 < p; ++q1) {
254 // top-level (not recursive) 230 int twidx = 0;
255 if (fstride==1 && p<=5) 231 Fout[k] = scratch[0];
256 { 232 for (q = 1; q < p; ++q) {
257 int k; 233 twidx += fstride * k;
258 234 if (twidx >= Norig) twidx -= Norig;
259 // execute the p different work units in different threads 235 C_MUL(t, scratch[q], twiddles[twidx]);
260# pragma omp parallel for 236 C_ADDTO(Fout[k], t);
261 for (k=0;k<p;++k) 237 }
262 kf_work( Fout +k*m, f+ fstride*in_stride*k,fstride*p,in_stride,factors,st); 238 k += m;
263 // all threads have joined by this point
264
265 switch (p) {
266 case 2: kf_bfly2(Fout,fstride,st,m); break;
267 case 3: kf_bfly3(Fout,fstride,st,m); break;
268 case 4: kf_bfly4(Fout,fstride,st,m); break;
269 case 5: kf_bfly5(Fout,fstride,st,m); break;
270 default: kf_bfly_generic(Fout,fstride,st,m,p); break;
271 }
272 return;
273 } 239 }
274#endif 240 }
241 KISS_FFT_TMP_FREE(scratch);
242}
275 243
276 if (m==1) { 244static void kf_work(kiss_fft_cpx *Fout, const kiss_fft_cpx *f,
277 do{ 245 const size_t fstride, int in_stride, int *factors,
278 *Fout = *f; 246 const kiss_fft_cfg st) {
279 f += fstride*in_stride; 247 kiss_fft_cpx *Fout_beg = Fout;
280 }while(++Fout != Fout_end ); 248 const int p = *factors++; /* the radix */
281 }else{ 249 const int m = *factors++; /* stage's fft length/p */
282 do{ 250 const kiss_fft_cpx *Fout_end = Fout + p * m;
283 // recursive call:
284 // DFT of size m*p performed by doing
285 // p instances of smaller DFTs of size m,
286 // each one takes a decimated version of the input
287 kf_work( Fout , f, fstride*p, in_stride, factors,st);
288 f += fstride*in_stride;
289 }while( (Fout += m) != Fout_end );
290 }
291 251
292 Fout=Fout_beg; 252#ifdef _OPENMP
253 // use openmp extensions at the
254 // top-level (not recursive)
255 if (fstride == 1 && p <= 5) {
256 int k;
257
258 // execute the p different work units in different threads
259#pragma omp parallel for
260 for (k = 0; k < p; ++k)
261 kf_work(Fout + k * m, f + fstride * in_stride * k, fstride * p, in_stride,
262 factors, st);
263 // all threads have joined by this point
293 264
294 // recombine the p smaller DFTs
295 switch (p) { 265 switch (p) {
296 case 2: kf_bfly2(Fout,fstride,st,m); break; 266 case 2:
297 case 3: kf_bfly3(Fout,fstride,st,m); break; 267 kf_bfly2(Fout, fstride, st, m);
298 case 4: kf_bfly4(Fout,fstride,st,m); break; 268 break;
299 case 5: kf_bfly5(Fout,fstride,st,m); break; 269 case 3:
300 default: kf_bfly_generic(Fout,fstride,st,m,p); break; 270 kf_bfly3(Fout, fstride, st, m);
271 break;
272 case 4:
273 kf_bfly4(Fout, fstride, st, m);
274 break;
275 case 5:
276 kf_bfly5(Fout, fstride, st, m);
277 break;
278 default:
279 kf_bfly_generic(Fout, fstride, st, m, p);
280 break;
301 } 281 }
282 return;
283 }
284#endif
285
286 if (m == 1) {
287 do {
288 *Fout = *f;
289 f += fstride * in_stride;
290 } while (++Fout != Fout_end);
291 } else {
292 do {
293 // recursive call:
294 // DFT of size m*p performed by doing
295 // p instances of smaller DFTs of size m,
296 // each one takes a decimated version of the input
297 kf_work(Fout, f, fstride * p, in_stride, factors, st);
298 f += fstride * in_stride;
299 } while ((Fout += m) != Fout_end);
300 }
301
302 Fout = Fout_beg;
303
304 // recombine the p smaller DFTs
305 switch (p) {
306 case 2:
307 kf_bfly2(Fout, fstride, st, m);
308 break;
309 case 3:
310 kf_bfly3(Fout, fstride, st, m);
311 break;
312 case 4:
313 kf_bfly4(Fout, fstride, st, m);
314 break;
315 case 5:
316 kf_bfly5(Fout, fstride, st, m);
317 break;
318 default:
319 kf_bfly_generic(Fout, fstride, st, m, p);
320 break;
321 }
302} 322}
303 323
304/* facbuf is populated by p1,m1,p2,m2, ... 324/* facbuf is populated by p1,m1,p2,m2, ...
305 where 325 where
306 p[i] * m[i] = m[i-1] 326 p[i] * m[i] = m[i-1]
307 m0 = n */ 327 m0 = n */
308static 328static void kf_factor(int n, int *facbuf) {
309void kf_factor(int n,int * facbuf) 329 int p = 4;
310{ 330 double floor_sqrt;
311 int p=4; 331 floor_sqrt = floorf(sqrtf((double)n));
312 double floor_sqrt; 332
313 floor_sqrt = floorf( sqrtf((double)n) ); 333 /*factor out powers of 4, powers of 2, then any remaining primes */
314 334 do {
315 /*factor out powers of 4, powers of 2, then any remaining primes */ 335 while (n % p) {
316 do { 336 switch (p) {
317 while (n % p) { 337 case 4:
318 switch (p) { 338 p = 2;
319 case 4: p = 2; break; 339 break;
320 case 2: p = 3; break; 340 case 2:
321 default: p += 2; break; 341 p = 3;
322 } 342 break;
323 if (p > floor_sqrt) 343 default:
324 p = n; /* no more factors, skip to end */ 344 p += 2;
325 } 345 break;
326 n /= p; 346 }
327 *facbuf++ = p; 347 if (p > floor_sqrt) p = n; /* no more factors, skip to end */
328 *facbuf++ = n; 348 }
329 } while (n > 1); 349 n /= p;
350 *facbuf++ = p;
351 *facbuf++ = n;
352 } while (n > 1);
330} 353}
331 354
332/* 355/*
333 * 356 *
334 * User-callable function to allocate all necessary storage space for the fft. 357 * User-callable function to allocate all necessary storage space for the fft.
335 * 358 *
336 * The return value is a contiguous block of memory, allocated with malloc. As such, 359 * The return value is a contiguous block of memory, allocated with malloc. As
337 * It can be freed with free(), rather than a kiss_fft-specific function. 360 * such, It can be freed with free(), rather than a kiss_fft-specific function.
338 * */ 361 * */
339kiss_fft_cfg kiss_fft_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem ) 362kiss_fft_cfg kiss_fft_alloc(int nfft, int inverse_fft, void *mem,
340{ 363 size_t *lenmem) {
341 kiss_fft_cfg st=NULL; 364 kiss_fft_cfg st = NULL;
342 size_t memneeded = sizeof(struct kiss_fft_state) 365 size_t memneeded = sizeof(struct kiss_fft_state) +
343 + sizeof(kiss_fft_cpx)*(nfft-1); /* twiddle factors*/ 366 sizeof(kiss_fft_cpx) * (nfft - 1); /* twiddle factors*/
344 367
345 if ( lenmem==NULL ) { 368 if (lenmem == NULL) {
346 st = ( kiss_fft_cfg)KISS_FFT_MALLOC( memneeded ); 369 st = (kiss_fft_cfg)KISS_FFT_MALLOC(memneeded);
347 }else{ 370 } else {
348 if (mem != NULL && *lenmem >= memneeded) 371 if (mem != NULL && *lenmem >= memneeded) st = (kiss_fft_cfg)mem;
349 st = (kiss_fft_cfg)mem; 372 *lenmem = memneeded;
350 *lenmem = memneeded; 373 }
351 } 374 if (st) {
352 if (st) { 375 int i;
353 int i; 376 st->nfft = nfft;
354 st->nfft=nfft; 377 st->inverse = inverse_fft;
355 st->inverse = inverse_fft; 378
356 379 for (i = 0; i < nfft; ++i) {
357 for (i=0;i<nfft;++i) { 380 const double pi =
358 const double pi=3.141592653589793238462643383279502884197169399375105820974944; 381 3.141592653589793238462643383279502884197169399375105820974944;
359 double phase = -2*pi*i / nfft; 382 double phase = -2 * pi * i / nfft;
360 if (st->inverse) 383 if (st->inverse) phase *= -1;
361 phase *= -1; 384 kf_cexp(st->twiddles + i, phase);
362 kf_cexp(st->twiddles+i, phase );
363 }
364
365 kf_factor(nfft,st->factors);
366 } 385 }
367 return st;
368}
369
370 386
371void kiss_fft_stride(kiss_fft_cfg st,const kiss_fft_cpx *fin,kiss_fft_cpx *fout,int in_stride) 387 kf_factor(nfft, st->factors);
372{ 388 }
373 if (fin == fout) { 389 return st;
374 //NOTE: this is not really an in-place FFT algorithm.
375 //It just performs an out-of-place FFT into a temp buffer
376 kiss_fft_cpx * tmpbuf = (kiss_fft_cpx*)KISS_FFT_TMP_ALLOC( sizeof(kiss_fft_cpx)*st->nfft);
377 kf_work(tmpbuf,fin,1,in_stride, st->factors,st);
378 memcpy(fout,tmpbuf,sizeof(kiss_fft_cpx)*st->nfft);
379 KISS_FFT_TMP_FREE(tmpbuf);
380 }else{
381 kf_work( fout, fin, 1,in_stride, st->factors,st );
382 }
383} 390}
384 391
385void kiss_fft(kiss_fft_cfg cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout) 392void kiss_fft_stride(kiss_fft_cfg st, const kiss_fft_cpx *fin,
386{ 393 kiss_fft_cpx *fout, int in_stride) {
387 kiss_fft_stride(cfg,fin,fout,1); 394 if (fin == fout) {
395 // NOTE: this is not really an in-place FFT algorithm.
396 // It just performs an out-of-place FFT into a temp buffer
397 kiss_fft_cpx *tmpbuf =
398 (kiss_fft_cpx *)KISS_FFT_TMP_ALLOC(sizeof(kiss_fft_cpx) * st->nfft);
399 kf_work(tmpbuf, fin, 1, in_stride, st->factors, st);
400 memcpy(fout, tmpbuf, sizeof(kiss_fft_cpx) * st->nfft);
401 KISS_FFT_TMP_FREE(tmpbuf);
402 } else {
403 kf_work(fout, fin, 1, in_stride, st->factors, st);
404 }
388} 405}
389 406
407void kiss_fft(kiss_fft_cfg cfg, const kiss_fft_cpx *fin, kiss_fft_cpx *fout) {
408 kiss_fft_stride(cfg, fin, fout, 1);
409}
390 410
391void kiss_fft_cleanup(void) 411void kiss_fft_cleanup(void) {
392{ 412 // nothing needed any more
393 // nothing needed any more
394} 413}
395 414
396int kiss_fft_next_fast_size(int n) 415int kiss_fft_next_fast_size(int n) {
397{ 416 while (1) {
398 while(1) { 417 int m = n;
399 int m=n; 418 while ((m % 2) == 0) m /= 2;
400 while ( (m%2) == 0 ) m/=2; 419 while ((m % 3) == 0) m /= 3;
401 while ( (m%3) == 0 ) m/=3; 420 while ((m % 5) == 0) m /= 5;
402 while ( (m%5) == 0 ) m/=5; 421 if (m <= 1)
403 if (m<=1) 422 break; /* n is completely factorable by twos, threes, and fives */
404 break; /* n is completely factorable by twos, threes, and fives */ 423 n++;
405 n++; 424 }
406 } 425 return n;
407 return n;
408} 426}
diff --git a/kiss_fft.h b/kiss_fft.h
index c01722c..ea1349e 100644
--- a/kiss_fft.h
+++ b/kiss_fft.h
@@ -1,9 +1,9 @@
1#ifndef KISS_FFT_H 1#ifndef KISS_FFT_H
2#define KISS_FFT_H 2#define KISS_FFT_H
3 3
4#include <stdlib.h>
5#include <stdio.h>
6#include <math.h> 4#include <math.h>
5#include <stdio.h>
6#include <stdlib.h>
7#include <string.h> 7#include <string.h>
8 8
9#ifdef __cplusplus 9#ifdef __cplusplus
@@ -24,36 +24,35 @@ extern "C" {
24*/ 24*/
25 25
26#ifdef USE_SIMD 26#ifdef USE_SIMD
27# include <xmmintrin.h> 27#include <xmmintrin.h>
28# define kiss_fft_scalar __m128 28#define kiss_fft_scalar __m128
29#define KISS_FFT_MALLOC(nbytes) _mm_malloc(nbytes,16) 29#define KISS_FFT_MALLOC(nbytes) _mm_malloc(nbytes, 16)
30#define KISS_FFT_FREE _mm_free 30#define KISS_FFT_FREE _mm_free
31#else 31#else
32#define KISS_FFT_MALLOC malloc 32#define KISS_FFT_MALLOC malloc
33#define KISS_FFT_FREE free 33#define KISS_FFT_FREE free
34#endif 34#endif
35 35
36
37#ifdef FIXED_POINT 36#ifdef FIXED_POINT
38#include <sys/types.h> 37#include <sys/types.h>
39# if (FIXED_POINT == 32) 38#if (FIXED_POINT == 32)
40# define kiss_fft_scalar int32_t 39#define kiss_fft_scalar int32_t
41# else
42# define kiss_fft_scalar int16_t
43# endif
44#else 40#else
45# ifndef kiss_fft_scalar 41#define kiss_fft_scalar int16_t
42#endif
43#else
44#ifndef kiss_fft_scalar
46/* default is float */ 45/* default is float */
47# define kiss_fft_scalar float 46#define kiss_fft_scalar float
48# endif 47#endif
49#endif 48#endif
50 49
51typedef struct { 50typedef struct {
52 kiss_fft_scalar r; 51 kiss_fft_scalar r;
53 kiss_fft_scalar i; 52 kiss_fft_scalar i;
54}kiss_fft_cpx; 53} kiss_fft_cpx;
55 54
56typedef struct kiss_fft_state* kiss_fft_cfg; 55typedef struct kiss_fft_state *kiss_fft_cfg;
57 56
58/* 57/*
59 * kiss_fft_alloc 58 * kiss_fft_alloc
@@ -65,8 +64,8 @@ typedef struct kiss_fft_state* kiss_fft_cfg;
65 * The return value from fft_alloc is a cfg buffer used internally 64 * The return value from fft_alloc is a cfg buffer used internally
66 * by the fft routine or NULL. 65 * by the fft routine or NULL.
67 * 66 *
68 * If lenmem is NULL, then kiss_fft_alloc will allocate a cfg buffer using malloc. 67 * If lenmem is NULL, then kiss_fft_alloc will allocate a cfg buffer using
69 * The returned value should be free()d when done to avoid memory leaks. 68 * malloc. The returned value should be free()d when done to avoid memory leaks.
70 * 69 *
71 * The state can be placed in a user supplied buffer 'mem': 70 * The state can be placed in a user supplied buffer 'mem':
72 * If lenmem is not NULL and mem is not NULL and *lenmem is large enough, 71 * If lenmem is not NULL and mem is not NULL and *lenmem is large enough,
@@ -78,7 +77,8 @@ typedef struct kiss_fft_state* kiss_fft_cfg;
78 * buffer size in *lenmem. 77 * buffer size in *lenmem.
79 * */ 78 * */
80 79
81kiss_fft_cfg kiss_fft_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem); 80kiss_fft_cfg kiss_fft_alloc(int nfft, int inverse_fft, void *mem,
81 size_t *lenmem);
82 82
83/* 83/*
84 * kiss_fft(cfg,in_out_buf) 84 * kiss_fft(cfg,in_out_buf)
@@ -90,32 +90,34 @@ kiss_fft_cfg kiss_fft_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem)
90 * Note that each element is complex and can be accessed like 90 * Note that each element is complex and can be accessed like
91 f[k].r and f[k].i 91 f[k].r and f[k].i
92 * */ 92 * */
93void kiss_fft(kiss_fft_cfg cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout); 93void kiss_fft(kiss_fft_cfg cfg, const kiss_fft_cpx *fin, kiss_fft_cpx *fout);
94 94
95/* 95/*
96 A more generic version of the above function. It reads its input from every Nth sample. 96 A more generic version of the above function. It reads its input from every Nth
97 sample.
97 * */ 98 * */
98void kiss_fft_stride(kiss_fft_cfg cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout,int fin_stride); 99void kiss_fft_stride(kiss_fft_cfg cfg, const kiss_fft_cpx *fin,
100 kiss_fft_cpx *fout, int fin_stride);
99 101
100/* If kiss_fft_alloc allocated a buffer, it is one contiguous 102/* If kiss_fft_alloc allocated a buffer, it is one contiguous
101 buffer and can be simply free()d when no longer needed*/ 103 buffer and can be simply free()d when no longer needed*/
102#define kiss_fft_free free 104#define kiss_fft_free free
103 105
104/* 106/*
105 Cleans up some memory that gets managed internally. Not necessary to call, but it might clean up 107 Cleans up some memory that gets managed internally. Not necessary to call, but
106 your compiler output to call this before you exit. 108 it might clean up your compiler output to call this before you exit.
107*/ 109*/
108void kiss_fft_cleanup(void); 110void kiss_fft_cleanup(void);
109 111
110
111/* 112/*
112 * Returns the smallest integer k, such that k>=n and k has only "fast" factors (2,3,5) 113 * Returns the smallest integer k, such that k>=n and k has only "fast" factors
114 * (2,3,5)
113 */ 115 */
114int kiss_fft_next_fast_size(int n); 116int kiss_fft_next_fast_size(int n);
115 117
116/* for real ffts, we need an even size */ 118/* for real ffts, we need an even size */
117#define kiss_fftr_next_fast_size_real(n) \ 119#define kiss_fftr_next_fast_size_real(n) \
118 (kiss_fft_next_fast_size( ((n)+1)>>1)<<1) 120 (kiss_fft_next_fast_size(((n) + 1) >> 1) << 1)
119 121
120#ifdef __cplusplus 122#ifdef __cplusplus
121} 123}
diff --git a/kiss_fftr.c b/kiss_fftr.c
index 7cc0286..e3aee37 100644
--- a/kiss_fftr.c
+++ b/kiss_fftr.c
@@ -3,152 +3,166 @@ Copyright (c) 2003-2004, Mark Borgerding
3 3
4All rights reserved. 4All rights reserved.
5 5
6Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: 6Redistribution and use in source and binary forms, with or without modification,
7 7are permitted provided that the following conditions are met:
8 * Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. 8
9 * Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. 9 * Redistributions of source code must retain the above copyright notice,
10 * Neither the author nor the names of any contributors may be used to endorse or promote products derived from this software without specific prior written permission. 10this list of conditions and the following disclaimer.
11 11 * Redistributions in binary form must reproduce the above copyright notice,
12THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 12this list of conditions and the following disclaimer in the documentation and/or
13other materials provided with the distribution.
14 * Neither the author nor the names of any contributors may be used to
15endorse or promote products derived from this software without specific prior
16written permission.
17
18THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
19ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
20WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
21DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
22ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
23(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
24LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
25ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
27SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
13*/ 28*/
14 29
15#include "kiss_fftr.h" 30#include "kiss_fftr.h"
31
16#include "_kiss_fft_guts.h" 32#include "_kiss_fft_guts.h"
17#include "assert.h" 33#include "assert.h"
18 34
19struct kiss_fftr_state{ 35struct kiss_fftr_state {
20 kiss_fft_cfg substate; 36 kiss_fft_cfg substate;
21 kiss_fft_cpx * tmpbuf; 37 kiss_fft_cpx *tmpbuf;
22 kiss_fft_cpx * super_twiddles; 38 kiss_fft_cpx *super_twiddles;
23#ifdef USE_SIMD 39#ifdef USE_SIMD
24 void * pad; 40 void *pad;
25#endif 41#endif
26}; 42};
27 43
28kiss_fftr_cfg kiss_fftr_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem) 44kiss_fftr_cfg kiss_fftr_alloc(int nfft, int inverse_fft, void *mem,
29{ 45 size_t *lenmem) {
30 int i; 46 int i;
31 kiss_fftr_cfg st = NULL; 47 kiss_fftr_cfg st = NULL;
32 size_t subsize, memneeded; 48 size_t subsize, memneeded;
33 49
34 if (nfft & 1) { 50 if (nfft & 1) {
35 fprintf(stderr,"Real FFT optimization must be even.\n"); 51 fprintf(stderr, "Real FFT optimization must be even.\n");
36 return NULL; 52 return NULL;
37 } 53 }
38 nfft >>= 1; 54 nfft >>= 1;
39 55
40 kiss_fft_alloc (nfft, inverse_fft, NULL, &subsize); 56 kiss_fft_alloc(nfft, inverse_fft, NULL, &subsize);
41 memneeded = sizeof(struct kiss_fftr_state) + subsize + sizeof(kiss_fft_cpx) * ( nfft * 3 / 2); 57 memneeded = sizeof(struct kiss_fftr_state) + subsize +
42 58 sizeof(kiss_fft_cpx) * (nfft * 3 / 2);
43 if (lenmem == NULL) { 59
44 st = (kiss_fftr_cfg) KISS_FFT_MALLOC (memneeded); 60 if (lenmem == NULL) {
45 } else { 61 st = (kiss_fftr_cfg)KISS_FFT_MALLOC(memneeded);
46 if (*lenmem >= memneeded) 62 } else {
47 st = (kiss_fftr_cfg) mem; 63 if (*lenmem >= memneeded) st = (kiss_fftr_cfg)mem;
48 *lenmem = memneeded; 64 *lenmem = memneeded;
49 } 65 }
50 if (!st) 66 if (!st) return NULL;
51 return NULL; 67
52 68 st->substate = (kiss_fft_cfg)(st + 1); /*just beyond kiss_fftr_state struct */
53 st->substate = (kiss_fft_cfg) (st + 1); /*just beyond kiss_fftr_state struct */ 69 st->tmpbuf = (kiss_fft_cpx *)(((char *)st->substate) + subsize);
54 st->tmpbuf = (kiss_fft_cpx *) (((char *) st->substate) + subsize); 70 st->super_twiddles = st->tmpbuf + nfft;
55 st->super_twiddles = st->tmpbuf + nfft; 71 kiss_fft_alloc(nfft, inverse_fft, st->substate, &subsize);
56 kiss_fft_alloc(nfft, inverse_fft, st->substate, &subsize); 72
57 73 for (i = 0; i < nfft / 2; ++i) {
58 for (i = 0; i < nfft/2; ++i) { 74 float phase =
59 float phase = 75 -3.14159265358979323846264338327 * ((float)(i + 1) / nfft + .5);
60 -3.14159265358979323846264338327 * ((float) (i+1) / nfft + .5); 76 if (inverse_fft) phase *= -1;
61 if (inverse_fft) 77 kf_cexp(st->super_twiddles + i, phase);
62 phase *= -1; 78 }
63 kf_cexp (st->super_twiddles+i,phase); 79 return st;
64 }
65 return st;
66} 80}
67 81
68void kiss_fftr(kiss_fftr_cfg st,const kiss_fft_scalar *timedata,kiss_fft_cpx *freqdata) 82void kiss_fftr(kiss_fftr_cfg st, const kiss_fft_scalar *timedata,
69{ 83 kiss_fft_cpx *freqdata) {
70 /* input buffer timedata is stored row-wise */ 84 /* input buffer timedata is stored row-wise */
71 int k,ncfft; 85 int k, ncfft;
72 kiss_fft_cpx fpnk,fpk,f1k,f2k,tw,tdc; 86 kiss_fft_cpx fpnk, fpk, f1k, f2k, tw, tdc;
73 87
74 assert(st->substate->inverse==0); 88 assert(st->substate->inverse == 0);
75 89
76 ncfft = st->substate->nfft; 90 ncfft = st->substate->nfft;
77 91
78 /*perform the parallel fft of two real signals packed in real,imag*/ 92 /*perform the parallel fft of two real signals packed in real,imag*/
79 kiss_fft( st->substate , (const kiss_fft_cpx*)timedata, st->tmpbuf ); 93 kiss_fft(st->substate, (const kiss_fft_cpx *)timedata, st->tmpbuf);
80 /* The real part of the DC element of the frequency spectrum in st->tmpbuf 94 /* The real part of the DC element of the frequency spectrum in st->tmpbuf
81 * contains the sum of the even-numbered elements of the input time sequence 95 * contains the sum of the even-numbered elements of the input time sequence
82 * The imag part is the sum of the odd-numbered elements 96 * The imag part is the sum of the odd-numbered elements
83 * 97 *
84 * The sum of tdc.r and tdc.i is the sum of the input time sequence. 98 * The sum of tdc.r and tdc.i is the sum of the input time sequence.
85 * yielding DC of input time sequence 99 * yielding DC of input time sequence
86 * The difference of tdc.r - tdc.i is the sum of the input (dot product) [1,-1,1,-1... 100 * The difference of tdc.r - tdc.i is the sum of the input (dot product)
87 * yielding Nyquist bin of input time sequence 101 * [1,-1,1,-1... yielding Nyquist bin of input time sequence
88 */ 102 */
89 103
90 tdc.r = st->tmpbuf[0].r; 104 tdc.r = st->tmpbuf[0].r;
91 tdc.i = st->tmpbuf[0].i; 105 tdc.i = st->tmpbuf[0].i;
92 C_FIXDIV(tdc,2); 106 C_FIXDIV(tdc, 2);
93 CHECK_OVERFLOW_OP(tdc.r ,+, tdc.i); 107 CHECK_OVERFLOW_OP(tdc.r, +, tdc.i);
94 CHECK_OVERFLOW_OP(tdc.r ,-, tdc.i); 108 CHECK_OVERFLOW_OP(tdc.r, -, tdc.i);
95 freqdata[0].r = tdc.r + tdc.i; 109 freqdata[0].r = tdc.r + tdc.i;
96 freqdata[ncfft].r = tdc.r - tdc.i; 110 freqdata[ncfft].r = tdc.r - tdc.i;
97#ifdef USE_SIMD 111#ifdef USE_SIMD
98 freqdata[ncfft].i = freqdata[0].i = _mm_set1_ps(0); 112 freqdata[ncfft].i = freqdata[0].i = _mm_set1_ps(0);
99#else 113#else
100 freqdata[ncfft].i = freqdata[0].i = 0; 114 freqdata[ncfft].i = freqdata[0].i = 0;
101#endif 115#endif
102 116
103 for ( k=1;k <= ncfft/2 ; ++k ) { 117 for (k = 1; k <= ncfft / 2; ++k) {
104 fpk = st->tmpbuf[k]; 118 fpk = st->tmpbuf[k];
105 fpnk.r = st->tmpbuf[ncfft-k].r; 119 fpnk.r = st->tmpbuf[ncfft - k].r;
106 fpnk.i = - st->tmpbuf[ncfft-k].i; 120 fpnk.i = -st->tmpbuf[ncfft - k].i;
107 C_FIXDIV(fpk,2); 121 C_FIXDIV(fpk, 2);
108 C_FIXDIV(fpnk,2); 122 C_FIXDIV(fpnk, 2);
109 123
110 C_ADD( f1k, fpk , fpnk ); 124 C_ADD(f1k, fpk, fpnk);
111 C_SUB( f2k, fpk , fpnk ); 125 C_SUB(f2k, fpk, fpnk);
112 C_MUL( tw , f2k , st->super_twiddles[k-1]); 126 C_MUL(tw, f2k, st->super_twiddles[k - 1]);
113 127
114 freqdata[k].r = HALF_OF(f1k.r + tw.r); 128 freqdata[k].r = HALF_OF(f1k.r + tw.r);
115 freqdata[k].i = HALF_OF(f1k.i + tw.i); 129 freqdata[k].i = HALF_OF(f1k.i + tw.i);
116 freqdata[ncfft-k].r = HALF_OF(f1k.r - tw.r); 130 freqdata[ncfft - k].r = HALF_OF(f1k.r - tw.r);
117 freqdata[ncfft-k].i = HALF_OF(tw.i - f1k.i); 131 freqdata[ncfft - k].i = HALF_OF(tw.i - f1k.i);
118 } 132 }
119} 133}
120 134
121void kiss_fftri(kiss_fftr_cfg st,const kiss_fft_cpx *freqdata,kiss_fft_scalar *timedata) 135void kiss_fftri(kiss_fftr_cfg st, const kiss_fft_cpx *freqdata,
122{ 136 kiss_fft_scalar *timedata) {
123 /* input buffer timedata is stored row-wise */ 137 /* input buffer timedata is stored row-wise */
124 int k, ncfft; 138 int k, ncfft;
125 139
126 assert(st->substate->inverse == 1); 140 assert(st->substate->inverse == 1);
127 141
128 ncfft = st->substate->nfft; 142 ncfft = st->substate->nfft;
129 143
130 st->tmpbuf[0].r = freqdata[0].r + freqdata[ncfft].r; 144 st->tmpbuf[0].r = freqdata[0].r + freqdata[ncfft].r;
131 st->tmpbuf[0].i = freqdata[0].r - freqdata[ncfft].r; 145 st->tmpbuf[0].i = freqdata[0].r - freqdata[ncfft].r;
132 C_FIXDIV(st->tmpbuf[0],2); 146 C_FIXDIV(st->tmpbuf[0], 2);
133 147
134 for (k = 1; k <= ncfft / 2; ++k) { 148 for (k = 1; k <= ncfft / 2; ++k) {
135 kiss_fft_cpx fk, fnkc, fek, fok, tmp; 149 kiss_fft_cpx fk, fnkc, fek, fok, tmp;
136 fk = freqdata[k]; 150 fk = freqdata[k];
137 fnkc.r = freqdata[ncfft - k].r; 151 fnkc.r = freqdata[ncfft - k].r;
138 fnkc.i = -freqdata[ncfft - k].i; 152 fnkc.i = -freqdata[ncfft - k].i;
139 C_FIXDIV( fk , 2 ); 153 C_FIXDIV(fk, 2);
140 C_FIXDIV( fnkc , 2 ); 154 C_FIXDIV(fnkc, 2);
141 155
142 C_ADD (fek, fk, fnkc); 156 C_ADD(fek, fk, fnkc);
143 C_SUB (tmp, fk, fnkc); 157 C_SUB(tmp, fk, fnkc);
144 C_MUL (fok, tmp, st->super_twiddles[k-1]); 158 C_MUL(fok, tmp, st->super_twiddles[k - 1]);
145 C_ADD (st->tmpbuf[k], fek, fok); 159 C_ADD(st->tmpbuf[k], fek, fok);
146 C_SUB (st->tmpbuf[ncfft - k], fek, fok); 160 C_SUB(st->tmpbuf[ncfft - k], fek, fok);
147#ifdef USE_SIMD 161#ifdef USE_SIMD
148 st->tmpbuf[ncfft - k].i *= _mm_set1_ps(-1.0); 162 st->tmpbuf[ncfft - k].i *= _mm_set1_ps(-1.0);
149#else 163#else
150 st->tmpbuf[ncfft - k].i *= -1; 164 st->tmpbuf[ncfft - k].i *= -1;
151#endif 165#endif
152 } 166 }
153 kiss_fft (st->substate, st->tmpbuf, (kiss_fft_cpx *) timedata); 167 kiss_fft(st->substate, st->tmpbuf, (kiss_fft_cpx *)timedata);
154} 168}
diff --git a/kiss_fftr.h b/kiss_fftr.h
index 72e5a57..4ab52db 100644
--- a/kiss_fftr.h
+++ b/kiss_fftr.h
@@ -6,33 +6,34 @@
6extern "C" { 6extern "C" {
7#endif 7#endif
8 8
9 9/*
10/* 10
11 11 Real optimized version can save about 45% cpu time vs. complex fft of a real
12 Real optimized version can save about 45% cpu time vs. complex fft of a real seq. 12 seq.
13
14
13 15
14
15
16 */ 16 */
17 17
18typedef struct kiss_fftr_state *kiss_fftr_cfg; 18typedef struct kiss_fftr_state *kiss_fftr_cfg;
19 19
20 20kiss_fftr_cfg kiss_fftr_alloc(int nfft, int inverse_fft, void *mem,
21kiss_fftr_cfg kiss_fftr_alloc(int nfft,int inverse_fft,void * mem, size_t * lenmem); 21 size_t *lenmem);
22/* 22/*
23 nfft must be even 23 nfft must be even
24 24
25 If you don't care to allocate space, use mem = lenmem = NULL 25 If you don't care to allocate space, use mem = lenmem = NULL
26*/ 26*/
27 27
28 28void kiss_fftr(kiss_fftr_cfg cfg, const kiss_fft_scalar *timedata,
29void kiss_fftr(kiss_fftr_cfg cfg,const kiss_fft_scalar *timedata,kiss_fft_cpx *freqdata); 29 kiss_fft_cpx *freqdata);
30/* 30/*
31 input timedata has nfft scalar points 31 input timedata has nfft scalar points
32 output freqdata has nfft/2+1 complex points 32 output freqdata has nfft/2+1 complex points
33*/ 33*/
34 34
35void kiss_fftri(kiss_fftr_cfg cfg,const kiss_fft_cpx *freqdata,kiss_fft_scalar *timedata); 35void kiss_fftri(kiss_fftr_cfg cfg, const kiss_fft_cpx *freqdata,
36 kiss_fft_scalar *timedata);
36/* 37/*
37 input freqdata has nfft/2+1 complex points 38 input freqdata has nfft/2+1 complex points
38 output timedata has nfft scalar points 39 output timedata has nfft scalar points
diff --git a/lpc.c b/lpc.c
index 31442cd..3117a1a 100644
--- a/lpc.c
+++ b/lpc.c
@@ -25,16 +25,18 @@
25 along with this program; if not, see <http://www.gnu.org/licenses/>. 25 along with this program; if not, see <http://www.gnu.org/licenses/>.
26*/ 26*/
27 27
28#define LPC_MAX_N 512 /* maximum no. of samples in frame */ 28#define LPC_MAX_N 512 /* maximum no. of samples in frame */
29#define PI 3.141592654 /* mathematical constant */ 29#define PI 3.141592654 /* mathematical constant */
30 30
31#define ALPHA 1.0 31#define ALPHA 1.0
32#define BETA 0.94 32#define BETA 0.94
33
34#include "lpc.h"
33 35
34#include <assert.h> 36#include <assert.h>
35#include <math.h> 37#include <math.h>
38
36#include "defines.h" 39#include "defines.h"
37#include "lpc.h"
38 40
39/*---------------------------------------------------------------------------*\ 41/*---------------------------------------------------------------------------*\
40 42
@@ -42,29 +44,25 @@
42 44
43 Pre-emphasise (high pass filter with zero close to 0 Hz) a frame of 45 Pre-emphasise (high pass filter with zero close to 0 Hz) a frame of
44 speech samples. Helps reduce dynamic range of LPC spectrum, giving 46 speech samples. Helps reduce dynamic range of LPC spectrum, giving
45 greater weight and hense a better match to low energy formants. 47 greater weight and hence a better match to low energy formants.
46 48
47 Should be balanced by de-emphasis of the output speech. 49 Should be balanced by de-emphasis of the output speech.
48 50
49\*---------------------------------------------------------------------------*/ 51\*---------------------------------------------------------------------------*/
50 52
51void pre_emp( 53void pre_emp(float Sn_pre[], /* output frame of speech samples */
52 float Sn_pre[], /* output frame of speech samples */ 54 float Sn[], /* input frame of speech samples */
53 float Sn[], /* input frame of speech samples */ 55 float *mem, /* Sn[-1]single sample memory */
54 float *mem, /* Sn[-1]single sample memory */ 56 int Nsam /* number of speech samples to use */
55 int Nsam /* number of speech samples to use */ 57) {
56) 58 int i;
57{
58 int i;
59
60 for(i=0; i<Nsam; i++) {
61 Sn_pre[i] = Sn[i] - ALPHA * mem[0];
62 mem[0] = Sn[i];
63 }
64 59
60 for (i = 0; i < Nsam; i++) {
61 Sn_pre[i] = Sn[i] - ALPHA * mem[0];
62 mem[0] = Sn[i];
63 }
65} 64}
66 65
67
68/*---------------------------------------------------------------------------*\ 66/*---------------------------------------------------------------------------*\
69 67
70 de_emp() 68 de_emp()
@@ -73,23 +71,19 @@ void pre_emp(
73 71
74\*---------------------------------------------------------------------------*/ 72\*---------------------------------------------------------------------------*/
75 73
76void de_emp( 74void de_emp(float Sn_de[], /* output frame of speech samples */
77 float Sn_de[], /* output frame of speech samples */ 75 float Sn[], /* input frame of speech samples */
78 float Sn[], /* input frame of speech samples */ 76 float *mem, /* Sn[-1]single sample memory */
79 float *mem, /* Sn[-1]single sample memory */ 77 int Nsam /* number of speech samples to use */
80 int Nsam /* number of speech samples to use */ 78) {
81) 79 int i;
82{
83 int i;
84
85 for(i=0; i<Nsam; i++) {
86 Sn_de[i] = Sn[i] + BETA * mem[0];
87 mem[0] = Sn_de[i];
88 }
89 80
81 for (i = 0; i < Nsam; i++) {
82 Sn_de[i] = Sn[i] + BETA * mem[0];
83 mem[0] = Sn_de[i];
84 }
90} 85}
91 86
92
93/*---------------------------------------------------------------------------*\ 87/*---------------------------------------------------------------------------*\
94 88
95 hanning_window() 89 hanning_window()
@@ -98,16 +92,14 @@ void de_emp(
98 92
99\*---------------------------------------------------------------------------*/ 93\*---------------------------------------------------------------------------*/
100 94
101void hanning_window( 95void hanning_window(float Sn[], /* input frame of speech samples */
102 float Sn[], /* input frame of speech samples */ 96 float Wn[], /* output frame of windowed samples */
103 float Wn[], /* output frame of windowed samples */ 97 int Nsam /* number of samples */
104 int Nsam /* number of samples */ 98) {
105) 99 int i; /* loop variable */
106{
107 int i; /* loop variable */
108 100
109 for(i=0; i<Nsam; i++) 101 for (i = 0; i < Nsam; i++)
110 Wn[i] = Sn[i]*(0.5 - 0.5*cosf(2*PI*(float)i/(Nsam-1))); 102 Wn[i] = Sn[i] * (0.5 - 0.5 * cosf(2 * PI * (float)i / (Nsam - 1)));
111} 103}
112 104
113/*---------------------------------------------------------------------------*\ 105/*---------------------------------------------------------------------------*\
@@ -119,19 +111,16 @@ void hanning_window(
119 111
120\*---------------------------------------------------------------------------*/ 112\*---------------------------------------------------------------------------*/
121 113
122void autocorrelate( 114void autocorrelate(float Sn[], /* frame of Nsam windowed speech samples */
123 float Sn[], /* frame of Nsam windowed speech samples */ 115 float Rn[], /* array of P+1 autocorrelation coefficients */
124 float Rn[], /* array of P+1 autocorrelation coefficients */ 116 int Nsam, /* number of windowed samples to use */
125 int Nsam, /* number of windowed samples to use */ 117 int order /* order of LPC analysis */
126 int order /* order of LPC analysis */ 118) {
127) 119 int i, j; /* loop variables */
128{
129 int i,j; /* loop variables */
130 120
131 for(j=0; j<order+1; j++) { 121 for (j = 0; j < order + 1; j++) {
132 Rn[j] = 0.0; 122 Rn[j] = 0.0;
133 for(i=0; i<Nsam-j; i++) 123 for (i = 0; i < Nsam - j; i++) Rn[j] += Sn[i] * Sn[i + j];
134 Rn[j] += Sn[i]*Sn[i+j];
135 } 124 }
136} 125}
137 126
@@ -150,36 +139,31 @@ void autocorrelate(
150 139
151\*---------------------------------------------------------------------------*/ 140\*---------------------------------------------------------------------------*/
152 141
153void levinson_durbin( 142void levinson_durbin(float R[], /* order+1 autocorrelation coeff */
154 float R[], /* order+1 autocorrelation coeff */ 143 float lpcs[], /* order+1 LPC's */
155 float lpcs[], /* order+1 LPC's */ 144 int order /* order of the LPC analysis */
156 int order /* order of the LPC analysis */ 145) {
157) 146 float a[order + 1][order + 1];
158{
159 float a[order+1][order+1];
160 float sum, e, k; 147 float sum, e, k;
161 int i,j; /* loop variables */ 148 int i, j; /* loop variables */
162 149
163 e = R[0]; /* Equation 38a, Makhoul */ 150 e = R[0]; /* Equation 38a, Makhoul */
164 151
165 for(i=1; i<=order; i++) { 152 for (i = 1; i <= order; i++) {
166 sum = 0.0; 153 sum = 0.0;
167 for(j=1; j<=i-1; j++) 154 for (j = 1; j <= i - 1; j++) sum += a[i - 1][j] * R[i - j];
168 sum += a[i-1][j]*R[i-j]; 155 k = -1.0 * (R[i] + sum) / e; /* Equation 38b, Makhoul */
169 k = -1.0*(R[i] + sum)/e; /* Equation 38b, Makhoul */ 156 if (fabsf(k) > 1.0) k = 0.0;
170 if (fabsf(k) > 1.0)
171 k = 0.0;
172 157
173 a[i][i] = k; 158 a[i][i] = k;
174 159
175 for(j=1; j<=i-1; j++) 160 for (j = 1; j <= i - 1; j++)
176 a[i][j] = a[i-1][j] + k*a[i-1][i-j]; /* Equation 38c, Makhoul */ 161 a[i][j] = a[i - 1][j] + k * a[i - 1][i - j]; /* Equation 38c, Makhoul */
177 162
178 e *= (1-k*k); /* Equation 38d, Makhoul */ 163 e *= (1 - k * k); /* Equation 38d, Makhoul */
179 } 164 }
180 165
181 for(i=1; i<=order; i++) 166 for (i = 1; i <= order; i++) lpcs[i] = a[order][i];
182 lpcs[i] = a[order][i];
183 lpcs[0] = 1.0; 167 lpcs[0] = 1.0;
184} 168}
185 169
@@ -194,20 +178,17 @@ void levinson_durbin(
194 178
195\*---------------------------------------------------------------------------*/ 179\*---------------------------------------------------------------------------*/
196 180
197void inverse_filter( 181void inverse_filter(float Sn[], /* Nsam input samples */
198 float Sn[], /* Nsam input samples */ 182 float a[], /* LPCs for this frame of samples */
199 float a[], /* LPCs for this frame of samples */ 183 int Nsam, /* number of samples */
200 int Nsam, /* number of samples */ 184 float res[], /* Nsam residual samples */
201 float res[], /* Nsam residual samples */ 185 int order /* order of LPC */
202 int order /* order of LPC */ 186) {
203) 187 int i, j; /* loop variables */
204{ 188
205 int i,j; /* loop variables */ 189 for (i = 0; i < Nsam; i++) {
206
207 for(i=0; i<Nsam; i++) {
208 res[i] = 0.0; 190 res[i] = 0.0;
209 for(j=0; j<=order; j++) 191 for (j = 0; j <= order; j++) res[i] += Sn[i - j] * a[j];
210 res[i] += Sn[i-j]*a[j];
211 } 192 }
212} 193}
213 194
@@ -223,7 +204,7 @@ void inverse_filter(
223 The synthesis filter has memory as well, this is treated in the same way 204 The synthesis filter has memory as well, this is treated in the same way
224 as the memory for the inverse filter (see inverse_filter() notes above). 205 as the memory for the inverse filter (see inverse_filter() notes above).
225 The difference is that the memory for the synthesis filter is stored in 206 The difference is that the memory for the synthesis filter is stored in
226 the output array, wheras the memory of the inverse filter is stored in the 207 the output array, whereas the memory of the inverse filter is stored in the
227 input array. 208 input array.
228 209
229 Note: the calling function must update the filter memory. 210 Note: the calling function must update the filter memory.
@@ -231,21 +212,19 @@ void inverse_filter(
231\*---------------------------------------------------------------------------*/ 212\*---------------------------------------------------------------------------*/
232 213
233void synthesis_filter( 214void synthesis_filter(
234 float res[], /* Nsam input residual (excitation) samples */ 215 float res[], /* Nsam input residual (excitation) samples */
235 float a[], /* LPCs for this frame of speech samples */ 216 float a[], /* LPCs for this frame of speech samples */
236 int Nsam, /* number of speech samples */ 217 int Nsam, /* number of speech samples */
237 int order, /* LPC order */ 218 int order, /* LPC order */
238 float Sn_[] /* Nsam output synthesised speech samples */ 219 float Sn_[] /* Nsam output synthesised speech samples */
239) 220) {
240{ 221 int i, j; /* loop variables */
241 int i,j; /* loop variables */
242 222
243 /* Filter Nsam samples */ 223 /* Filter Nsam samples */
244 224
245 for(i=0; i<Nsam; i++) { 225 for (i = 0; i < Nsam; i++) {
246 Sn_[i] = res[i]*a[0]; 226 Sn_[i] = res[i] * a[0];
247 for(j=1; j<=order; j++) 227 for (j = 1; j <= order; j++) Sn_[i] -= Sn_[i - j] * a[j];
248 Sn_[i] -= Sn_[i-j]*a[j];
249 } 228 }
250} 229}
251 230
@@ -258,29 +237,25 @@ void synthesis_filter(
258 237
259\*---------------------------------------------------------------------------*/ 238\*---------------------------------------------------------------------------*/
260 239
261void find_aks( 240void find_aks(float Sn[], /* Nsam samples with order sample memory */
262 float Sn[], /* Nsam samples with order sample memory */ 241 float a[], /* order+1 LPCs with first coeff 1.0 */
263 float a[], /* order+1 LPCs with first coeff 1.0 */ 242 int Nsam, /* number of input speech samples */
264 int Nsam, /* number of input speech samples */ 243 int order, /* order of the LPC analysis */
265 int order, /* order of the LPC analysis */ 244 float *E /* residual energy */
266 float *E /* residual energy */ 245) {
267) 246 float Wn[LPC_MAX_N]; /* windowed frame of Nsam speech samples */
268{ 247 float R[order + 1]; /* order+1 autocorrelation values of Sn[] */
269 float Wn[LPC_MAX_N]; /* windowed frame of Nsam speech samples */
270 float R[order+1]; /* order+1 autocorrelation values of Sn[] */
271 int i; 248 int i;
272 249
273 assert(Nsam < LPC_MAX_N); 250 assert(Nsam < LPC_MAX_N);
274 251
275 hanning_window(Sn,Wn,Nsam); 252 hanning_window(Sn, Wn, Nsam);
276 autocorrelate(Wn,R,Nsam,order); 253 autocorrelate(Wn, R, Nsam, order);
277 levinson_durbin(R,a,order); 254 levinson_durbin(R, a, order);
278 255
279 *E = 0.0; 256 *E = 0.0;
280 for(i=0; i<=order; i++) 257 for (i = 0; i <= order; i++) *E += a[i] * R[i];
281 *E += a[i]*R[i]; 258 if (*E < 0.0) *E = 1E-12;
282 if (*E < 0.0)
283 *E = 1E-12;
284} 259}
285 260
286/*---------------------------------------------------------------------------*\ 261/*---------------------------------------------------------------------------*\
@@ -291,16 +266,12 @@ void find_aks(
291 266
292\*---------------------------------------------------------------------------*/ 267\*---------------------------------------------------------------------------*/
293 268
294void weight( 269void weight(float ak[], /* vector of order+1 LPCs */
295 float ak[], /* vector of order+1 LPCs */ 270 float gamma, /* weighting factor */
296 float gamma, /* weighting factor */ 271 int order, /* num LPCs (excluding leading 1.0) */
297 int order, /* num LPCs (excluding leading 1.0) */ 272 float akw[] /* weighted vector of order+1 LPCs */
298 float akw[] /* weighted vector of order+1 LPCs */ 273) {
299)
300{
301 int i; 274 int i;
302 275
303 for(i=1; i<=order; i++) 276 for (i = 1; i <= order; i++) akw[i] = ak[i] * powf(gamma, (float)i);
304 akw[i] = ak[i]*powf(gamma,(float)i);
305} 277}
306
diff --git a/lpc.h b/lpc.h
index 482aa1f..8efbe9c 100644
--- a/lpc.h
+++ b/lpc.h
@@ -32,12 +32,12 @@
32 32
33void pre_emp(float Sn_pre[], float Sn[], float *mem, int Nsam); 33void pre_emp(float Sn_pre[], float Sn[], float *mem, int Nsam);
34void de_emp(float Sn_se[], float Sn[], float *mem, int Nsam); 34void de_emp(float Sn_se[], float Sn[], float *mem, int Nsam);
35void hanning_window(float Sn[], float Wn[], int Nsam); 35void hanning_window(float Sn[], float Wn[], int Nsam);
36void autocorrelate(float Sn[], float Rn[], int Nsam, int order); 36void autocorrelate(float Sn[], float Rn[], int Nsam, int order);
37void levinson_durbin(float R[], float lpcs[], int order); 37void levinson_durbin(float R[], float lpcs[], int order);
38void inverse_filter(float Sn[], float a[], int Nsam, float res[], int order); 38void inverse_filter(float Sn[], float a[], int Nsam, float res[], int order);
39void synthesis_filter(float res[], float a[], int Nsam, int order, float Sn_[]); 39void synthesis_filter(float res[], float a[], int Nsam, int order, float Sn_[]);
40void find_aks(float Sn[], float a[], int Nsam, int order, float *E); 40void find_aks(float Sn[], float a[], int Nsam, int order, float *E);
41void weight(float ak[], float gamma, int order, float akw[]); 41void weight(float ak[], float gamma, int order, float akw[]);
42 42
43#endif 43#endif
diff --git a/lsp.c b/lsp.c
index 05d190e..aa7bac3 100644
--- a/lsp.c
+++ b/lsp.c
@@ -28,12 +28,14 @@
28 along with this program; if not, see <http://www.gnu.org/licenses/>. 28 along with this program; if not, see <http://www.gnu.org/licenses/>.
29*/ 29*/
30 30
31#include "defines.h"
32#include "lsp.h" 31#include "lsp.h"
32
33#include <math.h> 33#include <math.h>
34#include <stdio.h> 34#include <stdio.h>
35#include <stdlib.h> 35#include <stdlib.h>
36 36
37#include "defines.h"
38
37/*---------------------------------------------------------------------------*\ 39/*---------------------------------------------------------------------------*\
38 40
39 Introduction to Line Spectrum Pairs (LSPs) 41 Introduction to Line Spectrum Pairs (LSPs)
@@ -77,49 +79,45 @@
77 AUTHOR......: David Rowe 79 AUTHOR......: David Rowe
78 DATE CREATED: 24/2/93 80 DATE CREATED: 24/2/93
79 81
80 This function evalutes a series of chebyshev polynomials 82 This function evaluates a series of chebyshev polynomials
81 83
82 FIXME: performing memory allocation at run time is very inefficient, 84 FIXME: performing memory allocation at run time is very inefficient,
83 replace with stack variables of MAX_P size. 85 replace with stack variables of MAX_P size.
84 86
85\*---------------------------------------------------------------------------*/ 87\*---------------------------------------------------------------------------*/
86 88
87 89static float cheb_poly_eva(float *coef, float x, int order)
88static float
89cheb_poly_eva(float *coef,float x,int order)
90/* float coef[] coefficients of the polynomial to be evaluated */ 90/* float coef[] coefficients of the polynomial to be evaluated */
91/* float x the point where polynomial is to be evaluated */ 91/* float x the point where polynomial is to be evaluated */
92/* int order order of the polynomial */ 92/* int order order of the polynomial */
93{ 93{
94 int i; 94 int i;
95 float *t,*u,*v,sum; 95 float *t, *u, *v, sum;
96 float T[(order / 2) + 1]; 96 float T[(order / 2) + 1];
97 97
98 /* Initialise pointers */ 98 /* Initialise pointers */
99 99
100 t = T; /* T[i-2] */ 100 t = T; /* T[i-2] */
101 *t++ = 1.0; 101 *t++ = 1.0;
102 u = t--; /* T[i-1] */ 102 u = t--; /* T[i-1] */
103 *u++ = x; 103 *u++ = x;
104 v = u--; /* T[i] */ 104 v = u--; /* T[i] */
105 105
106 /* Evaluate chebyshev series formulation using iterative approach */ 106 /* Evaluate chebyshev series formulation using iterative approach */
107 107
108 for(i=2;i<=order/2;i++) 108 for (i = 2; i <= order / 2; i++)
109 *v++ = (2*x)*(*u++) - *t++; /* T[i] = 2*x*T[i-1] - T[i-2] */ 109 *v++ = (2 * x) * (*u++) - *t++; /* T[i] = 2*x*T[i-1] - T[i-2] */
110 110
111 sum=0.0; /* initialise sum to zero */ 111 sum = 0.0; /* initialise sum to zero */
112 t = T; /* reset pointer */ 112 t = T; /* reset pointer */
113 113
114 /* Evaluate polynomial and return value also free memory space */ 114 /* Evaluate polynomial and return value also free memory space */
115 115
116 for(i=0;i<=order/2;i++) 116 for (i = 0; i <= order / 2; i++) sum += coef[(order / 2) - i] * *t++;
117 sum+=coef[(order/2)-i]**t++;
118 117
119 return sum; 118 return sum;
120} 119}
121 120
122
123/*---------------------------------------------------------------------------*\ 121/*---------------------------------------------------------------------------*\
124 122
125 FUNCTION....: lpc_to_lsp() 123 FUNCTION....: lpc_to_lsp()
@@ -130,121 +128,118 @@ cheb_poly_eva(float *coef,float x,int order)
130 128
131\*---------------------------------------------------------------------------*/ 129\*---------------------------------------------------------------------------*/
132 130
133int lpc_to_lsp (float *a, int order, float *freq, int nb, float delta) 131int lpc_to_lsp(float *a, int order, float *freq, int nb, float delta)
134/* float *a lpc coefficients */ 132/* float *a lpc coefficients */
135/* int order order of LPC coefficients (10) */ 133/* int order order of LPC coefficients (10) */
136/* float *freq LSP frequencies in radians */ 134/* float *freq LSP frequencies in radians */
137/* int nb number of sub-intervals (4) */ 135/* int nb number of sub-intervals (4) */
138/* float delta grid spacing interval (0.02) */ 136/* float delta grid spacing interval (0.02) */
139{ 137{
140 float psuml,psumr,psumm,temp_xr,xl,xr,xm = 0; 138 float psuml, psumr, psumm, temp_xr, xl, xr, xm = 0;
141 float temp_psumr; 139 float temp_psumr;
142 int i,j,m,flag,k; 140 int i, j, m, flag, k;
143 float *px; /* ptrs of respective P'(z) & Q'(z) */ 141 float *px; /* ptrs of respective P'(z) & Q'(z) */
144 float *qx; 142 float *qx;
145 float *p; 143 float *p;
146 float *q; 144 float *q;
147 float *pt; /* ptr used for cheb_poly_eval() 145 float *pt; /* ptr used for cheb_poly_eval()
148 whether P' or Q' */ 146 whether P' or Q' */
149 int roots=0; /* number of roots found */ 147 int roots = 0; /* number of roots found */
150 float Q[order + 1]; 148 float Q[order + 1];
151 float P[order + 1]; 149 float P[order + 1];
152 150
151 flag = 1;
152 m = order / 2; /* order of P'(z) & Q'(z) polynimials */
153
154 /* Allocate memory space for polynomials */
155
156 /* determine P'(z)'s and Q'(z)'s coefficients where
157 P'(z) = P(z)/(1 + z^(-1)) and Q'(z) = Q(z)/(1-z^(-1)) */
158
159 px = P; /* initilaise ptrs */
160 qx = Q;
161 p = px;
162 q = qx;
163 *px++ = 1.0;
164 *qx++ = 1.0;
165 for (i = 1; i <= m; i++) {
166 *px++ = a[i] + a[order + 1 - i] - *p++;
167 *qx++ = a[i] - a[order + 1 - i] + *q++;
168 }
169 px = P;
170 qx = Q;
171 for (i = 0; i < m; i++) {
172 *px = 2 * *px;
173 *qx = 2 * *qx;
174 px++;
175 qx++;
176 }
177 px = P; /* re-initialise ptrs */
178 qx = Q;
179
180 /* Search for a zero in P'(z) polynomial first and then alternate to Q'(z).
181 Keep alternating between the two polynomials as each zero is found */
182
183 xr = 0; /* initialise xr to zero */
184 xl = 1.0; /* start at point xl = 1 */
185
186 for (j = 0; j < order; j++) {
187 if (j % 2) /* determines whether P' or Q' is eval. */
188 pt = qx;
189 else
190 pt = px;
191
192 psuml = cheb_poly_eva(pt, xl, order); /* evals poly. at xl */
153 flag = 1; 193 flag = 1;
154 m = order/2; /* order of P'(z) & Q'(z) polynimials */ 194 while (flag && (xr >= -1.0)) {
155 195 xr = xl - delta; /* interval spacing */
156 /* Allocate memory space for polynomials */ 196 psumr = cheb_poly_eva(pt, xr, order); /* poly(xl-delta_x) */
157 197 temp_psumr = psumr;
158 /* determine P'(z)'s and Q'(z)'s coefficients where 198 temp_xr = xr;
159 P'(z) = P(z)/(1 + z^(-1)) and Q'(z) = Q(z)/(1-z^(-1)) */ 199
160 200 /* if no sign change increment xr and re-evaluate
161 px = P; /* initilaise ptrs */ 201 poly(xr). Repeat til sign change. if a sign change has
162 qx = Q; 202 occurred the interval is bisected and then checked again
163 p = px; 203 for a sign change which determines in which interval the
164 q = qx; 204 zero lies in. If there is no sign change between poly(xm)
165 *px++ = 1.0; 205 and poly(xl) set interval between xm and xr else set
166 *qx++ = 1.0; 206 interval between xl and xr and repeat till root is located
167 for(i=1;i<=m;i++){ 207 within the specified limits */
168 *px++ = a[i]+a[order+1-i]-*p++; 208
169 *qx++ = a[i]-a[order+1-i]+*q++; 209 if (((psumr * psuml) < 0.0) || (psumr == 0.0)) {
170 } 210 roots++;
171 px = P; 211
172 qx = Q; 212 psumm = psuml;
173 for(i=0;i<m;i++){ 213 for (k = 0; k <= nb; k++) {
174 *px = 2**px; 214 xm = (xl + xr) / 2; /* bisect the interval */
175 *qx = 2**qx; 215 psumm = cheb_poly_eva(pt, xm, order);
176 px++; 216 if (psumm * psuml > 0.) {
177 qx++; 217 psuml = psumm;
178 } 218 xl = xm;
179 px = P; /* re-initialise ptrs */ 219 } else {
180 qx = Q; 220 psumr = psumm;
181 221 xr = xm;
182 /* Search for a zero in P'(z) polynomial first and then alternate to Q'(z). 222 }
183 Keep alternating between the two polynomials as each zero is found */ 223 }
184 224
185 xr = 0; /* initialise xr to zero */ 225 /* once zero is found, reset initial interval to xr */
186 xl = 1.0; /* start at point xl = 1 */ 226 freq[j] = (xm);
187 227 xl = xm;
188 228 flag = 0; /* reset flag for next search */
189 for(j=0;j<order;j++){ 229 } else {
190 if(j%2) /* determines whether P' or Q' is eval. */ 230 psuml = temp_psumr;
191 pt = qx; 231 xl = temp_xr;
192 else 232 }
193 pt = px;
194
195 psuml = cheb_poly_eva(pt,xl,order); /* evals poly. at xl */
196 flag = 1;
197 while(flag && (xr >= -1.0)){
198 xr = xl - delta ; /* interval spacing */
199 psumr = cheb_poly_eva(pt,xr,order);/* poly(xl-delta_x) */
200 temp_psumr = psumr;
201 temp_xr = xr;
202
203 /* if no sign change increment xr and re-evaluate
204 poly(xr). Repeat til sign change. if a sign change has
205 occurred the interval is bisected and then checked again
206 for a sign change which determines in which interval the
207 zero lies in. If there is no sign change between poly(xm)
208 and poly(xl) set interval between xm and xr else set
209 interval between xl and xr and repeat till root is located
210 within the specified limits */
211
212 if(((psumr*psuml)<0.0) || (psumr == 0.0)){
213 roots++;
214
215 psumm=psuml;
216 for(k=0;k<=nb;k++){
217 xm = (xl+xr)/2; /* bisect the interval */
218 psumm=cheb_poly_eva(pt,xm,order);
219 if(psumm*psuml>0.){
220 psuml=psumm;
221 xl=xm;
222 }
223 else{
224 psumr=psumm;
225 xr=xm;
226 }
227 }
228
229 /* once zero is found, reset initial interval to xr */
230 freq[j] = (xm);
231 xl = xm;
232 flag = 0; /* reset flag for next search */
233 }
234 else{
235 psuml=temp_psumr;
236 xl=temp_xr;
237 }
238 }
239 } 233 }
234 }
240 235
241 /* convert from x domain to radians */ 236 /* convert from x domain to radians */
242 237
243 for(i=0; i<order; i++) { 238 for (i = 0; i < order; i++) {
244 freq[i] = acosf(freq[i]); 239 freq[i] = acosf(freq[i]);
245 } 240 }
246 241
247 return(roots); 242 return (roots);
248} 243}
249 244
250/*---------------------------------------------------------------------------*\ 245/*---------------------------------------------------------------------------*\
@@ -263,59 +258,56 @@ void lsp_to_lpc(float *lsp, float *ak, int order)
263/* float *ak array of LPC coefficients */ 258/* float *ak array of LPC coefficients */
264/* int order order of LPC coefficients */ 259/* int order order of LPC coefficients */
265 260
266
267{ 261{
268 int i,j; 262 int i, j;
269 float xout1,xout2,xin1,xin2; 263 float xout1, xout2, xin1, xin2;
270 float *pw,*n1,*n2,*n3,*n4 = 0; 264 float *pw, *n1, *n2, *n3, *n4 = 0;
271 float freq[order]; 265 float freq[order];
272 float Wp[(order * 4) + 2]; 266 float Wp[(order * 4) + 2];
273 267
274 /* convert from radians to the x=cos(w) domain */ 268 /* convert from radians to the x=cos(w) domain */
275 269
276 for(i=0; i<order; i++) 270 for (i = 0; i < order; i++) freq[i] = cosf(lsp[i]);
277 freq[i] = cosf(lsp[i]); 271
278 272 pw = Wp;
279 pw = Wp; 273
280 274 /* initialise contents of array */
281 /* initialise contents of array */ 275
282 276 for (i = 0; i <= 4 * (order / 2) + 1; i++) { /* set contents of buffer to 0 */
283 for(i=0;i<=4*(order/2)+1;i++){ /* set contents of buffer to 0 */ 277 *pw++ = 0.0;
284 *pw++ = 0.0; 278 }
285 } 279
286 280 /* Set pointers up */
287 /* Set pointers up */ 281
288 282 pw = Wp;
289 pw = Wp; 283 xin1 = 1.0;
290 xin1 = 1.0; 284 xin2 = 1.0;
291 xin2 = 1.0; 285
292 286 /* reconstruct P(z) and Q(z) by cascading second order polynomials
293 /* reconstruct P(z) and Q(z) by cascading second order polynomials 287 in form 1 - 2xz(-1) +z(-2), where x is the LSP coefficient */
294 in form 1 - 2xz(-1) +z(-2), where x is the LSP coefficient */ 288
295 289 for (j = 0; j <= order; j++) {
296 for(j=0;j<=order;j++){ 290 for (i = 0; i < (order / 2); i++) {
297 for(i=0;i<(order/2);i++){ 291 n1 = pw + (i * 4);
298 n1 = pw+(i*4); 292 n2 = n1 + 1;
299 n2 = n1 + 1; 293 n3 = n2 + 1;
300 n3 = n2 + 1; 294 n4 = n3 + 1;
301 n4 = n3 + 1; 295 xout1 = xin1 - 2 * (freq[2 * i]) * *n1 + *n2;
302 xout1 = xin1 - 2*(freq[2*i]) * *n1 + *n2; 296 xout2 = xin2 - 2 * (freq[2 * i + 1]) * *n3 + *n4;
303 xout2 = xin2 - 2*(freq[2*i+1]) * *n3 + *n4; 297 *n2 = *n1;
304 *n2 = *n1; 298 *n4 = *n3;
305 *n4 = *n3; 299 *n1 = xin1;
306 *n1 = xin1; 300 *n3 = xin2;
307 *n3 = xin2; 301 xin1 = xout1;
308 xin1 = xout1; 302 xin2 = xout2;
309 xin2 = xout2;
310 }
311 xout1 = xin1 + *(n4+1);
312 xout2 = xin2 - *(n4+2);
313 ak[j] = (xout1 + xout2)*0.5;
314 *(n4+1) = xin1;
315 *(n4+2) = xin2;
316
317 xin1 = 0.0;
318 xin2 = 0.0;
319 } 303 }
304 xout1 = xin1 + *(n4 + 1);
305 xout2 = xin2 - *(n4 + 2);
306 ak[j] = (xout1 + xout2) * 0.5;
307 *(n4 + 1) = xin1;
308 *(n4 + 2) = xin2;
309
310 xin1 = 0.0;
311 xin2 = 0.0;
312 }
320} 313}
321
diff --git a/lsp.h b/lsp.h
index 5acef01..d473e18 100644
--- a/lsp.h
+++ b/lsp.h
@@ -31,7 +31,7 @@
31#ifndef __LSP__ 31#ifndef __LSP__
32#define __LSP__ 32#define __LSP__
33 33
34int lpc_to_lsp (float *a, int lpcrdr, float *freq, int nb, float delta); 34int lpc_to_lsp(float *a, int lpcrdr, float *freq, int nb, float delta);
35void lsp_to_lpc(float *freq, float *ak, int lpcrdr); 35void lsp_to_lpc(float *freq, float *ak, int lpcrdr);
36 36
37#endif 37#endif
diff --git a/machdep.h b/machdep.h
index 4dff9ba..e17b5fa 100644
--- a/machdep.h
+++ b/machdep.h
@@ -4,7 +4,7 @@
4 AUTHOR......: David Rowe 4 AUTHOR......: David Rowe
5 DATE CREATED: May 2 2013 5 DATE CREATED: May 2 2013
6 6
7 Machine dependant functions, e.g. profiling that requires access to a clock 7 Machine dependent functions, e.g. profiling that requires access to a clock
8 counter register. 8 counter register.
9 9
10\*---------------------------------------------------------------------------*/ 10\*---------------------------------------------------------------------------*/
@@ -33,9 +33,9 @@
33#define PROFILE_VAR(...) unsigned int __VA_ARGS__ 33#define PROFILE_VAR(...) unsigned int __VA_ARGS__
34#define PROFILE_SAMPLE(timestamp) timestamp = machdep_profile_sample() 34#define PROFILE_SAMPLE(timestamp) timestamp = machdep_profile_sample()
35#define PROFILE_SAMPLE_AND_LOG(timestamp, prev_timestamp, label) \ 35#define PROFILE_SAMPLE_AND_LOG(timestamp, prev_timestamp, label) \
36 timestamp = machdep_profile_sample_and_log(prev_timestamp, label) 36 timestamp = machdep_profile_sample_and_log(prev_timestamp, label)
37#define PROFILE_SAMPLE_AND_LOG2(prev_timestamp, label) \ 37#define PROFILE_SAMPLE_AND_LOG2(prev_timestamp, label) \
38 machdep_profile_sample_and_log(prev_timestamp, label) 38 machdep_profile_sample_and_log(prev_timestamp, label)
39#else 39#else
40#define PROFILE_VAR(...) 40#define PROFILE_VAR(...)
41#define PROFILE_SAMPLE(timestamp) 41#define PROFILE_SAMPLE(timestamp)
@@ -43,10 +43,10 @@
43#define PROFILE_SAMPLE_AND_LOG2(prev_timestamp, label) 43#define PROFILE_SAMPLE_AND_LOG2(prev_timestamp, label)
44#endif 44#endif
45 45
46void machdep_profile_init(void); 46void machdep_profile_init(void);
47void machdep_profile_reset(void); 47void machdep_profile_reset(void);
48unsigned int machdep_profile_sample(void); 48unsigned int machdep_profile_sample(void);
49unsigned int machdep_profile_sample_and_log(unsigned int start, char s[]); 49unsigned int machdep_profile_sample_and_log(unsigned int start, char s[]);
50void machdep_profile_print_logged_samples(void); 50void machdep_profile_print_logged_samples(void);
51 51
52#endif 52#endif
diff --git a/main.c b/main.c
index 10a0818..359c45b 100644
--- a/main.c
+++ b/main.c
@@ -14,6 +14,8 @@ int main() {
14 int off = 0; 14 int off = 0;
15 unsigned char bits[128]; 15 unsigned char bits[128];
16 16
17 fprintf(stderr, "%d samples per frame, %d bits per frame\n", nsam, nbit);
18
17 fread(input, 976692, 1, f); 19 fread(input, 976692, 1, f);
18 fclose(f); 20 fclose(f);
19 21
diff --git a/mbest.c b/mbest.c
index 6216b77..9c1a6c5 100644
--- a/mbest.c
+++ b/mbest.c
@@ -27,42 +27,48 @@
27 27
28*/ 28*/
29 29
30#include "mbest.h"
31
30#include <assert.h> 32#include <assert.h>
31#include <math.h> 33#include <math.h>
32#include <stdio.h> 34#include <stdio.h>
33#include <stdlib.h> 35#include <stdlib.h>
34#include <string.h> 36#include <string.h>
35 37
36#include "mbest.h"
37
38struct MBEST *mbest_create(int entries) { 38struct MBEST *mbest_create(int entries) {
39 int i,j; 39 int i, j;
40 struct MBEST *mbest; 40 struct MBEST *mbest;
41 41
42 assert(entries > 0); 42 assert(entries > 0);
43 mbest = (struct MBEST *)malloc(sizeof(struct MBEST)); 43 mbest = (struct MBEST *)malloc(sizeof(struct MBEST));
44 assert(mbest != NULL); 44 assert(mbest != NULL);
45 45
46 mbest->entries = entries; 46 mbest->entries = entries;
47 mbest->list = (struct MBEST_LIST *)malloc(entries*sizeof(struct MBEST_LIST)); 47 mbest->list =
48 assert(mbest->list != NULL); 48 (struct MBEST_LIST *)malloc(entries * sizeof(struct MBEST_LIST));
49 assert(mbest->list != NULL);
49 50
50 for(i=0; i<mbest->entries; i++) { 51 for (i = 0; i < mbest->entries; i++) {
51 for(j=0; j<MBEST_STAGES; j++) 52 for (j = 0; j < MBEST_STAGES; j++) mbest->list[i].index[j] = 0;
52 mbest->list[i].index[j] = 0; 53 mbest->list[i].error = 1E32;
53 mbest->list[i].error = 1E32; 54 }
54 }
55 55
56 return mbest; 56 return mbest;
57} 57}
58 58
59
60void mbest_destroy(struct MBEST *mbest) { 59void mbest_destroy(struct MBEST *mbest) {
61 assert(mbest != NULL); 60 assert(mbest != NULL);
62 free(mbest->list); 61 free(mbest->list);
63 free(mbest); 62 free(mbest);
64} 63}
65 64
65/* apply weighting to VQ for efficient VQ search */
66
67void mbest_precompute_weight(float cb[], float w[], int k, int m) {
68 for (int j = 0; j < m; j++) {
69 for (int i = 0; i < k; i++) cb[k * j + i] *= w[i];
70 }
71}
66 72
67/*---------------------------------------------------------------------------*\ 73/*---------------------------------------------------------------------------*\
68 74
@@ -75,36 +81,31 @@ void mbest_destroy(struct MBEST *mbest) {
75\*---------------------------------------------------------------------------*/ 81\*---------------------------------------------------------------------------*/
76 82
77void mbest_insert(struct MBEST *mbest, int index[], float error) { 83void mbest_insert(struct MBEST *mbest, int index[], float error) {
78 int i, j, found; 84 int i, found;
79 struct MBEST_LIST *list = mbest->list; 85 struct MBEST_LIST *list = mbest->list;
80 int entries = mbest->entries; 86 int entries = mbest->entries;
81 87
82 found = 0; 88 found = 0;
83 for(i=0; i<entries && !found; i++) 89 for (i = 0; i < entries && !found; i++)
84 if (error < list[i].error) { 90 if (error < list[i].error) {
85 found = 1; 91 found = 1;
86 for(j=entries-1; j>i; j--) 92 memmove(&list[i + 1], &list[i],
87 list[j] = list[j-1]; 93 sizeof(struct MBEST_LIST) * (entries - i - 1));
88 for(j=0; j<MBEST_STAGES; j++) 94 memcpy(&list[i].index[0], &index[0], sizeof(int) * MBEST_STAGES);
89 list[i].index[j] = index[j]; 95 list[i].error = error;
90 list[i].error = error; 96 }
91 }
92} 97}
93 98
99void mbest_print(char title[], struct MBEST *mbest) {
100 int i, j;
94 101
95#ifdef MBEST_PRINT_OUT 102 fprintf(stderr, "%s\n", title);
96static void mbest_print(char title[], struct MBEST *mbest) { 103 for (i = 0; i < mbest->entries; i++) {
97 int i,j; 104 for (j = 0; j < MBEST_STAGES; j++)
98 105 fprintf(stderr, " %4d ", mbest->list[i].index[j]);
99 fprintf(stderr, "%s\n", title); 106 fprintf(stderr, " %f\n", (double)mbest->list[i].error);
100 for(i=0; i<mbest->entries; i++) { 107 }
101 for(j=0; j<MBEST_STAGES; j++)
102 fprintf(stderr, " %4d ", mbest->list[i].index[j]);
103 fprintf(stderr, " %f\n", mbest->list[i].error);
104 }
105} 108}
106#endif
107
108 109
109/*---------------------------------------------------------------------------*\ 110/*---------------------------------------------------------------------------*\
110 111
@@ -115,31 +116,48 @@ static void mbest_print(char title[], struct MBEST *mbest) {
115 116
116\*---------------------------------------------------------------------------*/ 117\*---------------------------------------------------------------------------*/
117 118
118void mbest_search( 119void mbest_search(const float *cb, /* VQ codebook to search */
119 const float *cb, /* VQ codebook to search */ 120 float vec[], /* target vector */
120 float vec[], /* target vector */ 121 int k, /* dimension of vector */
121 float w[], /* weighting vector */ 122 int m, /* number on entries in codebook */
122 int k, /* dimension of vector */ 123 struct MBEST *mbest, /* list of closest matches */
123 int m, /* number on entries in codebook */ 124 int index[] /* indexes that lead us here */
124 struct MBEST *mbest, /* list of closest matches */ 125) {
125 int index[] /* indexes that lead us here */ 126 int j;
126) 127
127{ 128 /* note weighting can be applied externally by modifying cb[] and vec:
128 float e; 129
129 int i,j; 130 float e = 0.0;
130 float diff; 131 for(i=0; i<k; i++)
131 132 e += pow(w[i]*(cb[j*k+i] - vec[i]),2.0)
132 for(j=0; j<m; j++) { 133
133 e = 0.0; 134 |
134 for(i=0; i<k; i++) { 135 \|/
135 diff = cb[j*k+i]-vec[i]; 136
136 e += diff*w[i]*diff*w[i]; 137 for(i=0; i<k; i++)
137 } 138 e += pow(w[i]*cb[j*k+i] - w[i]*vec[i]),2.0)
138 index[0] = j; 139
139 mbest_insert(mbest, index, e); 140 |
140 } 141 \|/
141}
142 142
143 for(i=0; i<k; i++)
144 e += pow(cb1[j*k+i] - vec1[i]),2.0)
145
146 where cb1[j*k+i] = w[i]*cb[j*k+i], and vec1[i] = w[i]*vec[i]
147 */
148
149 for (j = 0; j < m; j++) {
150 float e = 0.0;
151 for (int i = 0; i < k; i++) {
152 float diff = *cb++ - vec[i];
153 e += diff * diff;
154 }
155
156 index[0] = j;
157 if (e < mbest->list[mbest->entries - 1].error)
158 mbest_insert(mbest, index, e);
159 }
160}
143 161
144/*---------------------------------------------------------------------------*\ 162/*---------------------------------------------------------------------------*\
145 163
@@ -148,26 +166,26 @@ void mbest_search(
148 Searches vec[] to a codebbook of vectors, and maintains a list of the mbest 166 Searches vec[] to a codebbook of vectors, and maintains a list of the mbest
149 closest matches. Only searches the first NewAmp2_K Vectors 167 closest matches. Only searches the first NewAmp2_K Vectors
150 168
151 \*---------------------------------------------------------------------------*/ 169\*---------------------------------------------------------------------------*/
152 170
153void mbest_search450(const float *cb, float vec[], float w[], int k,int shorterK, int m, struct MBEST *mbest, int index[]) 171void mbest_search450(const float *cb, float vec[], float w[], int k,
172 int shorterK, int m, struct MBEST *mbest, int index[])
154 173
155{ 174{
156 float e; 175 float e;
157 int i,j; 176 int i, j;
158 float diff; 177 float diff;
159 178
160 for(j=0; j<m; j++) { 179 for (j = 0; j < m; j++) {
161 e = 0.0; 180 e = 0.0;
162 for(i=0; i<k; i++) { 181 for (i = 0; i < k; i++) {
163 //Only search first NEWAMP2_K Vectors 182 // Only search first NEWAMP2_K Vectors
164 if(i<shorterK){ 183 if (i < shorterK) {
165 diff = cb[j*k+i]-vec[i]; 184 diff = cb[j * k + i] - vec[i];
166 e += diff*w[i] * diff*w[i]; 185 e += diff * w[i] * diff * w[i];
167 } 186 }
168 }
169 index[0] = j;
170 mbest_insert(mbest, index, e);
171 } 187 }
188 index[0] = j;
189 mbest_insert(mbest, index, e);
190 }
172} 191}
173
diff --git a/mbest.h b/mbest.h
index fb7664f..2851fb9 100644
--- a/mbest.h
+++ b/mbest.h
@@ -33,27 +33,24 @@
33#define MBEST_STAGES 4 33#define MBEST_STAGES 4
34 34
35struct MBEST_LIST { 35struct MBEST_LIST {
36 int index[MBEST_STAGES]; /* index of each stage that lead us to this error */ 36 int index[MBEST_STAGES]; /* index of each stage that lead us to this error */
37 float error; 37 float error;
38}; 38};
39 39
40struct MBEST { 40struct MBEST {
41 int entries; /* number of entries in mbest list */ 41 int entries; /* number of entries in mbest list */
42 struct MBEST_LIST *list; 42 struct MBEST_LIST *list;
43}; 43};
44 44
45struct MBEST *mbest_create(int entries); 45struct MBEST *mbest_create(int entries);
46void mbest_destroy(struct MBEST *mbest); 46void mbest_destroy(struct MBEST *mbest);
47void mbest_precompute_weight(float cb[], float w[], int k, int m);
47void mbest_insert(struct MBEST *mbest, int index[], float error); 48void mbest_insert(struct MBEST *mbest, int index[], float error);
48void mbest_search(const float *cb, float vec[], float w[], int k, int m, struct MBEST *mbest, int index[]); 49void mbest_search(const float *cb, float vec[], int k, int m,
49void mbest_search450(const float *cb, float vec[], float w[], int k,int shorterK, int m, struct MBEST *mbest, int index[]); 50 struct MBEST *mbest, int index[]);
50 51void mbest_search450(const float *cb, float vec[], float w[], int k,
51// #define MBEST_PRINT_OUT 52 int shorterK, int m, struct MBEST *mbest, int index[]);
52#ifdef MBEST_PRINT_OUT
53 #define MBEST_PRINT(a,b) mbest_print((a),(b))
54#else
55 #define MBEST_PRINT(a,b)
56#endif
57 53
54void mbest_print(char title[], struct MBEST *mbest);
58 55
59#endif 56#endif
diff --git a/newamp1.c b/newamp1.c
index 8980ac6..3ba2de0 100644
--- a/newamp1.c
+++ b/newamp1.c
@@ -28,19 +28,18 @@
28 28
29*/ 29*/
30 30
31#include "newamp1.h"
32
31#include <assert.h> 33#include <assert.h>
34#include <math.h>
32#include <stdio.h> 35#include <stdio.h>
33#include <stdlib.h> 36#include <stdlib.h>
34#include <string.h> 37#include <string.h>
35#include <math.h>
36 38
37#include "defines.h" 39#include "defines.h"
40#include "mbest.h"
38#include "phase.h" 41#include "phase.h"
39#include "quantise.h" 42#include "quantise.h"
40#include "mbest.h"
41#include "newamp1.h"
42
43#define NEWAMP1_VQ_MBEST_DEPTH 5 /* how many candidates we keep for each stage of mbest search */
44 43
45/*---------------------------------------------------------------------------*\ 44/*---------------------------------------------------------------------------*\
46 45
@@ -48,39 +47,44 @@
48 AUTHOR......: David Rowe 47 AUTHOR......: David Rowe
49 DATE CREATED: Jan 2017 48 DATE CREATED: Jan 2017
50 49
51 General 2nd order parabolic interpolator. Used splines orginally, 50 General 2nd order parabolic interpolator. Used splines originally,
52 but this is much simpler and we don't need much accuracy. Given two 51 but this is much simpler and we don't need much accuracy. Given two
53 vectors of points xp and yp, find interpolated values y at points x. 52 vectors of points xp and yp, find interpolated values y at points x.
54 53
55\*---------------------------------------------------------------------------*/ 54\*---------------------------------------------------------------------------*/
56 55
57void interp_para(float y[], float xp[], float yp[], int np, float x[], int n) 56void interp_para(float y[], float xp[], float yp[], int np, float x[], int n) {
58{ 57 assert(np >= 3);
59 assert(np >= 3);
60 58
61 int k,i; 59 int k, i;
62 float xi, x1, y1, x2, y2, x3, y3, a, b; 60 float xi, x1, y1, x2, y2, x3, y3, a, b;
63 61
64 k = 0; 62 k = 0;
65 for (i=0; i<n; i++) { 63 for (i = 0; i < n; i++) {
66 xi = x[i]; 64 xi = x[i];
67 65
68 /* k is index into xp of where we start 3 points used to form parabola */ 66 /* k is index into xp of where we start 3 points used to form parabola */
69 67
70 while ((xp[k+1] < xi) && (k < (np-3))) 68 while ((xp[k + 1] < xi) && (k < (np - 3))) k++;
71 k++;
72
73 x1 = xp[k]; y1 = yp[k]; x2 = xp[k+1]; y2 = yp[k+1]; x3 = xp[k+2]; y3 = yp[k+2];
74 69
75 //printf("k: %d np: %d i: %d xi: %f x1: %f y1: %f\n", k, np, i, xi, x1, y1); 70 x1 = xp[k];
71 y1 = yp[k];
72 x2 = xp[k + 1];
73 y2 = yp[k + 1];
74 x3 = xp[k + 2];
75 y3 = yp[k + 2];
76 76
77 a = ((y3-y2)/(x3-x2)-(y2-y1)/(x2-x1))/(x3-x1); 77 // printf("k: %d np: %d i: %d xi: %f x1: %f y1: %f\n", k, np, i, xi, x1,
78 b = ((y3-y2)/(x3-x2)*(x2-x1)+(y2-y1)/(x2-x1)*(x3-x2))/(x3-x1); 78 // y1);
79
80 y[i] = a*(xi-x2)*(xi-x2) + b*(xi-x2) + y2;
81 }
82}
83 79
80 a = ((y3 - y2) / (x3 - x2) - (y2 - y1) / (x2 - x1)) / (x3 - x1);
81 b = ((y3 - y2) / (x3 - x2) * (x2 - x1) +
82 (y2 - y1) / (x2 - x1) * (x3 - x2)) /
83 (x3 - x1);
84
85 y[i] = a * (xi - x2) * (xi - x2) + b * (xi - x2) + y2;
86 }
87}
84 88
85/*---------------------------------------------------------------------------*\ 89/*---------------------------------------------------------------------------*\
86 90
@@ -94,24 +98,23 @@ void interp_para(float y[], float xp[], float yp[], int np, float x[], int n)
94\*---------------------------------------------------------------------------*/ 98\*---------------------------------------------------------------------------*/
95 99
96float ftomel(float fHz) { 100float ftomel(float fHz) {
97 float mel = floorf(2595.0*log10f(1.0 + fHz/700.0)+0.5); 101 float mel = floorf(2595.0 * log10f(1.0 + fHz / 700.0) + 0.5);
98 return mel; 102 return mel;
99} 103}
100 104
101void mel_sample_freqs_kHz(float rate_K_sample_freqs_kHz[], int K, float mel_start, float mel_end) 105void mel_sample_freqs_kHz(float rate_K_sample_freqs_kHz[], int K,
102{ 106 float mel_start, float mel_end) {
103 float step = (mel_end-mel_start)/(K-1); 107 float step = (mel_end - mel_start) / (K - 1);
104 float mel; 108 float mel;
105 int k; 109 int k;
106 110
107 mel = mel_start; 111 mel = mel_start;
108 for (k=0; k<K; k++) { 112 for (k = 0; k < K; k++) {
109 rate_K_sample_freqs_kHz[k] = 0.7*(POW10F(mel/2595.0) - 1.0); 113 rate_K_sample_freqs_kHz[k] = 0.7 * (POW10F(mel / 2595.0) - 1.0);
110 mel += step; 114 mel += step;
111 } 115 }
112} 116}
113 117
114
115/*---------------------------------------------------------------------------*\ 118/*---------------------------------------------------------------------------*\
116 119
117 FUNCTION....: resample_const_rate_f() 120 FUNCTION....: resample_const_rate_f()
@@ -122,35 +125,36 @@ void mel_sample_freqs_kHz(float rate_K_sample_freqs_kHz[], int K, float mel_star
122 125
123\*---------------------------------------------------------------------------*/ 126\*---------------------------------------------------------------------------*/
124 127
125void resample_const_rate_f(C2CONST *c2const, MODEL *model, float rate_K_vec[], float rate_K_sample_freqs_kHz[], int K) 128void resample_const_rate_f(C2CONST *c2const, MODEL *model, float rate_K_vec[],
126{ 129 float rate_K_sample_freqs_kHz[], int K) {
127 int m; 130 int m;
128 float AmdB[MAX_AMP+1], rate_L_sample_freqs_kHz[MAX_AMP+1], AmdB_peak; 131 float AmdB[MAX_AMP + 1], rate_L_sample_freqs_kHz[MAX_AMP + 1], AmdB_peak;
129 132
130 /* convert rate L=pi/Wo amplitude samples to fixed rate K */ 133 /* convert rate L=pi/Wo amplitude samples to fixed rate K */
131 134
132 AmdB_peak = -100.0; 135 AmdB_peak = -100.0;
133 for(m=1; m<=model->L; m++) { 136 for (m = 1; m <= model->L; m++) {
134 AmdB[m] = 20.0*log10f(model->A[m]+1E-16); 137 AmdB[m] = 20.0 * log10f(model->A[m] + 1E-16);
135 if (AmdB[m] > AmdB_peak) { 138 if (AmdB[m] > AmdB_peak) {
136 AmdB_peak = AmdB[m]; 139 AmdB_peak = AmdB[m];
137 }
138 rate_L_sample_freqs_kHz[m] = m*model->Wo*(c2const->Fs/2000.0)/M_PI;
139 //printf("m: %d AmdB: %f AmdB_peak: %f sf: %f\n", m, AmdB[m], AmdB_peak, rate_L_sample_freqs_kHz[m]);
140 } 140 }
141 141 rate_L_sample_freqs_kHz[m] = m * model->Wo * (c2const->Fs / 2000.0) / M_PI;
142 /* clip between peak and peak -50dB, to reduce dynamic range */ 142 // printf("m: %d AmdB: %f AmdB_peak: %f sf: %f\n", m, AmdB[m], AmdB_peak,
143 // rate_L_sample_freqs_kHz[m]);
144 }
145
146 /* clip between peak and peak -50dB, to reduce dynamic range */
143 147
144 for(m=1; m<=model->L; m++) { 148 for (m = 1; m <= model->L; m++) {
145 if (AmdB[m] < (AmdB_peak-50.0)) { 149 if (AmdB[m] < (AmdB_peak - 50.0)) {
146 AmdB[m] = AmdB_peak-50.0; 150 AmdB[m] = AmdB_peak - 50.0;
147 }
148 } 151 }
152 }
149 153
150 interp_para(rate_K_vec, &rate_L_sample_freqs_kHz[1], &AmdB[1], model->L, rate_K_sample_freqs_kHz, K); 154 interp_para(rate_K_vec, &rate_L_sample_freqs_kHz[1], &AmdB[1], model->L,
155 rate_K_sample_freqs_kHz, K);
151} 156}
152 157
153
154/*---------------------------------------------------------------------------*\ 158/*---------------------------------------------------------------------------*\
155 159
156 FUNCTION....: rate_K_mbest_encode 160 FUNCTION....: rate_K_mbest_encode
@@ -161,64 +165,55 @@ void resample_const_rate_f(C2CONST *c2const, MODEL *model, float rate_K_vec[], f
161 165
162\*---------------------------------------------------------------------------*/ 166\*---------------------------------------------------------------------------*/
163 167
164float rate_K_mbest_encode(int *indexes, float *x, float *xq, int ndim, int mbest_entries) 168float rate_K_mbest_encode(int *indexes, float *x, float *xq, int ndim,
165{ 169 int mbest_entries) {
166 int i, j, n1, n2; 170 int i, j, n1, n2;
167 const float *codebook1 = newamp1vq_cb[0].cb; 171 const float *codebook1 = newamp1vq_cb[0].cb;
168 const float *codebook2 = newamp1vq_cb[1].cb; 172 const float *codebook2 = newamp1vq_cb[1].cb;
169 struct MBEST *mbest_stage1, *mbest_stage2; 173 struct MBEST *mbest_stage1, *mbest_stage2;
170 float target[ndim]; 174 float target[ndim];
171 float w[ndim]; 175 int index[MBEST_STAGES];
172 int index[MBEST_STAGES];
173 float mse, tmp; 176 float mse, tmp;
174 177
175 /* codebook is compiled for a fixed K */ 178 /* codebook is compiled for a fixed K */
176 179
177 assert(ndim == newamp1vq_cb[0].k); 180 assert(ndim == newamp1vq_cb[0].k);
178 181
179 /* equal weights, could be argued mel freq axis gives freq dep weighting */
180
181 for(i=0; i<ndim; i++)
182 w[i] = 1.0;
183
184 mbest_stage1 = mbest_create(mbest_entries); 182 mbest_stage1 = mbest_create(mbest_entries);
185 mbest_stage2 = mbest_create(mbest_entries); 183 mbest_stage2 = mbest_create(mbest_entries);
186 for(i=0; i<MBEST_STAGES; i++) 184 for (i = 0; i < MBEST_STAGES; i++) index[i] = 0;
187 index[i] = 0;
188 185
189 /* Stage 1 */ 186 /* Stage 1 */
190 187
191 mbest_search(codebook1, x, w, ndim, newamp1vq_cb[0].m, mbest_stage1, index); 188 mbest_search(codebook1, x, ndim, newamp1vq_cb[0].m, mbest_stage1, index);
192 MBEST_PRINT("Stage 1:", mbest_stage1);
193 189
194 /* Stage 2 */ 190 /* Stage 2 */
195 191
196 for (j=0; j<mbest_entries; j++) { 192 for (j = 0; j < mbest_entries; j++) {
197 index[1] = n1 = mbest_stage1->list[j].index[0]; 193 index[1] = n1 = mbest_stage1->list[j].index[0];
198 for(i=0; i<ndim; i++) 194 for (i = 0; i < ndim; i++) target[i] = x[i] - codebook1[ndim * n1 + i];
199 target[i] = x[i] - codebook1[ndim*n1+i]; 195 mbest_search(codebook2, target, ndim, newamp1vq_cb[1].m, mbest_stage2,
200 mbest_search(codebook2, target, w, ndim, newamp1vq_cb[1].m, mbest_stage2, index); 196 index);
201 } 197 }
202 MBEST_PRINT("Stage 2:", mbest_stage2);
203 198
204 n1 = mbest_stage2->list[0].index[1]; 199 n1 = mbest_stage2->list[0].index[1];
205 n2 = mbest_stage2->list[0].index[0]; 200 n2 = mbest_stage2->list[0].index[0];
206 mse = 0.0; 201 mse = 0.0;
207 for (i=0;i<ndim;i++) { 202 for (i = 0; i < ndim; i++) {
208 tmp = codebook1[ndim*n1+i] + codebook2[ndim*n2+i]; 203 tmp = codebook1[ndim * n1 + i] + codebook2[ndim * n2 + i];
209 mse += (x[i]-tmp)*(x[i]-tmp); 204 mse += (x[i] - tmp) * (x[i] - tmp);
210 xq[i] = tmp; 205 xq[i] = tmp;
211 } 206 }
212 207
213 mbest_destroy(mbest_stage1); 208 mbest_destroy(mbest_stage1);
214 mbest_destroy(mbest_stage2); 209 mbest_destroy(mbest_stage2);
215 210
216 indexes[0] = n1; indexes[1] = n2; 211 indexes[0] = n1;
212 indexes[1] = n2;
217 213
218 return mse; 214 return mse;
219} 215}
220 216
221
222/*---------------------------------------------------------------------------*\ 217/*---------------------------------------------------------------------------*\
223 218
224 FUNCTION....: post_filter 219 FUNCTION....: post_filter
@@ -226,7 +221,7 @@ float rate_K_mbest_encode(int *indexes, float *x, float *xq, int ndim, int mbest
226 DATE CREATED: Jan 2017 221 DATE CREATED: Jan 2017
227 222
228 Post Filter, has a big impact on speech quality after VQ. When used 223 Post Filter, has a big impact on speech quality after VQ. When used
229 on a mean removed rate K vector, it raises formants, and supresses 224 on a mean removed rate K vector, it raises formants, and suppresses
230 anti-formants. As it manipulates amplitudes, we normalise energy to 225 anti-formants. As it manipulates amplitudes, we normalise energy to
231 prevent clipping or large level variations. pf_gain of 1.2 to 1.5 226 prevent clipping or large level variations. pf_gain of 1.2 to 1.5
232 (dB) seems to work OK. Good area for further investigations and 227 (dB) seems to work OK. Good area for further investigations and
@@ -234,37 +229,36 @@ float rate_K_mbest_encode(int *indexes, float *x, float *xq, int ndim, int mbest
234 229
235\*---------------------------------------------------------------------------*/ 230\*---------------------------------------------------------------------------*/
236 231
237void post_filter_newamp1(float vec[], float sample_freq_kHz[], int K, float pf_gain) 232void post_filter_newamp1(float vec[], float sample_freq_kHz[], int K,
238{ 233 float pf_gain) {
239 int k; 234 int k;
240 235
241 /* 236 /*
242 vec is rate K vector describing spectrum of current frame lets 237 vec is rate K vector describing spectrum of current frame lets
243 pre-emp before applying PF. 20dB/dec over 300Hz. Postfilter 238 pre-emp before applying PF. 20dB/dec over 300Hz. Postfilter
244 affects energy of frame so we measure energy before and after 239 affects energy of frame so we measure energy before and after
245 and normalise. Plenty of room for experiment here as well. 240 and normalise. Plenty of room for experimentation here.
246 */ 241 */
247 242
248 float pre[K]; 243 float pre[K];
249 float e_before = 0.0; 244 float e_before = 0.0;
250 float e_after = 0.0; 245 float e_after = 0.0;
251 for(k=0; k<K; k++) { 246 for (k = 0; k < K; k++) {
252 pre[k] = 20.0*log10f(sample_freq_kHz[k]/0.3); 247 pre[k] = 20.0 * log10f(sample_freq_kHz[k] / 0.3);
253 vec[k] += pre[k]; 248 vec[k] += pre[k];
254 e_before += POW10F(vec[k]/10.0); 249 e_before += POW10F(vec[k] / 10.0);
255 vec[k] *= pf_gain; 250 vec[k] *= pf_gain;
256 e_after += POW10F(vec[k]/10.0); 251 e_after += POW10F(vec[k] / 10.0);
257 } 252 }
258 253
259 float gain = e_after/e_before; 254 float gain = e_after / e_before;
260 float gaindB = 10*log10f(gain); 255 float gaindB = 10 * log10f(gain);
261
262 for(k=0; k<K; k++) {
263 vec[k] -= gaindB;
264 vec[k] -= pre[k];
265 }
266}
267 256
257 for (k = 0; k < K; k++) {
258 vec[k] -= gaindB;
259 vec[k] -= pre[k];
260 }
261}
268 262
269/*---------------------------------------------------------------------------*\ 263/*---------------------------------------------------------------------------*\
270 264
@@ -273,49 +267,46 @@ void post_filter_newamp1(float vec[], float sample_freq_kHz[], int K, float pf_g
273 DATE CREATED: Jan 2017 267 DATE CREATED: Jan 2017
274 268
275 Decoder side interpolation of Wo and voicing, to go from 25 Hz 269 Decoder side interpolation of Wo and voicing, to go from 25 Hz
276 sample rate used over channle to 100Hz internal sample rate of Codec 2. 270 sample rate used over channel to 100Hz internal sample rate of Codec 2.
277 271
278\*---------------------------------------------------------------------------*/ 272\*---------------------------------------------------------------------------*/
279 273
280void interp_Wo_v(float Wo_[], int L_[], int voicing_[], float Wo1, float Wo2, int voicing1, int voicing2) 274void interp_Wo_v(float Wo_[], int L_[], int voicing_[], float Wo1, float Wo2,
281{ 275 int voicing1, int voicing2) {
282 int i; 276 int i;
283 int M = 4; /* interpolation rate */ 277 int M = 4; /* interpolation rate */
284 278
285 for(i=0; i<M; i++) 279 for (i = 0; i < M; i++) voicing_[i] = 0;
286 voicing_[i] = 0;
287 280
288 if (!voicing1 && !voicing2) { 281 if (!voicing1 && !voicing2) {
289 for(i=0; i<M; i++) 282 for (i = 0; i < M; i++) Wo_[i] = 2.0 * M_PI / 100.0;
290 Wo_[i] = 2.0*M_PI/100.0; 283 }
291 }
292 284
293 if (voicing1 && !voicing2) { 285 if (voicing1 && !voicing2) {
294 Wo_[0] = Wo_[1] = Wo1; 286 Wo_[0] = Wo_[1] = Wo1;
295 Wo_[2] = Wo_[3] = 2.0*M_PI/100.0; 287 Wo_[2] = Wo_[3] = 2.0 * M_PI / 100.0;
296 voicing_[0] = voicing_[1] = 1; 288 voicing_[0] = voicing_[1] = 1;
297 } 289 }
298 290
299 if (!voicing1 && voicing2) { 291 if (!voicing1 && voicing2) {
300 Wo_[0] = Wo_[1] = 2.0*M_PI/100.0; 292 Wo_[0] = Wo_[1] = 2.0 * M_PI / 100.0;
301 Wo_[2] = Wo_[3] = Wo2; 293 Wo_[2] = Wo_[3] = Wo2;
302 voicing_[2] = voicing_[3] = 1; 294 voicing_[2] = voicing_[3] = 1;
303 } 295 }
304 296
305 if (voicing1 && voicing2) { 297 if (voicing1 && voicing2) {
306 float c; 298 float c;
307 for(i=0,c=1.0; i<M; i++,c-=1.0/M) { 299 for (i = 0, c = 1.0; i < M; i++, c -= 1.0 / M) {
308 Wo_[i] = Wo1*c + Wo2*(1.0-c); 300 Wo_[i] = Wo1 * c + Wo2 * (1.0 - c);
309 voicing_[i] = 1; 301 voicing_[i] = 1;
310 }
311 } 302 }
303 }
312 304
313 for(i=0; i<M; i++) { 305 for (i = 0; i < M; i++) {
314 L_[i] = floorf(M_PI/Wo_[i]); 306 L_[i] = floorf(M_PI / Wo_[i]);
315 } 307 }
316} 308}
317 309
318
319/*---------------------------------------------------------------------------*\ 310/*---------------------------------------------------------------------------*\
320 311
321 FUNCTION....: resample_rate_L 312 FUNCTION....: resample_rate_L
@@ -326,36 +317,37 @@ void interp_Wo_v(float Wo_[], int L_[], int voicing_[], float Wo1, float Wo2, in
326 317
327\*---------------------------------------------------------------------------*/ 318\*---------------------------------------------------------------------------*/
328 319
329void resample_rate_L(C2CONST *c2const, MODEL *model, float rate_K_vec[], float rate_K_sample_freqs_kHz[], int K) 320void resample_rate_L(C2CONST *c2const, MODEL *model, float rate_K_vec[],
330{ 321 float rate_K_sample_freqs_kHz[], int K) {
331 float rate_K_vec_term[K+2], rate_K_sample_freqs_kHz_term[K+2]; 322 float rate_K_vec_term[K + 2], rate_K_sample_freqs_kHz_term[K + 2];
332 float AmdB[MAX_AMP+1], rate_L_sample_freqs_kHz[MAX_AMP+1]; 323 float AmdB[MAX_AMP + 1], rate_L_sample_freqs_kHz[MAX_AMP + 1];
333 int m,k; 324 int m, k;
334
335 /* terminate either end of the rate K vecs with 0dB points */
336
337 rate_K_vec_term[0] = rate_K_vec_term[K+1] = 0.0;
338 rate_K_sample_freqs_kHz_term[0] = 0.0;
339 rate_K_sample_freqs_kHz_term[K+1] = 4.0;
340
341 for(k=0; k<K; k++) {
342 rate_K_vec_term[k+1] = rate_K_vec[k];
343 rate_K_sample_freqs_kHz_term[k+1] = rate_K_sample_freqs_kHz[k];
344
345 //printf("k: %d f: %f rate_K: %f\n", k, rate_K_sample_freqs_kHz[k], rate_K_vec[k]);
346 }
347
348 for(m=1; m<=model->L; m++) {
349 rate_L_sample_freqs_kHz[m] = m*model->Wo*(c2const->Fs/2000.0)/M_PI;
350 }
351
352 interp_para(&AmdB[1], rate_K_sample_freqs_kHz_term, rate_K_vec_term, K+2, &rate_L_sample_freqs_kHz[1], model->L);
353 for(m=1; m<=model->L; m++) {
354 model->A[m] = POW10F(AmdB[m]/20.0);
355 // printf("m: %d f: %f AdB: %f A: %f\n", m, rate_L_sample_freqs_kHz[m], AmdB[m], model->A[m]);
356 }
357}
358 325
326 /* terminate either end of the rate K vecs with 0dB points */
327
328 rate_K_vec_term[0] = rate_K_vec_term[K + 1] = 0.0;
329 rate_K_sample_freqs_kHz_term[0] = 0.0;
330 rate_K_sample_freqs_kHz_term[K + 1] = 4.0;
331
332 for (k = 0; k < K; k++) {
333 rate_K_vec_term[k + 1] = rate_K_vec[k];
334 rate_K_sample_freqs_kHz_term[k + 1] = rate_K_sample_freqs_kHz[k];
335 // printf("k: %d f: %f rate_K: %f\n", k, rate_K_sample_freqs_kHz[k],
336 // rate_K_vec[k]);
337 }
338
339 for (m = 1; m <= model->L; m++) {
340 rate_L_sample_freqs_kHz[m] = m * model->Wo * (c2const->Fs / 2000.0) / M_PI;
341 }
342
343 interp_para(&AmdB[1], rate_K_sample_freqs_kHz_term, rate_K_vec_term, K + 2,
344 &rate_L_sample_freqs_kHz[1], model->L);
345 for (m = 1; m <= model->L; m++) {
346 model->A[m] = POW10F(AmdB[m] / 20.0);
347 // printf("m: %d f: %f AdB: %f A: %f\n", m, rate_L_sample_freqs_kHz[m],
348 // AmdB[m], model->A[m]);
349 }
350}
359 351
360/*---------------------------------------------------------------------------*\ 352/*---------------------------------------------------------------------------*\
361 353
@@ -368,34 +360,100 @@ void resample_rate_L(C2CONST *c2const, MODEL *model, float rate_K_vec[], float r
368 360
369\*---------------------------------------------------------------------------*/ 361\*---------------------------------------------------------------------------*/
370 362
371void determine_phase(C2CONST *c2const, COMP H[], MODEL *model, int Nfft, codec2_fft_cfg fwd_cfg, codec2_fft_cfg inv_cfg) 363void determine_phase(C2CONST *c2const, COMP H[], MODEL *model, int Nfft,
372{ 364 codec2_fft_cfg fwd_cfg, codec2_fft_cfg inv_cfg) {
373 int i,m,b; 365 int i, m, b;
374 int Ns = Nfft/2+1; 366 int Ns = Nfft / 2 + 1;
375 float Gdbfk[Ns], sample_freqs_kHz[Ns], phase[Ns]; 367 float Gdbfk[Ns], sample_freqs_kHz[Ns], phase[Ns];
376 float AmdB[MAX_AMP+1], rate_L_sample_freqs_kHz[MAX_AMP+1]; 368 float AmdB[MAX_AMP + 1], rate_L_sample_freqs_kHz[MAX_AMP + 1];
377 369
378 for(m=1; m<=model->L; m++) { 370 for (m = 1; m <= model->L; m++) {
379 assert(model->A[m] != 0.0); 371 assert(model->A[m] != 0.0);
380 AmdB[m] = 20.0*log10f(model->A[m]); 372 AmdB[m] = 20.0 * log10f(model->A[m]);
381 rate_L_sample_freqs_kHz[m] = (float)m*model->Wo*(c2const->Fs/2000.0)/M_PI; 373 rate_L_sample_freqs_kHz[m] =
382 } 374 (float)m * model->Wo * (c2const->Fs / 2000.0) / M_PI;
383 375 }
384 for(i=0; i<Ns; i++) {
385 sample_freqs_kHz[i] = (c2const->Fs/1000.0)*(float)i/Nfft;
386 }
387 376
388 interp_para(Gdbfk, &rate_L_sample_freqs_kHz[1], &AmdB[1], model->L, sample_freqs_kHz, Ns); 377 for (i = 0; i < Ns; i++) {
389 mag_to_phase(phase, Gdbfk, Nfft, fwd_cfg, inv_cfg); 378 sample_freqs_kHz[i] = (c2const->Fs / 1000.0) * (float)i / Nfft;
379 }
390 380
391 for(m=1; m<=model->L; m++) { 381 interp_para(Gdbfk, &rate_L_sample_freqs_kHz[1], &AmdB[1], model->L,
392 b = floorf(0.5+m*model->Wo*Nfft/(2.0*M_PI)); 382 sample_freqs_kHz, Ns);
393 H[m].real = cosf(phase[b]); H[m].imag = sinf(phase[b]); 383 mag_to_phase(phase, Gdbfk, Nfft, fwd_cfg, inv_cfg);
394 } 384
385 for (m = 1; m <= model->L; m++) {
386 b = floorf(0.5 + m * model->Wo * Nfft / (2.0 * M_PI));
387 H[m].real = cosf(phase[b]);
388 H[m].imag = sinf(phase[b]);
389 }
395} 390}
396 391
392/*---------------------------------------------------------------------------* \
397 393
398/*---------------------------------------------------------------------------*\ 394 FUNCTION....: determine_autoc
395 AUTHOR......: David Rowe
396 DATE CREATED: April 2020
397
398 Determine autocorrelation coefficients from model params, for machine
399 learning experiments.
400
401\*---------------------------------------------------------------------------*/
402
403void determine_autoc(C2CONST *c2const, float Rk[], int order, MODEL *model,
404 int Nfft, codec2_fft_cfg fwd_cfg, codec2_fft_cfg inv_cfg) {
405 int i, m;
406 int Ns = Nfft / 2 + 1;
407 float Gdbfk[Ns], sample_freqs_kHz[Ns];
408 float AmdB[MAX_AMP + 1], rate_L_sample_freqs_kHz[MAX_AMP + 1];
409
410 /* interpolate in the log domain */
411 for (m = 1; m <= model->L; m++) {
412 assert(model->A[m] != 0.0);
413 AmdB[m] = 20.0 * log10f(model->A[m]);
414 rate_L_sample_freqs_kHz[m] =
415 (float)m * model->Wo * (c2const->Fs / 2000.0) / M_PI;
416 }
417
418 for (i = 0; i < Ns; i++) {
419 sample_freqs_kHz[i] = (c2const->Fs / 1000.0) * (float)i / Nfft;
420 }
421
422 interp_para(Gdbfk, &rate_L_sample_freqs_kHz[1], &AmdB[1], model->L,
423 sample_freqs_kHz, Ns);
424
425 COMP S[Nfft], R[Nfft];
426
427 /* install negative frequency components, convert to mag squared of spectrum
428 */
429 S[0].real = pow(10.0, Gdbfk[0] / 10.0);
430 S[0].imag = 0.0;
431 for (i = 1; i < Ns; i++) {
432 S[i].real = S[Nfft - i].real = pow(10.0, Gdbfk[i] / 10.0);
433 S[i].imag = S[Nfft - i].imag = 0.0;
434 }
435
436 /* IDFT of mag squared is autocorrelation function */
437 codec2_fft(inv_cfg, S, R);
438 for (int k = 0; k < order + 1; k++) Rk[k] = R[k].real;
439}
440
441/* update and optionally run "front eq" equaliser on before VQ */
442void newamp1_eq(float rate_K_vec_no_mean[], float eq[], int K, int eq_en) {
443 static float ideal[] = {8, 10, 12, 14, 14, 14, 14, 14, 14, 14,
444 14, 14, 14, 14, 14, 14, 14, 14, 14, -20};
445 float gain = 0.02;
446 float update;
447
448 for (int k = 0; k < K; k++) {
449 update = rate_K_vec_no_mean[k] - ideal[k];
450 eq[k] = (1.0 - gain) * eq[k] + gain * update;
451 if (eq[k] < 0.0) eq[k] = 0.0;
452 if (eq_en) rate_K_vec_no_mean[k] -= eq[k];
453 }
454}
455
456/*---------------------------------------------------------------------------* \
399 457
400 FUNCTION....: newamp1_model_to_indexes 458 FUNCTION....: newamp1_model_to_indexes
401 AUTHOR......: David Rowe 459 AUTHOR......: David Rowe
@@ -406,78 +464,53 @@ void determine_phase(C2CONST *c2const, COMP H[], MODEL *model, int Nfft, codec2_
406 464
407\*---------------------------------------------------------------------------*/ 465\*---------------------------------------------------------------------------*/
408 466
409void newamp1_model_to_indexes(C2CONST *c2const, 467void newamp1_model_to_indexes(C2CONST *c2const, int indexes[], MODEL *model,
410 int indexes[], 468 float rate_K_vec[],
411 MODEL *model, 469 float rate_K_sample_freqs_kHz[], int K,
412 float rate_K_vec[], 470 float *mean, float rate_K_vec_no_mean[],
413 float rate_K_sample_freqs_kHz[], 471 float rate_K_vec_no_mean_[], float *se, float *eq,
414 int K, 472 int eq_en) {
415 float *mean, 473 int k;
416 float rate_K_vec_no_mean[], 474
417 float rate_K_vec_no_mean_[], 475 /* convert variable rate L to fixed rate K */
418 float *se, 476 resample_const_rate_f(c2const, model, rate_K_vec, rate_K_sample_freqs_kHz, K);
419 float *eq, 477
420 int eq_en 478 /* remove mean */
421 ) 479 float sum = 0.0;
422{ 480 for (k = 0; k < K; k++) sum += rate_K_vec[k];
423 int k; 481 *mean = sum / K;
424 482 for (k = 0; k < K; k++) rate_K_vec_no_mean[k] = rate_K_vec[k] - *mean;
425 /* convert variable rate L to fixed rate K */ 483
426 resample_const_rate_f(c2const, model, rate_K_vec, rate_K_sample_freqs_kHz, K); 484 /* update and optionally run "front eq" equaliser on before VQ */
427 485 newamp1_eq(rate_K_vec_no_mean, eq, K, eq_en);
428 /* remove mean */ 486
429 float sum = 0.0; 487 /* two stage VQ */
430 for(k=0; k<K; k++) 488 rate_K_mbest_encode(indexes, rate_K_vec_no_mean, rate_K_vec_no_mean_, K,
431 sum += rate_K_vec[k]; 489 NEWAMP1_VQ_MBEST_DEPTH);
432 *mean = sum/K; 490
433 for(k=0; k<K; k++) 491 /* running sum of squared error for variance calculation */
434 rate_K_vec_no_mean[k] = rate_K_vec[k] - *mean; 492 for (k = 0; k < K; k++)
435 493 *se += (float)pow(rate_K_vec_no_mean[k] - rate_K_vec_no_mean_[k], 2.0);
436 /* update and optionally run "front eq" equaliser on before VQ */ 494
437 static float ideal[] = {8,10,12,14,14,14,14,14,14,14,14,14,14,14,14,14,14,14,14,-20}; 495 /* scalar quantise mean (effectively the frame energy) */
438 float gain = 0.02; 496 float w[1] = {1.0};
439 float update; 497 float se_mean;
440 498 indexes[2] =
441 for(k=0; k<K; k++) { 499 quantise(newamp1_energy_cb[0].cb, mean, w, newamp1_energy_cb[0].k,
442 update = rate_K_vec_no_mean[k] - ideal[k]; 500 newamp1_energy_cb[0].m, &se_mean);
443 eq[k] = (1.0-gain)*eq[k] + gain*update; 501
444 if (eq[k] < 0.0) eq[k] = 0.0; 502 /* scalar quantise Wo. We steal the smallest Wo index to signal
445 if (eq_en) 503 an unvoiced frame */
446 rate_K_vec_no_mean[k] -= eq[k]; 504 if (model->voiced) {
447 } 505 int index = encode_log_Wo(c2const, model->Wo, 6);
448 506 if (index == 0) {
449 /* two stage VQ */ 507 index = 1;
450 rate_K_mbest_encode(indexes, rate_K_vec_no_mean, rate_K_vec_no_mean_, K, NEWAMP1_VQ_MBEST_DEPTH);
451
452 /* running sum of squared error for variance calculation */
453 for(k=0; k<K; k++)
454 *se += pow(rate_K_vec_no_mean[k]-rate_K_vec_no_mean_[k],2.0);
455
456 /* scalar quantise mean (effectively the frame energy) */
457 float w[1] = {1.0};
458 float se_mean;
459 indexes[2] = quantise(newamp1_energy_cb[0].cb,
460 mean,
461 w,
462 newamp1_energy_cb[0].k,
463 newamp1_energy_cb[0].m,
464 &se_mean);
465
466 /* scalar quantise Wo. We steal the smallest Wo index to signal
467 an unvoiced frame */
468 if (model->voiced) {
469 int index = encode_log_Wo(c2const, model->Wo, 6);
470 if (index == 0) {
471 index = 1;
472 }
473 indexes[3] = index;
474 }
475 else {
476 indexes[3] = 0;
477 } 508 }
478 509 indexes[3] = index;
479 } 510 } else {
480 511 indexes[3] = 0;
512 }
513}
481 514
482/*---------------------------------------------------------------------------*\ 515/*---------------------------------------------------------------------------*\
483 516
@@ -487,22 +520,22 @@ void newamp1_model_to_indexes(C2CONST *c2const,
487 520
488\*---------------------------------------------------------------------------*/ 521\*---------------------------------------------------------------------------*/
489 522
490void newamp1_interpolate(float interpolated_surface_[], float left_vec[], float right_vec[], int K) 523void newamp1_interpolate(float interpolated_surface_[], float left_vec[],
491{ 524 float right_vec[], int K) {
492 int i, k; 525 int i, k;
493 int M = 4; 526 int M = 4;
494 float c; 527 float c;
495 528
496 /* (linearly) interpolate 25Hz amplitude vectors back to 100Hz */ 529 /* (linearly) interpolate 25Hz amplitude vectors back to 100Hz */
497 530
498 for(i=0,c=1.0; i<M; i++,c-=1.0/M) { 531 for (i = 0, c = 1.0; i < M; i++, c -= 1.0 / M) {
499 for(k=0; k<K; k++) { 532 for (k = 0; k < K; k++) {
500 interpolated_surface_[i*K+k] = left_vec[k]*c + right_vec[k]*(1.0-c); 533 interpolated_surface_[i * K + k] =
501 } 534 left_vec[k] * c + right_vec[k] * (1.0 - c);
502 } 535 }
536 }
503} 537}
504 538
505
506/*---------------------------------------------------------------------------*\ 539/*---------------------------------------------------------------------------*\
507 540
508 FUNCTION....: newamp1_indexes_to_rate_K_vec 541 FUNCTION....: newamp1_indexes_to_rate_K_vec
@@ -514,42 +547,39 @@ void newamp1_interpolate(float interpolated_surface_[], float left_vec[], float
514 547
515\*---------------------------------------------------------------------------*/ 548\*---------------------------------------------------------------------------*/
516 549
517void newamp1_indexes_to_rate_K_vec(float rate_K_vec_[], 550void newamp1_indexes_to_rate_K_vec(float rate_K_vec_[],
518 float rate_K_vec_no_mean_[], 551 float rate_K_vec_no_mean_[],
519 float rate_K_sample_freqs_kHz[], 552 float rate_K_sample_freqs_kHz[], int K,
520 int K, 553 float *mean_, int indexes[],
521 float *mean_,
522 int indexes[],
523 float user_rate_K_vec_no_mean_[], 554 float user_rate_K_vec_no_mean_[],
524 int post_filter_en) 555 int post_filter_en) {
525{ 556 int k;
526 int k; 557 const float *codebook1 = newamp1vq_cb[0].cb;
527 const float *codebook1 = newamp1vq_cb[0].cb; 558 const float *codebook2 = newamp1vq_cb[1].cb;
528 const float *codebook2 = newamp1vq_cb[1].cb; 559 int n1 = indexes[0];
529 int n1 = indexes[0]; 560 int n2 = indexes[1];
530 int n2 = indexes[1]; 561
531 562 if (user_rate_K_vec_no_mean_ == NULL) {
532 if (user_rate_K_vec_no_mean_ == NULL) { 563 /* normal operation */
533 /* normal operation */ 564 for (k = 0; k < K; k++) {
534 for(k=0; k<K; k++) { 565 rate_K_vec_no_mean_[k] = codebook1[K * n1 + k] + codebook2[K * n2 + k];
535 rate_K_vec_no_mean_[k] = codebook1[K*n1+k] + codebook2[K*n2+k];
536 }
537 } else {
538 /* for development we can optionally inject the quantised rate K vector here */
539 for(k=0; k<K; k++)
540 rate_K_vec_no_mean_[k] = user_rate_K_vec_no_mean_[k];
541 } 566 }
542 567 } else {
543 if (post_filter_en) 568 /* for development we can optionally inject the quantised rate K vector here
544 post_filter_newamp1(rate_K_vec_no_mean_, rate_K_sample_freqs_kHz, K, 1.5); 569 */
570 for (k = 0; k < K; k++)
571 rate_K_vec_no_mean_[k] = user_rate_K_vec_no_mean_[k];
572 }
545 573
546 *mean_ = newamp1_energy_cb[0].cb[indexes[2]]; 574 if (post_filter_en)
575 post_filter_newamp1(rate_K_vec_no_mean_, rate_K_sample_freqs_kHz, K, 1.5);
547 576
548 for(k=0; k<K; k++) { 577 *mean_ = newamp1_energy_cb[0].cb[indexes[2]];
549 rate_K_vec_[k] = rate_K_vec_no_mean_[k] + *mean_;
550 }
551}
552 578
579 for (k = 0; k < K; k++) {
580 rate_K_vec_[k] = rate_K_vec_no_mean_[k] + *mean_;
581 }
582}
553 583
554/*---------------------------------------------------------------------------*\ 584/*---------------------------------------------------------------------------*\
555 585
@@ -561,78 +591,66 @@ void newamp1_indexes_to_rate_K_vec(float rate_K_vec_[],
561 591
562\*---------------------------------------------------------------------------*/ 592\*---------------------------------------------------------------------------*/
563 593
564void newamp1_indexes_to_model(C2CONST *c2const, 594void newamp1_indexes_to_model(C2CONST *c2const, MODEL model_[], COMP H[],
565 MODEL model_[],
566 COMP H[],
567 float *interpolated_surface_, 595 float *interpolated_surface_,
568 float prev_rate_K_vec_[], 596 float prev_rate_K_vec_[], float *Wo_left,
569 float *Wo_left, 597 int *voicing_left,
570 int *voicing_left, 598 float rate_K_sample_freqs_kHz[], int K,
571 float rate_K_sample_freqs_kHz[], 599 codec2_fft_cfg fwd_cfg, codec2_fft_cfg inv_cfg,
572 int K, 600 int indexes[], float user_rate_K_vec_no_mean_[],
573 codec2_fft_cfg fwd_cfg, 601 int post_filter_en) {
574 codec2_fft_cfg inv_cfg, 602 float rate_K_vec_[K], rate_K_vec_no_mean_[K], mean_, Wo_right;
575 int indexes[], 603 int voicing_right, k;
576 float user_rate_K_vec_no_mean_[], 604 int M = 4;
577 int post_filter_en) 605
578{ 606 /* extract latest rate K vector */
579 float rate_K_vec_[K], rate_K_vec_no_mean_[K], mean_, Wo_right; 607
580 int voicing_right, k; 608 newamp1_indexes_to_rate_K_vec(rate_K_vec_, rate_K_vec_no_mean_,
581 int M = 4; 609 rate_K_sample_freqs_kHz, K, &mean_, indexes,
582 610 user_rate_K_vec_no_mean_, post_filter_en);
583 /* extract latest rate K vector */ 611
584 612 /* decode latest Wo and voicing */
585 newamp1_indexes_to_rate_K_vec(rate_K_vec_, 613
586 rate_K_vec_no_mean_, 614 if (indexes[3]) {
587 rate_K_sample_freqs_kHz, 615 Wo_right = decode_log_Wo(c2const, indexes[3], 6);
588 K, 616 voicing_right = 1;
589 &mean_, 617 } else {
590 indexes, 618 Wo_right = 2.0 * M_PI / 100.0;
591 user_rate_K_vec_no_mean_, 619 voicing_right = 0;
592 post_filter_en); 620 }
593
594 /* decode latest Wo and voicing */
595
596 if (indexes[3]) {
597 Wo_right = decode_log_Wo(c2const, indexes[3], 6);
598 voicing_right = 1;
599 }
600 else {
601 Wo_right = 2.0*M_PI/100.0;
602 voicing_right = 0;
603 }
604
605 /* interpolate 25Hz rate K vec back to 100Hz */
606 621
607 float *left_vec = prev_rate_K_vec_; 622 /* interpolate 25Hz rate K vec back to 100Hz */
608 float *right_vec = rate_K_vec_;
609 newamp1_interpolate(interpolated_surface_, left_vec, right_vec, K);
610 623
611 /* interpolate 25Hz v and Wo back to 100Hz */ 624 float *left_vec = prev_rate_K_vec_;
625 float *right_vec = rate_K_vec_;
626 newamp1_interpolate(interpolated_surface_, left_vec, right_vec, K);
612 627
613 float aWo_[M]; 628 /* interpolate 25Hz v and Wo back to 100Hz */
614 int avoicing_[M], aL_[M], i;
615 629
616 interp_Wo_v(aWo_, aL_, avoicing_, *Wo_left, Wo_right, *voicing_left, voicing_right); 630 float aWo_[M];
631 int avoicing_[M], aL_[M], i;
617 632
618 /* back to rate L amplitudes, synthesis phase for each frame */ 633 interp_Wo_v(aWo_, aL_, avoicing_, *Wo_left, Wo_right, *voicing_left,
634 voicing_right);
619 635
620 for(i=0; i<M; i++) { 636 /* back to rate L amplitudes, synthesise phase for each frame */
621 model_[i].Wo = aWo_[i];
622 model_[i].L = aL_[i];
623 model_[i].voiced = avoicing_[i];
624 637
625 resample_rate_L(c2const, &model_[i], &interpolated_surface_[K*i], rate_K_sample_freqs_kHz, K); 638 for (i = 0; i < M; i++) {
626 determine_phase(c2const, &H[(MAX_AMP+1)*i], &model_[i], NEWAMP1_PHASE_NFFT, fwd_cfg, inv_cfg); 639 model_[i].Wo = aWo_[i];
627 } 640 model_[i].L = aL_[i];
641 model_[i].voiced = avoicing_[i];
628 642
629 /* update memories for next time */ 643 resample_rate_L(c2const, &model_[i], &interpolated_surface_[K * i],
644 rate_K_sample_freqs_kHz, K);
645 determine_phase(c2const, &H[(MAX_AMP + 1) * i], &model_[i],
646 NEWAMP1_PHASE_NFFT, fwd_cfg, inv_cfg);
647 }
630 648
631 for(k=0; k<K; k++) { 649 /* update memories for next time */
632 prev_rate_K_vec_[k] = rate_K_vec_[k];
633 }
634 *Wo_left = Wo_right;
635 *voicing_left = voicing_right;
636 650
651 for (k = 0; k < K; k++) {
652 prev_rate_K_vec_[k] = rate_K_vec_[k];
653 }
654 *Wo_left = Wo_right;
655 *voicing_left = voicing_right;
637} 656}
638
diff --git a/newamp1.h b/newamp1.h
index 8c18813..476f51b 100644
--- a/newamp1.h
+++ b/newamp1.h
@@ -30,57 +30,57 @@
30#ifndef __NEWAMP1__ 30#ifndef __NEWAMP1__
31#define __NEWAMP1__ 31#define __NEWAMP1__
32 32
33#define NEWAMP1_N_INDEXES 4 /* Number of indexes to pack: vq1, vq2, energy, Wo */ 33#define NEWAMP1_N_INDEXES \
34#define NEWAMP1_PHASE_NFFT 128 /* size of FFT used for phase synthesis */ 34 4 /* Number of indexes to pack: vq1, vq2, energy, Wo */
35#define NEWAMP1_K 20 /* rate K vector length */ 35#define NEWAMP1_PHASE_NFFT \
36 128 /* size of FFT used for phase synthesis */
37#define NEWAMP1_K 20 /* rate K vector length */
38#define NEWAMP1_VQ_MBEST_DEPTH \
39 5 /* how many candidates we keep for each stage of mbest search */
36 40
37#include "codec2_fft.h" 41#include "codec2_fft.h"
38#include "comp.h" 42#include "comp.h"
39 43
40void interp_para(float y[], float xp[], float yp[], int np, float x[], int n); 44void interp_para(float y[], float xp[], float yp[], int np, float x[], int n);
41float ftomel(float fHz); 45float ftomel(float fHz);
42void mel_sample_freqs_kHz(float rate_K_sample_freqs_kHz[], int K, float mel_start, float mel_end); 46void mel_sample_freqs_kHz(float rate_K_sample_freqs_kHz[], int K,
43void resample_const_rate_f(C2CONST *c2const, MODEL *model, float rate_K_vec[], float rate_K_sample_freqs_kHz[], int K); 47 float mel_start, float mel_end);
44float rate_K_mbest_encode(int *indexes, float *x, float *xq, int ndim, int mbest_entries); 48void resample_const_rate_f(C2CONST *c2const, MODEL *model, float rate_K_vec[],
45void post_filter_newamp1(float vec[], float sample_freq_kHz[], int K, float pf_gain); 49 float rate_K_sample_freqs_kHz[], int K);
46void interp_Wo_v(float Wo_[], int L_[], int voicing_[], float Wo1, float Wo2, int voicing1, int voicing2); 50float rate_K_mbest_encode(int *indexes, float *x, float *xq, int ndim,
47void resample_rate_L(C2CONST *c2const, MODEL *model, float rate_K_vec[], float rate_K_sample_freqs_kHz[], int K); 51 int mbest_entries);
48void determine_phase(C2CONST *c2const, COMP H[], MODEL *model, int Nfft, codec2_fft_cfg fwd_cfg, codec2_fft_cfg inv_cfg); 52void post_filter_newamp1(float vec[], float sample_freq_kHz[], int K,
49void newamp1_model_to_indexes(C2CONST *c2const, 53 float pf_gain);
50 int indexes[], 54void interp_Wo_v(float Wo_[], int L_[], int voicing_[], float Wo1, float Wo2,
51 MODEL *model, 55 int voicing1, int voicing2);
52 float rate_K_vec[], 56void resample_rate_L(C2CONST *c2const, MODEL *model, float rate_K_vec[],
53 float rate_K_sample_freqs_kHz[], 57 float rate_K_sample_freqs_kHz[], int K);
54 int K, 58void determine_phase(C2CONST *c2const, COMP H[], MODEL *model, int Nfft,
55 float *mean, 59 codec2_fft_cfg fwd_cfg, codec2_fft_cfg inv_cfg);
56 float rate_K_vec_no_mean[], 60void determine_autoc(C2CONST *c2const, float Rk[], int order, MODEL *model,
57 float rate_K_vec_no_mean_[], 61 int Nfft, codec2_fft_cfg fwd_cfg, codec2_fft_cfg inv_cfg);
58 float *se, 62void newamp1_model_to_indexes(C2CONST *c2const, int indexes[], MODEL *model,
59 float *eq, 63 float rate_K_vec[],
60 int eq_en); 64 float rate_K_sample_freqs_kHz[], int K,
61void newamp1_indexes_to_rate_K_vec(float rate_K_vec_[], 65 float *mean, float rate_K_vec_no_mean[],
62 float rate_K_vec_no_mean_[], 66 float rate_K_vec_no_mean_[], float *se, float *eq,
63 float rate_K_sample_freqs_kHz[], 67 int eq_en);
64 int K, 68void newamp1_indexes_to_rate_K_vec(float rate_K_vec_[],
65 float *mean_, 69 float rate_K_vec_no_mean_[],
66 int indexes[], 70 float rate_K_sample_freqs_kHz[], int K,
71 float *mean_, int indexes[],
67 float user_rate_K_vec_no_mean_[], 72 float user_rate_K_vec_no_mean_[],
68 int post_filter_en); 73 int post_filter_en);
69void newamp1_interpolate(float interpolated_surface_[], float left_vec[], float right_vec[], int K); 74void newamp1_interpolate(float interpolated_surface_[], float left_vec[],
70 75 float right_vec[], int K);
71void newamp1_indexes_to_model(C2CONST *c2const, 76void newamp1_eq(float rate_K_vec_no_mean[], float eq[], int K, int eq_en);
72 MODEL model_[], 77void newamp1_indexes_to_model(C2CONST *c2const, MODEL model_[], COMP H[],
73 COMP H[], 78 float interpolated_surface_[],
74 float interpolated_surface_[], 79 float prev_rate_K_vec_[], float *Wo_left,
75 float prev_rate_K_vec_[], 80 int *voicing_left,
76 float *Wo_left, 81 float rate_K_sample_freqs_kHz[], int K,
77 int *voicing_left, 82 codec2_fft_cfg fwd_cfg, codec2_fft_cfg inv_cfg,
78 float rate_K_sample_freqs_kHz[], 83 int indexes[], float user_rate_K_vec_no_mean_[],
79 int K,
80 codec2_fft_cfg fwd_cfg,
81 codec2_fft_cfg inv_cfg,
82 int indexes[],
83 float user_rate_K_vec_no_mean_[],
84 int post_filter_en); 84 int post_filter_en);
85 85
86#endif 86#endif
diff --git a/newamp2.c b/newamp2.c
deleted file mode 100644
index 6550557..0000000
--- a/newamp2.c
+++ /dev/null
@@ -1,570 +0,0 @@
1/*---------------------------------------------------------------------------*\
2
3 FILE........: newamp2.c
4 AUTHOR......: Thomas Kurin and Stefan Erhardt
5 INSTITUTE...: Institute for Electronics Engineering, University of Erlangen-Nuremberg
6 DATE CREATED: July 2018
7 BASED ON....: "newamp1" by David Rowe
8
9 Quantisation functions for the sinusoidal coder, using "newamp1"
10 algorithm that resamples variable rate L [Am} to a fixed rate K then
11 VQs.
12
13\*---------------------------------------------------------------------------*/
14
15/*
16 Copyright David Rowe 2017
17
18 All rights reserved.
19
20 This program is free software; you can redistribute it and/or modify
21 it under the terms of the GNU Lesser General Public License version 2.1, as
22 published by the Free Software Foundation. This program is
23 distributed in the hope that it will be useful, but WITHOUT ANY
24 WARRANTY; without even the implied warranty of MERCHANTABILITY or
25 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public
26 License for more details.
27
28 You should have received a copy of the GNU Lesser General Public License
29 along with this program; if not, see <http://www.gnu.org/licenses/>.
30
31*/
32
33#include <assert.h>
34#include <stdio.h>
35#include <stdlib.h>
36#include <string.h>
37#include <math.h>
38
39#include "defines.h"
40#include "phase.h"
41#include "quantise.h"
42#include "mbest.h"
43#include "newamp1.h"
44#include "newamp2.h"
45
46/*---------------------------------------------------------------------------*\
47
48 FUNCTION....: n2_mel_sample_freqs_kHz()
49 AUTHOR......: Thomas Kurin and Stefan Erhardt
50 INSTITUTE...: Institute for Electronics Engineering, University of Erlangen-Nuremberg
51 DATE CREATED: July 2018
52
53 Outputs fixed frequencies for the K-Vectors to be able to work with both 8k and 16k mode.
54
55\*---------------------------------------------------------------------------*/
56
57void n2_mel_sample_freqs_kHz(float rate_K_sample_freqs_kHz[], int K)
58{
59 float freq[] = {0.199816, 0.252849, 0.309008, 0.368476, 0.431449, 0.498134, 0.568749, 0.643526, 0.722710, 0.806561, 0.895354, 0.989380,
60 1.088948, 1.194384, 1.306034, 1.424264, 1.549463, 1.682041, 1.822432, 1.971098, 2.128525, 2.295232, 2.471763, 2.658699,
61 2.856652, 3.066272, 3.288246, 3.523303, 3.772214, 4.035795, 4.314912, 4.610478, 4.923465, 5.254899, 5.605865, 5.977518,
62 6.371075, 6.787827, 7.229141, 7.696465};
63 int k;
64 //printf("\n\n");
65 for (k=0; k<K; k++) {
66 rate_K_sample_freqs_kHz[k] = freq[k];
67 // printf("%f ",mel);
68 // printf("%f \n",rate_K_sample_freqs_kHz[k]);
69 }
70
71}
72
73
74/*---------------------------------------------------------------------------*\
75
76 FUNCTION....: n2_resample_const_rate_f() still equal to resample_const_rate_f()
77 AUTHOR......: David Rowe
78 DATE CREATED: Jan 2017
79
80 Resample Am from time-varying rate L=floor(pi/Wo) to fixed rate K.
81
82\*---------------------------------------------------------------------------*/
83
84void n2_resample_const_rate_f(C2CONST *c2const, MODEL *model, float rate_K_vec[], float rate_K_sample_freqs_kHz[], int K)
85{
86 int m;
87 float AmdB[MAX_AMP+1], rate_L_sample_freqs_kHz[MAX_AMP+1], AmdB_peak;
88
89 /* convert rate L=pi/Wo amplitude samples to fixed rate K */
90
91 AmdB_peak = -100.0;
92 for(m=1; m<=model->L; m++) {
93 AmdB[m] = 20.0*log10(model->A[m]+1E-16);
94 if (AmdB[m] > AmdB_peak) {
95 AmdB_peak = AmdB[m];
96 }
97 rate_L_sample_freqs_kHz[m] = m*model->Wo*(c2const->Fs/2000.0)/M_PI;
98 //printf("m: %d AmdB: %f AmdB_peak: %f sf: %f\n", m, AmdB[m], AmdB_peak, rate_L_sample_freqs_kHz[m]);
99 }
100
101 /* clip between peak and peak -50dB, to reduce dynamic range */
102
103 for(m=1; m<=model->L; m++) {
104 if (AmdB[m] < (AmdB_peak-50.0)) {
105 AmdB[m] = AmdB_peak-50.0;
106 }
107 }
108
109 interp_para(rate_K_vec, &rate_L_sample_freqs_kHz[1], &AmdB[1], model->L, rate_K_sample_freqs_kHz, K);
110}
111
112
113/*---------------------------------------------------------------------------*\
114
115 FUNCTION....: n2_rate_K_mbest_encode
116 AUTHOR......: Thomas Kurin and Stefan Erhardt
117 INSTITUTE...: Institute for Electronics Engineering, University of Erlangen-Nuremberg
118 DATE CREATED: July 2018
119
120 One stage rate K newamp2 VQ quantiser using mbest search.
121
122\*---------------------------------------------------------------------------*/
123
124void n2_rate_K_mbest_encode(int *indexes, float *x, float *xq, int ndim)
125{
126 int i, n1;
127 const float *codebook1 = newamp2vq_cb[0].cb;
128 struct MBEST *mbest_stage1;
129 float w[ndim];
130 int index[1];
131
132 /* codebook is compiled for a fixed K */
133
134 //assert(ndim == newamp2vq_cb[0].k);
135
136 /* equal weights, could be argued mel freq axis gives freq dep weighting */
137
138 for(i=0; i<ndim; i++)
139 w[i] = 1.0;
140
141 mbest_stage1 = mbest_create(1);
142
143 index[0] = 0;
144
145 /* Stage 1 */
146
147 mbest_search450(codebook1, x, w, ndim,NEWAMP2_K, newamp2vq_cb[0].m, mbest_stage1, index);
148 MBEST_PRINT("Stage 1:", mbest_stage1);
149 n1 = mbest_stage1->list[0].index[0];
150
151 mbest_destroy(mbest_stage1);
152
153 //indexes[1]: legacy from newamp1
154 indexes[0] = n1; indexes[1] = n1;
155
156}
157
158
159/*---------------------------------------------------------------------------*\
160
161 FUNCTION....: n2_resample_rate_L
162 AUTHOR......: Thomas Kurin and Stefan Erhardt
163 INSTITUTE...: Institute for Electronics Engineering, University of Erlangen-Nuremberg
164 DATE CREATED: July 2018
165
166 Decoder side conversion of rate K vector back to rate L.
167 Plosives are set to zero for the first 2 of 4 frames.
168
169\*---------------------------------------------------------------------------*/
170
171void n2_resample_rate_L(C2CONST *c2const, MODEL *model, float rate_K_vec[], float rate_K_sample_freqs_kHz[], int K,int plosive_flag)
172{
173 float rate_K_vec_term[K+2], rate_K_sample_freqs_kHz_term[K+2];
174 float AmdB[MAX_AMP+1], rate_L_sample_freqs_kHz[MAX_AMP+1];
175 int m,k;
176
177 /* terminate either end of the rate K vecs with 0dB points */
178
179 rate_K_vec_term[0] = rate_K_vec_term[K+1] = 0.0;
180 rate_K_sample_freqs_kHz_term[0] = 0.0;
181 rate_K_sample_freqs_kHz_term[K+1] = 4.0;
182
183 for(k=0; k<K; k++) {
184 rate_K_vec_term[k+1] = rate_K_vec[k];
185 rate_K_sample_freqs_kHz_term[k+1] = rate_K_sample_freqs_kHz[k];
186
187 //printf("k: %d f: %f rate_K: %f\n", k, rate_K_sample_freqs_kHz[k], rate_K_vec[k]);
188 }
189
190 for(m=1; m<=model->L; m++) {
191 rate_L_sample_freqs_kHz[m] = m*model->Wo*(c2const->Fs/2000.0)/M_PI;
192 }
193
194 interp_para(&AmdB[1], rate_K_sample_freqs_kHz_term, rate_K_vec_term, K+2, &rate_L_sample_freqs_kHz[1], model->L);
195 for(m=1; m<=model->L; m++) {
196 if(plosive_flag==0){
197 model->A[m] = pow(10.0, AmdB[m]/20.0);
198 }else{
199 model->A[m] = 0.1;
200 }
201 // printf("m: %d f: %f AdB: %f A: %f\n", m, rate_L_sample_freqs_kHz[m], AmdB[m], model->A[m]);
202 }
203}
204
205/*---------------------------------------------------------------------------*\
206
207 FUNCTION....: n2_post_filter_newamp2
208 AUTHOR......: Thomas Kurin and Stefan Erhardt
209 INSTITUTE...: Institute for Electronics Engineering, University of Erlangen-Nuremberg
210 DATE CREATED: July 2018
211
212 Postfilter for the pseudo wideband mode. Still has to be adapted!
213
214\*---------------------------------------------------------------------------*/
215
216void n2_post_filter_newamp2(float vec[], float sample_freq_kHz[], int K, float pf_gain)
217{
218 int k;
219
220 /*
221 vec is rate K vector describing spectrum of current frame lets
222 pre-emp before applying PF. 20dB/dec over 300Hz. Postfilter
223 affects energy of frame so we measure energy before and after
224 and normalise. Plenty of room for experiment here as well.
225 */
226
227 float pre[K];
228 float e_before = 0.0;
229 float e_after = 0.0;
230 for(k=0; k<K; k++) {
231 pre[k] = 20.0*log10f(sample_freq_kHz[k]/0.3);
232 vec[k] += pre[k];
233 e_before += POW10F(vec[k]/10.0);
234 vec[k] *= pf_gain;
235 e_after += POW10F(vec[k]/10.0);
236 }
237
238 float gain = e_after/e_before;
239 float gaindB = 10*log10f(gain);
240
241 for(k=0; k<K; k++) {
242 vec[k] -= gaindB;
243 vec[k] -= pre[k];
244 }
245}
246
247
248/*---------------------------------------------------------------------------*\
249
250 FUNCTION....: newamp2_model_to_indexes
251 AUTHOR......: Thomas Kurin and Stefan Erhardt
252 INSTITUTE...: Institute for Electronics Engineering, University of Erlangen-Nuremberg
253 DATE CREATED: July 2018
254
255 newamp2 encoder: Encodes the 8k sampled samples using mbest search (one stage)
256
257\*---------------------------------------------------------------------------*/
258
259void newamp2_model_to_indexes(C2CONST *c2const,
260 int indexes[],
261 MODEL *model,
262 float rate_K_vec[],
263 float rate_K_sample_freqs_kHz[],
264 int K,
265 float *mean,
266 float rate_K_vec_no_mean[],
267 float rate_K_vec_no_mean_[],
268 int plosive
269 )
270{
271 int k;
272
273 /* convert variable rate L to fixed rate K */
274
275 resample_const_rate_f(c2const, model, rate_K_vec, rate_K_sample_freqs_kHz, K);
276
277 /* remove mean and two stage VQ */
278
279 float sum = 0.0;
280 for(k=0; k<K; k++)
281 sum += rate_K_vec[k];
282 *mean = sum/K;
283 for(k=0; k<K; k++)
284 {
285 rate_K_vec_no_mean[k] = rate_K_vec[k] - *mean;
286 }
287 //NEWAMP2_16K_K+1 because the last vector is not a vector for VQ (and not included in the constant)
288 //but a calculated medium mean value
289 n2_rate_K_mbest_encode(indexes, rate_K_vec_no_mean, rate_K_vec_no_mean_, NEWAMP2_16K_K+1);
290
291 /* scalar quantise mean (effectively the frame energy) */
292
293 float w[1] = {1.0};
294 float se;
295 indexes[2] = quantise(newamp2_energy_cb[0].cb,
296 mean,
297 w,
298 newamp2_energy_cb[0].k,
299 newamp2_energy_cb[0].m,
300 &se);
301
302 /* scalar quantise Wo. We steal the smallest Wo index to signal
303 an unvoiced frame */
304
305 if (model->voiced) {
306 int index = encode_log_Wo(c2const, model->Wo, 6);
307 if (index == 0) {
308 index = 1;
309 }
310 if (index == 63) {
311 index = 62;
312 }
313 indexes[3] = index;
314 }
315 else {
316 indexes[3] = 0;
317 }
318 if(plosive != 0){
319 indexes[3] = 63;
320 }
321 }
322
323
324/*---------------------------------------------------------------------------*\
325
326 FUNCTION....: newamp2_indexes_to_rate_K_vec
327 AUTHOR......: Thomas Kurin and Stefan Erhardt
328 INSTITUTE...: Institute for Electronics Engineering, University of Erlangen-Nuremberg
329 DATE CREATED: July 2018
330
331 newamp2 decoder for amplitudes {Am}. Given the rate K VQ and energy
332 indexes, outputs rate K vector. Equal to newamp1 but using only one stage VQ.
333
334\*---------------------------------------------------------------------------*/
335
336void newamp2_indexes_to_rate_K_vec(float rate_K_vec_[],
337 float rate_K_vec_no_mean_[],
338 float rate_K_sample_freqs_kHz[],
339 int K,
340 float *mean_,
341 int indexes[],
342 float pf_gain)
343{
344 int k;
345 const float *codebook1 = newamp2vq_cb[0].cb;
346 int n1 = indexes[0];
347
348 for(k=0; k<K; k++) {
349 rate_K_vec_no_mean_[k] = codebook1[(NEWAMP2_16K_K+1)*n1+k];
350 }
351
352 post_filter_newamp1(rate_K_vec_no_mean_, rate_K_sample_freqs_kHz, K, pf_gain);
353
354 *mean_ = newamp2_energy_cb[0].cb[indexes[2]];
355
356 for(k=0; k<K; k++) {
357 rate_K_vec_[k] = rate_K_vec_no_mean_[k] + *mean_;
358 }
359}
360
361/*---------------------------------------------------------------------------*\
362
363 FUNCTION....: newamp2_16k_indexes_to_rate_K_vec
364 AUTHOR......: Thomas Kurin and Stefan Erhardt
365 INSTITUTE...: Institute for Electronics Engineering, University of Erlangen-Nuremberg
366 DATE CREATED: July 2018
367
368 newamp2 decoder for amplitudes {Am}. Given the rate K VQ and energy
369 indexes, outputs rate K vector. Extends the sample rate by looking up the corresponding
370 higher frequency values with their energy difference to the base energy (=>mean2)
371
372\*---------------------------------------------------------------------------*/
373
374void newamp2_16k_indexes_to_rate_K_vec(float rate_K_vec_[],
375 float rate_K_vec_no_mean_[],
376 float rate_K_sample_freqs_kHz[],
377 int K,
378 float *mean_,
379 int indexes[],
380 float pf_gain)
381{
382 int k;
383 const float *codebook1 = newamp2vq_cb[0].cb;
384 float mean2 = 0;
385 int n1 = indexes[0];
386
387 for(k=0; k<K; k++) {
388 rate_K_vec_no_mean_[k] = codebook1[(K+1)*n1+k];
389 }
390
391 n2_post_filter_newamp2(rate_K_vec_no_mean_, rate_K_sample_freqs_kHz, K, pf_gain);
392
393 *mean_ = newamp2_energy_cb[0].cb[indexes[2]];
394 mean2 = *mean_ + codebook1[(K+1)*n1+K] -10;
395
396 //HF ear Protection
397 if(mean2>50){
398 mean2 = 50;
399 }
400
401 for(k=0; k<K; k++) {
402 if(k<NEWAMP2_K){
403 rate_K_vec_[k] = rate_K_vec_no_mean_[k] + *mean_;
404 }
405 else{
406 //Amplify or Reduce ??
407 rate_K_vec_[k] = rate_K_vec_no_mean_[k] + mean2;
408 }
409 }
410}
411/*---------------------------------------------------------------------------*\
412
413 FUNCTION....: newamp2_interpolate
414 AUTHOR......: Thomas Kurin and Stefan Erhardt
415 INSTITUTE...: Institute for Electronics Engineering, University of Erlangen-Nuremberg
416 DATE CREATED: July 2018
417
418 Interpolates to the 4 10ms Frames and leaves the forst 2 empty for plosives
419
420\*---------------------------------------------------------------------------*/
421
422void newamp2_interpolate(float interpolated_surface_[], float left_vec[], float right_vec[], int K, int plosive_flag)
423{
424 int i, k;
425 int M = 4;
426 float c;
427
428 /* (linearly) interpolate 25Hz amplitude vectors back to 100Hz */
429
430 if(plosive_flag == 0){
431 for(i=0,c=1.0; i<M; i++,c-=1.0/M) {
432 for(k=0; k<K; k++) {
433 interpolated_surface_[i*K+k] = left_vec[k]*c + right_vec[k]*(1.0-c);
434 }
435 }
436 }
437 else{
438 for(i=0,c=1.0; i<M; i++,c-=1.0/M) {
439 for(k=0; k<K; k++) {
440 if(i<2){
441 interpolated_surface_[i*K+k] = 0;
442 }
443 else{
444 //perhaps add some dB ?
445 interpolated_surface_[i*K+k] = right_vec[k];
446 }
447 }
448 }
449
450 }
451}
452
453
454/*---------------------------------------------------------------------------*\
455
456 FUNCTION....: newamp2_indexes_to_model
457 AUTHOR......: Thomas Kurin and Stefan Erhardt
458 INSTITUTE...: Institute for Electronics Engineering, University of Erlangen-Nuremberg
459 DATE CREATED: July 2018
460
461 newamp2 decoder. Chooses whether to decode to 16k mode or to 8k mode
462
463\*---------------------------------------------------------------------------*/
464
465void newamp2_indexes_to_model(C2CONST *c2const,
466 MODEL model_[],
467 COMP H[],
468 float *interpolated_surface_,
469 float prev_rate_K_vec_[],
470 float *Wo_left,
471 int *voicing_left,
472 float rate_K_sample_freqs_kHz[],
473 int K,
474 codec2_fft_cfg fwd_cfg,
475 codec2_fft_cfg inv_cfg,
476 int indexes[],
477 float pf_gain,
478 int flag16k)
479{
480 float rate_K_vec_[K], rate_K_vec_no_mean_[K], mean_, Wo_right;
481 int voicing_right, k;
482 int M = 4;
483
484 /* extract latest rate K vector */
485
486 if(flag16k == 0){
487 newamp2_indexes_to_rate_K_vec(rate_K_vec_,
488 rate_K_vec_no_mean_,
489 rate_K_sample_freqs_kHz,
490 K,
491 &mean_,
492 indexes,
493 pf_gain);
494 }else{
495 newamp2_16k_indexes_to_rate_K_vec(rate_K_vec_,
496 rate_K_vec_no_mean_,
497 rate_K_sample_freqs_kHz,
498 K,
499 &mean_,
500 indexes,
501 pf_gain);
502 }
503
504
505 /* decode latest Wo and voicing and plosive */
506 int plosive_flag = 0;
507
508 //Voiced with Wo
509 if (indexes[3]>0 && indexes[3]<63) {
510 Wo_right = decode_log_Wo(c2const, indexes[3], 6);
511 voicing_right = 1;
512 }
513 //Unvoiced
514 else if(indexes[3] == 0){
515 Wo_right = 2.0*M_PI/100.0;
516 voicing_right = 0;
517 }
518 //indexes[3]=63 (= Plosive) and unvoiced
519 else {
520 Wo_right = 2.0*M_PI/100.0;
521 voicing_right = 0;
522 plosive_flag = 1;
523 }
524
525 /* interpolate 25Hz rate K vec back to 100Hz */
526
527 float *left_vec = prev_rate_K_vec_;
528 float *right_vec = rate_K_vec_;
529 newamp2_interpolate(interpolated_surface_, left_vec, right_vec, K,plosive_flag);
530
531 /* interpolate 25Hz v and Wo back to 100Hz */
532
533 float aWo_[M];
534 int avoicing_[M], aL_[M], i;
535
536 interp_Wo_v(aWo_, aL_, avoicing_, *Wo_left, Wo_right, *voicing_left, voicing_right);
537
538 /* back to rate L amplitudes, synthesis phase for each frame */
539
540 for(i=0; i<M; i++) {
541 model_[i].Wo = aWo_[i];
542 model_[i].L = aL_[i];
543 model_[i].voiced = avoicing_[i];
544 //Plosive Detected
545 if(plosive_flag>0){
546 //First two frames are set to zero
547 if (i<2){
548 n2_resample_rate_L(c2const, &model_[i], &interpolated_surface_[K*i], rate_K_sample_freqs_kHz, K,1);
549 }
550 else{
551 n2_resample_rate_L(c2const, &model_[i], &interpolated_surface_[K*i], rate_K_sample_freqs_kHz, K,0);
552 }
553 }
554 //No Plosive, standard resample
555 else{
556 n2_resample_rate_L(c2const, &model_[i], &interpolated_surface_[K*i], rate_K_sample_freqs_kHz, K,0);
557 }
558 determine_phase(c2const, &H[(MAX_AMP+1)*i], &model_[i], NEWAMP2_PHASE_NFFT, fwd_cfg, inv_cfg);
559 }
560
561 /* update memories for next time */
562
563 for(k=0; k<K; k++) {
564 prev_rate_K_vec_[k] = rate_K_vec_[k];
565 }
566 *Wo_left = Wo_right;
567 *voicing_left = voicing_right;
568
569}
570
diff --git a/newamp2.h b/newamp2.h
deleted file mode 100644
index ec14566..0000000
--- a/newamp2.h
+++ /dev/null
@@ -1,89 +0,0 @@
1/*---------------------------------------------------------------------------*\
2
3 FILE........: newamp2.h
4 AUTHOR......: Thomas Kurin and Stefan Erhardt
5 INSTITUTE...: Institute for Electronics Engineering, University of Erlangen-Nuremberg
6 DATE CREATED: July 2018
7 BASED ON....: "newamp1.h" by David Rowe
8
9 Quantisation functions for the sinusoidal coder, using "newamp1"
10 algorithm that resamples variable rate L [Am} to a fixed rate K then
11 VQs.
12
13\*---------------------------------------------------------------------------*/
14
15/*
16 Copyright Thomas Kurin and Stefan Erhardt 2018
17
18 All rights reserved.
19
20 This program is free software; you can redistribute it and/or modify
21 it under the terms of the GNU Lesser General Public License version 2.1, as
22 published by the Free Software Foundation. This program is
23 distributed in the hope that it will be useful, but WITHOUT ANY
24 WARRANTY; without even the implied warranty of MERCHANTABILITY or
25 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public
26 License for more details.
27
28 You should have received a copy of the GNU Lesser General Public License
29 along with this program; if not, see <http://www.gnu.org/licenses/>.
30*/
31
32#ifndef __NEWAMP2__
33#define __NEWAMP2__
34
35#define NEWAMP2_N_INDEXES 4 /* Number of indexes to pack: vq1, vq2, energy, Wo */
36#define NEWAMP2_PHASE_NFFT 128 /* size of FFT used for phase synthesis */
37#define NEWAMP2_K 29 /* rate K vector length */
38#define NEWAMP2_16K_K 40 /* rate K vector length for 16k Mode */
39
40#include "codec2_fft.h"
41#include "comp.h"
42
43void n2_mel_sample_freqs_kHz(float rate_K_sample_freqs_kHz[], int K);
44void n2_resample_const_rate_f(C2CONST *c2const, MODEL *model, float rate_K_vec[], float rate_K_sample_freqs_kHz[], int K);
45void n2_rate_K_mbest_encode(int *indexes, float *x, float *xq, int ndim);
46void n2_resample_rate_L(C2CONST *c2const, MODEL *model, float rate_K_vec[], float rate_K_sample_freqs_kHz[], int K,int plosive_flag);
47void n2_post_filter_newamp2(float vec[], float sample_freq_kHz[], int K, float pf_gain);
48void newamp2_interpolate(float interpolated_surface_[], float left_vec[], float right_vec[], int K,int plosive_flag);
49void newamp2_model_to_indexes(C2CONST *c2const,
50 int indexes[],
51 MODEL *model,
52 float rate_K_vec[],
53 float rate_K_sample_freqs_kHz[],
54 int K,
55 float *mean,
56 float rate_K_vec_no_mean[],
57 float rate_K_vec_no_mean_[],
58 int plosiv
59 );
60void newamp2_indexes_to_rate_K_vec(float rate_K_vec_[],
61 float rate_K_vec_no_mean_[],
62 float rate_K_sample_freqs_kHz[],
63 int K,
64 float *mean_,
65 int indexes[],
66 float pf_gain);
67void newamp2_16k_indexes_to_rate_K_vec(float rate_K_vec_[],
68 float rate_K_vec_no_mean_[],
69 float rate_K_sample_freqs_kHz[],
70 int K,
71 float *mean_,
72 int indexes[],
73 float pf_gain);
74void newamp2_indexes_to_model(C2CONST *c2const,
75 MODEL model_[],
76 COMP H[],
77 float interpolated_surface_[],
78 float prev_rate_K_vec_[],
79 float *Wo_left,
80 int *voicing_left,
81 float rate_K_sample_freqs_kHz[],
82 int K,
83 codec2_fft_cfg fwd_cfg,
84 codec2_fft_cfg inv_cfg,
85 int indexes[],
86 float pf_gain,
87 int flag16k);
88
89#endif
diff --git a/nlp.c b/nlp.c
index 036f6be..53cf040 100644
--- a/nlp.c
+++ b/nlp.c
@@ -25,115 +25,79 @@
25 along with this program; if not, see <http://www.gnu.org/licenses/>. 25 along with this program; if not, see <http://www.gnu.org/licenses/>.
26*/ 26*/
27 27
28#include "defines.h"
29#include "nlp.h" 28#include "nlp.h"
30#include "dump.h" 29
31#include "codec2_fft.h" 30#include "codec2_fft.h"
31#include "defines.h"
32#include "dump.h"
32#undef PROFILE 33#undef PROFILE
33#include "machdep.h"
34#include "os.h"
35
36#include <assert.h> 34#include <assert.h>
37#include <math.h> 35#include <math.h>
38#include <stdlib.h> 36#include <stdlib.h>
39 37
38#include "machdep.h"
39#include "os.h"
40
40/*---------------------------------------------------------------------------*\ 41/*---------------------------------------------------------------------------*\
41 42
42 DEFINES 43 DEFINES
43 44
44\*---------------------------------------------------------------------------*/ 45\*---------------------------------------------------------------------------*/
45 46
46#define PMAX_M 320 /* maximum NLP analysis window size */ 47#define PMAX_M 320 /* maximum NLP analysis window size */
47#define COEFF 0.95 /* notch filter parameter */ 48#define COEFF 0.95 /* notch filter parameter */
48#define PE_FFT_SIZE 512 /* DFT size for pitch estimation */ 49#define PE_FFT_SIZE 512 /* DFT size for pitch estimation */
49#define DEC 5 /* decimation factor */ 50#define DEC 5 /* decimation factor */
50#define SAMPLE_RATE 8000 51#define SAMPLE_RATE 8000
51#define PI 3.141592654 /* mathematical constant */ 52#define PI 3.141592654 /* mathematical constant */
52#define T 0.1 /* threshold for local minima candidate */ 53#define T 0.1 /* threshold for local minima candidate */
53#define F0_MAX 500 54#define F0_MAX 500
54#define CNLP 0.3 /* post processor constant */ 55#define CNLP 0.3 /* post processor constant */
55#define NLP_NTAP 48 /* Decimation LPF order */ 56#define NLP_NTAP 48 /* Decimation LPF order */
56 57
57/* 8 to 16 kHz sample rate conversion */ 58/* 8 to 16 kHz sample rate conversion */
58 59
59#define FDMDV_OS 2 /* oversampling rate */ 60#define FDMDV_OS 2 /* oversampling rate */
60#define FDMDV_OS_TAPS_16K 48 /* number of OS filter taps at 16kHz */ 61#define FDMDV_OS_TAPS_16K 48 /* number of OS filter taps at 16kHz */
61#define FDMDV_OS_TAPS_8K (FDMDV_OS_TAPS_16K/FDMDV_OS) /* number of OS filter taps at 8kHz */ 62#define FDMDV_OS_TAPS_8K \
63 (FDMDV_OS_TAPS_16K / FDMDV_OS) /* number of OS filter taps at 8kHz */
62 64
63/*---------------------------------------------------------------------------*\ 65/*---------------------------------------------------------------------------*\
64 66
65 GLOBALS 67 GLOBALS
66 68
67\*---------------------------------------------------------------------------*/ 69\*---------------------------------------------------------------------------*/
68 70
69/* 48 tap 600Hz low pass FIR filter coefficients */ 71/* 48 tap 600Hz low pass FIR filter coefficients */
70 72
71const float nlp_fir[] = { 73const float nlp_fir[] = {
72 -1.0818124e-03, 74 -1.0818124e-03, -1.1008344e-03, -9.2768838e-04, -4.2289438e-04,
73 -1.1008344e-03, 75 5.5034190e-04, 2.0029849e-03, 3.7058509e-03, 5.1449415e-03,
74 -9.2768838e-04, 76 5.5924666e-03, 4.3036754e-03, 8.0284511e-04, -4.8204610e-03,
75 -4.2289438e-04, 77 -1.1705810e-02, -1.8199275e-02, -2.2065282e-02, -2.0920610e-02,
76 5.5034190e-04, 78 -1.2808831e-02, 3.2204775e-03, 2.6683811e-02, 5.5520624e-02,
77 2.0029849e-03, 79 8.6305944e-02, 1.1480192e-01, 1.3674206e-01, 1.4867556e-01,
78 3.7058509e-03, 80 1.4867556e-01, 1.3674206e-01, 1.1480192e-01, 8.6305944e-02,
79 5.1449415e-03, 81 5.5520624e-02, 2.6683811e-02, 3.2204775e-03, -1.2808831e-02,
80 5.5924666e-03, 82 -2.0920610e-02, -2.2065282e-02, -1.8199275e-02, -1.1705810e-02,
81 4.3036754e-03, 83 -4.8204610e-03, 8.0284511e-04, 4.3036754e-03, 5.5924666e-03,
82 8.0284511e-04, 84 5.1449415e-03, 3.7058509e-03, 2.0029849e-03, 5.5034190e-04,
83 -4.8204610e-03, 85 -4.2289438e-04, -9.2768838e-04, -1.1008344e-03, -1.0818124e-03};
84 -1.1705810e-02,
85 -1.8199275e-02,
86 -2.2065282e-02,
87 -2.0920610e-02,
88 -1.2808831e-02,
89 3.2204775e-03,
90 2.6683811e-02,
91 5.5520624e-02,
92 8.6305944e-02,
93 1.1480192e-01,
94 1.3674206e-01,
95 1.4867556e-01,
96 1.4867556e-01,
97 1.3674206e-01,
98 1.1480192e-01,
99 8.6305944e-02,
100 5.5520624e-02,
101 2.6683811e-02,
102 3.2204775e-03,
103 -1.2808831e-02,
104 -2.0920610e-02,
105 -2.2065282e-02,
106 -1.8199275e-02,
107 -1.1705810e-02,
108 -4.8204610e-03,
109 8.0284511e-04,
110 4.3036754e-03,
111 5.5924666e-03,
112 5.1449415e-03,
113 3.7058509e-03,
114 2.0029849e-03,
115 5.5034190e-04,
116 -4.2289438e-04,
117 -9.2768838e-04,
118 -1.1008344e-03,
119 -1.0818124e-03
120};
121 86
122typedef struct { 87typedef struct {
123 int Fs; /* sample rate in Hz */ 88 int Fs; /* sample rate in Hz */
124 int m; 89 int m;
125 float w[PMAX_M/DEC]; /* DFT window */ 90 float w[PMAX_M / DEC]; /* DFT window */
126 float sq[PMAX_M]; /* squared speech samples */ 91 float sq[PMAX_M]; /* squared speech samples */
127 float mem_x,mem_y; /* memory for notch filter */ 92 float mem_x, mem_y; /* memory for notch filter */
128 float mem_fir[NLP_NTAP]; /* decimation FIR filter memory */ 93 float mem_fir[NLP_NTAP]; /* decimation FIR filter memory */
129 codec2_fft_cfg fft_cfg; /* kiss FFT config */ 94 codec2_fft_cfg fft_cfg; /* kiss FFT config */
130 float *Sn16k; /* Fs=16kHz input speech vector */ 95 float *Sn16k; /* Fs=16kHz input speech vector */
131 FILE *f; 96 FILE *f;
132} NLP; 97} NLP;
133 98
134float post_process_sub_multiples(COMP Fw[], 99float post_process_sub_multiples(COMP Fw[], int pmin, int pmax, float gmax,
135 int pmin, int pmax, float gmax, int gmax_bin, 100 int gmax_bin, float *prev_f0);
136 float *prev_f0);
137static void fdmdv_16_to_8(float out8k[], float in16k[], int n); 101static void fdmdv_16_to_8(float out8k[], float in16k[], int n);
138 102
139/*---------------------------------------------------------------------------*\ 103/*---------------------------------------------------------------------------*\
@@ -144,56 +108,53 @@ static void fdmdv_16_to_8(float out8k[], float in16k[], int n);
144 108
145\*---------------------------------------------------------------------------*/ 109\*---------------------------------------------------------------------------*/
146 110
147void *nlp_create(C2CONST *c2const) 111void *nlp_create(C2CONST *c2const) {
148{ 112 NLP *nlp;
149 NLP *nlp; 113 int i;
150 int i; 114 int m = c2const->m_pitch;
151 int m = c2const->m_pitch; 115 int Fs = c2const->Fs;
152 int Fs = c2const->Fs; 116
117 nlp = (NLP *)malloc(sizeof(NLP));
118 if (nlp == NULL) return NULL;
153 119
154 nlp = (NLP*)malloc(sizeof(NLP)); 120 assert((Fs == 8000) || (Fs == 16000));
155 if (nlp == NULL) 121 nlp->Fs = Fs;
156 return NULL;
157 122
158 assert((Fs == 8000) || (Fs == 16000)); 123 nlp->m = m;
159 nlp->Fs = Fs;
160 124
161 nlp->m = m; 125 /* if running at 16kHz allocate storage for decimating filter memory */
162 126
163 /* if running at 16kHz allocate storage for decimating filter memory */ 127 if (Fs == 16000) {
128 nlp->Sn16k =
129 (float *)malloc(sizeof(float) * (FDMDV_OS_TAPS_16K + c2const->n_samp));
130 for (i = 0; i < FDMDV_OS_TAPS_16K; i++) {
131 nlp->Sn16k[i] = 0.0;
132 }
133 if (nlp->Sn16k == NULL) {
134 free(nlp);
135 return NULL;
136 }
164 137
165 if (Fs == 16000) { 138 /* most processing occurs at 8 kHz sample rate so halve m */
166 nlp->Sn16k = (float*)malloc(sizeof(float)*(FDMDV_OS_TAPS_16K + c2const->n_samp));
167 for(i=0; i<FDMDV_OS_TAPS_16K; i++) {
168 nlp->Sn16k[i] = 0.0;
169 }
170 if (nlp->Sn16k == NULL) {
171 free(nlp);
172 return NULL;
173 }
174 139
175 /* most processing occurs at 8 kHz sample rate so halve m */ 140 m /= 2;
141 }
176 142
177 m /= 2; 143 assert(m <= PMAX_M);
178 }
179 144
180 assert(m <= PMAX_M); 145 for (i = 0; i < m / DEC; i++) {
181 146 nlp->w[i] = 0.5 - 0.5 * cosf(2 * PI * i / (m / DEC - 1));
182 for(i=0; i<m/DEC; i++) { 147 }
183 nlp->w[i] = 0.5 - 0.5*cosf(2*PI*i/(m/DEC-1));
184 }
185 148
186 for(i=0; i<PMAX_M; i++) 149 for (i = 0; i < PMAX_M; i++) nlp->sq[i] = 0.0;
187 nlp->sq[i] = 0.0; 150 nlp->mem_x = 0.0;
188 nlp->mem_x = 0.0; 151 nlp->mem_y = 0.0;
189 nlp->mem_y = 0.0; 152 for (i = 0; i < NLP_NTAP; i++) nlp->mem_fir[i] = 0.0;
190 for(i=0; i<NLP_NTAP; i++)
191 nlp->mem_fir[i] = 0.0;
192 153
193 nlp->fft_cfg = codec2_fft_alloc (PE_FFT_SIZE, 0, NULL, NULL); 154 nlp->fft_cfg = codec2_fft_alloc(PE_FFT_SIZE, 0, NULL, NULL);
194 assert(nlp->fft_cfg != NULL); 155 assert(nlp->fft_cfg != NULL);
195 156
196 return (void*)nlp; 157 return (void *)nlp;
197} 158}
198 159
199/*---------------------------------------------------------------------------*\ 160/*---------------------------------------------------------------------------*\
@@ -204,17 +165,16 @@ void *nlp_create(C2CONST *c2const)
204 165
205\*---------------------------------------------------------------------------*/ 166\*---------------------------------------------------------------------------*/
206 167
207void nlp_destroy(void *nlp_state) 168void nlp_destroy(void *nlp_state) {
208{ 169 NLP *nlp;
209 NLP *nlp; 170 assert(nlp_state != NULL);
210 assert(nlp_state != NULL); 171 nlp = (NLP *)nlp_state;
211 nlp = (NLP*)nlp_state;
212 172
213 codec2_fft_free(nlp->fft_cfg); 173 codec2_fft_free(nlp->fft_cfg);
214 if (nlp->Fs == 16000) { 174 if (nlp->Fs == 16000) {
215 free(nlp->Sn16k); 175 free(nlp->Sn16k);
216 } 176 }
217 free(nlp_state); 177 free(nlp_state);
218} 178}
219 179
220/*---------------------------------------------------------------------------*\ 180/*---------------------------------------------------------------------------*\
@@ -248,162 +208,157 @@ void nlp_destroy(void *nlp_state)
248\*---------------------------------------------------------------------------*/ 208\*---------------------------------------------------------------------------*/
249 209
250float nlp( 210float nlp(
251 void *nlp_state, 211 void *nlp_state, float Sn[], /* input speech vector */
252 float Sn[], /* input speech vector */ 212 int n, /* frames shift (no. new samples in Sn[]) */
253 int n, /* frames shift (no. new samples in Sn[]) */ 213 float *pitch, /* estimated pitch period in samples at current Fs */
254 float *pitch, /* estimated pitch period in samples at current Fs */ 214 COMP Sw[], /* Freq domain version of Sn[] */
255 COMP Sw[], /* Freq domain version of Sn[] */ 215 float W[], /* Freq domain window */
256 float W[], /* Freq domain window */ 216 float *prev_f0 /* previous pitch f0 in Hz, memory for pitch tracking */
257 float *prev_f0 /* previous pitch f0 in Hz, memory for pitch tracking */ 217) {
258) 218 NLP *nlp;
259{ 219 float notch; /* current notch filter output */
260 NLP *nlp; 220 COMP Fw[PE_FFT_SIZE]; /* DFT of squared signal (input/output) */
261 float notch; /* current notch filter output */ 221 float gmax;
262 COMP Fw[PE_FFT_SIZE]; /* DFT of squared signal (input/output) */ 222 int gmax_bin;
263 float gmax; 223 int m, i, j;
264 int gmax_bin; 224 float best_f0;
265 int m, i, j; 225 PROFILE_VAR(start, tnotch, filter, peakpick, window, fft, magsq, shiftmem);
266 float best_f0; 226
267 PROFILE_VAR(start, tnotch, filter, peakpick, window, fft, magsq, shiftmem); 227 assert(nlp_state != NULL);
268 228 nlp = (NLP *)nlp_state;
269 assert(nlp_state != NULL); 229 m = nlp->m;
270 nlp = (NLP*)nlp_state; 230
271 m = nlp->m; 231 /* Square, notch filter at DC, and LP filter vector */
272 232
273 /* Square, notch filter at DC, and LP filter vector */ 233 /* If running at 16 kHz decimate to 8 kHz, as NLP ws designed for
274 234 Fs = 8kHz. The decimating filter introduces about 3ms of delay,
275 /* If running at 16 kHz decimate to 8 kHz, as NLP ws designed for 235 that shouldn't be a problem as pitch changes slowly. */
276 Fs = 8kHz. The decimating filter introduces about 3ms of delay, 236
277 that shouldn't be a problem as pitch changes slowly. */ 237 if (nlp->Fs == 8000) {
278 238 /* Square latest input samples */
279 if (nlp->Fs == 8000) { 239
280 /* Square latest input samples */ 240 for (i = m - n; i < m; i++) {
281 241 nlp->sq[i] = Sn[i] * Sn[i];
282 for(i=m-n; i<m; i++) {
283 nlp->sq[i] = Sn[i]*Sn[i];
284 }
285 } 242 }
286 else { 243 } else {
287 assert(nlp->Fs == 16000); 244 assert(nlp->Fs == 16000);
288
289 /* re-sample at 8 KHz */
290
291 for(i=0; i<n; i++) {
292 nlp->Sn16k[FDMDV_OS_TAPS_16K+i] = Sn[m-n+i];
293 }
294
295 m /= 2; n /= 2;
296 245
297 float Sn8k[n]; 246 /* re-sample at 8 KHz */
298 fdmdv_16_to_8(Sn8k, &nlp->Sn16k[FDMDV_OS_TAPS_16K], n);
299 247
300 /* Square latest input samples */ 248 for (i = 0; i < n; i++) {
301 249 nlp->Sn16k[FDMDV_OS_TAPS_16K + i] = Sn[m - n + i];
302 for(i=m-n, j=0; i<m; i++, j++) {
303 nlp->sq[i] = Sn8k[j]*Sn8k[j];
304 }
305 assert(j <= n);
306 }
307 //fprintf(stderr, "n: %d m: %d\n", n, m);
308
309 PROFILE_SAMPLE(start);
310
311 for(i=m-n; i<m; i++) { /* notch filter at DC */
312 notch = nlp->sq[i] - nlp->mem_x;
313 notch += COEFF*nlp->mem_y;
314 nlp->mem_x = nlp->sq[i];
315 nlp->mem_y = notch;
316 nlp->sq[i] = notch + 1.0; /* With 0 input vectors to codec,
317 kiss_fft() would take a long
318 time to execute when running in
319 real time. Problem was traced
320 to kiss_fft function call in
321 this function. Adding this small
322 constant fixed problem. Not
323 exactly sure why. */
324 } 250 }
325 251
326 PROFILE_SAMPLE_AND_LOG(tnotch, start, " square and notch"); 252 m /= 2;
253 n /= 2;
327 254
328 for(i=m-n; i<m; i++) { /* FIR filter vector */ 255 float Sn8k[n];
256 fdmdv_16_to_8(Sn8k, &nlp->Sn16k[FDMDV_OS_TAPS_16K], n);
329 257
330 for(j=0; j<NLP_NTAP-1; j++) 258 /* Square latest input samples */
331 nlp->mem_fir[j] = nlp->mem_fir[j+1];
332 nlp->mem_fir[NLP_NTAP-1] = nlp->sq[i];
333 259
334 nlp->sq[i] = 0.0; 260 for (i = m - n, j = 0; i < m; i++, j++) {
335 for(j=0; j<NLP_NTAP; j++) 261 nlp->sq[i] = Sn8k[j] * Sn8k[j];
336 nlp->sq[i] += nlp->mem_fir[j]*nlp_fir[j];
337 } 262 }
338 263 assert(j <= n);
339 PROFILE_SAMPLE_AND_LOG(filter, tnotch, " filter"); 264 }
340 265 // fprintf(stderr, "n: %d m: %d\n", n, m);
341 /* Decimate and DFT */ 266
342 267 PROFILE_SAMPLE(start);
343 for(i=0; i<PE_FFT_SIZE; i++) { 268
344 Fw[i].real = 0.0; 269 for (i = m - n; i < m; i++) { /* notch filter at DC */
345 Fw[i].imag = 0.0; 270 notch = nlp->sq[i] - nlp->mem_x;
346 } 271 notch += COEFF * nlp->mem_y;
347 for(i=0; i<m/DEC; i++) { 272 nlp->mem_x = nlp->sq[i];
348 Fw[i].real = nlp->sq[i*DEC]*nlp->w[i]; 273 nlp->mem_y = notch;
349 } 274 nlp->sq[i] = notch + 1.0; /* With 0 input vectors to codec,
350 PROFILE_SAMPLE_AND_LOG(window, filter, " window"); 275 kiss_fft() would take a long
351 #ifdef DUMP 276 time to execute when running in
352 dump_dec(Fw); 277 real time. Problem was traced
353 #endif 278 to kiss_fft function call in
354 279 this function. Adding this small
355 // FIXME: check if this can be converted to a real fft 280 constant fixed problem. Not
356 // since all imag inputs are 0 281 exactly sure why. */
357 codec2_fft_inplace(nlp->fft_cfg, Fw); 282 }
358 PROFILE_SAMPLE_AND_LOG(fft, window, " fft"); 283
359 284 PROFILE_SAMPLE_AND_LOG(tnotch, start, " square and notch");
360 for(i=0; i<PE_FFT_SIZE; i++) 285
361 Fw[i].real = Fw[i].real*Fw[i].real + Fw[i].imag*Fw[i].imag; 286 for (i = m - n; i < m; i++) { /* FIR filter vector */
362 287
363 PROFILE_SAMPLE_AND_LOG(magsq, fft, " mag sq"); 288 for (j = 0; j < NLP_NTAP - 1; j++) nlp->mem_fir[j] = nlp->mem_fir[j + 1];
364 #ifdef DUMP 289 nlp->mem_fir[NLP_NTAP - 1] = nlp->sq[i];
365 dump_sq(m, nlp->sq); 290
366 dump_Fw(Fw); 291 nlp->sq[i] = 0.0;
367 #endif 292 for (j = 0; j < NLP_NTAP; j++) nlp->sq[i] += nlp->mem_fir[j] * nlp_fir[j];
368 293 }
369 /* todo: express everything in f0, as pitch in samples is dep on Fs */ 294
370 295 PROFILE_SAMPLE_AND_LOG(filter, tnotch, " filter");
371 int pmin = floor(SAMPLE_RATE*P_MIN_S); 296
372 int pmax = floor(SAMPLE_RATE*P_MAX_S); 297 /* Decimate and DFT */
373 298
374 /* find global peak */ 299 for (i = 0; i < PE_FFT_SIZE; i++) {
375 300 Fw[i].real = 0.0;
376 gmax = 0.0; 301 Fw[i].imag = 0.0;
377 gmax_bin = PE_FFT_SIZE*DEC/pmax; 302 }
378 for(i=PE_FFT_SIZE*DEC/pmax; i<=PE_FFT_SIZE*DEC/pmin; i++) { 303 for (i = 0; i < m / DEC; i++) {
379 if (Fw[i].real > gmax) { 304 Fw[i].real = nlp->sq[i * DEC] * nlp->w[i];
380 gmax = Fw[i].real; 305 }
381 gmax_bin = i; 306 PROFILE_SAMPLE_AND_LOG(window, filter, " window");
382 } 307#ifdef DUMP
308 dump_dec(Fw);
309#endif
310
311 // FIXME: check if this can be converted to a real fft
312 // since all imag inputs are 0
313 codec2_fft_inplace(nlp->fft_cfg, Fw);
314 PROFILE_SAMPLE_AND_LOG(fft, window, " fft");
315
316 for (i = 0; i < PE_FFT_SIZE; i++)
317 Fw[i].real = Fw[i].real * Fw[i].real + Fw[i].imag * Fw[i].imag;
318
319 PROFILE_SAMPLE_AND_LOG(magsq, fft, " mag sq");
320#ifdef DUMP
321 dump_sq(m, nlp->sq);
322 dump_Fw(Fw);
323#endif
324
325 /* todo: express everything in f0, as pitch in samples is dep on Fs */
326
327 int pmin = floor(SAMPLE_RATE * P_MIN_S);
328 int pmax = floor(SAMPLE_RATE * P_MAX_S);
329
330 /* find global peak */
331
332 gmax = 0.0;
333 gmax_bin = PE_FFT_SIZE * DEC / pmax;
334 for (i = PE_FFT_SIZE * DEC / pmax; i <= PE_FFT_SIZE * DEC / pmin; i++) {
335 if (Fw[i].real > gmax) {
336 gmax = Fw[i].real;
337 gmax_bin = i;
383 } 338 }
339 }
384 340
385 PROFILE_SAMPLE_AND_LOG(peakpick, magsq, " peak pick"); 341 PROFILE_SAMPLE_AND_LOG(peakpick, magsq, " peak pick");
386 342
387 best_f0 = post_process_sub_multiples(Fw, pmin, pmax, gmax, gmax_bin, prev_f0); 343 best_f0 = post_process_sub_multiples(Fw, pmin, pmax, gmax, gmax_bin, prev_f0);
388 344
389 PROFILE_SAMPLE_AND_LOG(shiftmem, peakpick, " post process"); 345 PROFILE_SAMPLE_AND_LOG(shiftmem, peakpick, " post process");
390 346
391 /* Shift samples in buffer to make room for new samples */ 347 /* Shift samples in buffer to make room for new samples */
392 348
393 for(i=0; i<m-n; i++) 349 for (i = 0; i < m - n; i++) nlp->sq[i] = nlp->sq[i + n];
394 nlp->sq[i] = nlp->sq[i+n];
395 350
396 /* return pitch period in samples and F0 estimate */ 351 /* return pitch period in samples and F0 estimate */
397 352
398 *pitch = (float)nlp->Fs/best_f0; 353 *pitch = (float)nlp->Fs / best_f0;
399 354
400 PROFILE_SAMPLE_AND_LOG2(shiftmem, " shift mem"); 355 PROFILE_SAMPLE_AND_LOG2(shiftmem, " shift mem");
401 356
402 PROFILE_SAMPLE_AND_LOG2(start, " nlp int"); 357 PROFILE_SAMPLE_AND_LOG2(start, " nlp int");
403 358
404 *prev_f0 = best_f0; 359 *prev_f0 = best_f0;
405 360
406 return(best_f0); 361 return (best_f0);
407} 362}
408 363
409/*---------------------------------------------------------------------------*\ 364/*---------------------------------------------------------------------------*\
@@ -427,59 +382,55 @@ float nlp(
427 382
428\*---------------------------------------------------------------------------*/ 383\*---------------------------------------------------------------------------*/
429 384
430float post_process_sub_multiples(COMP Fw[], 385float post_process_sub_multiples(COMP Fw[], int pmin, int pmax, float gmax,
431 int pmin, int pmax, float gmax, int gmax_bin, 386 int gmax_bin, float *prev_f0) {
432 float *prev_f0) 387 int min_bin, cmax_bin;
433{ 388 int mult;
434 int min_bin, cmax_bin; 389 float thresh, best_f0;
435 int mult; 390 int b, bmin, bmax, lmax_bin;
436 float thresh, best_f0; 391 float lmax;
437 int b, bmin, bmax, lmax_bin; 392 int prev_f0_bin;
438 float lmax; 393
439 int prev_f0_bin; 394 /* post process estimate by searching submultiples */
440 395
441 /* post process estimate by searching submultiples */ 396 mult = 2;
442 397 min_bin = PE_FFT_SIZE * DEC / pmax;
443 mult = 2; 398 cmax_bin = gmax_bin;
444 min_bin = PE_FFT_SIZE*DEC/pmax; 399 prev_f0_bin = *prev_f0 * (PE_FFT_SIZE * DEC) / SAMPLE_RATE;
445 cmax_bin = gmax_bin; 400
446 prev_f0_bin = *prev_f0*(PE_FFT_SIZE*DEC)/SAMPLE_RATE; 401 while (gmax_bin / mult >= min_bin) {
447 402 b = gmax_bin / mult; /* determine search interval */
448 while(gmax_bin/mult >= min_bin) { 403 bmin = 0.8 * b;
449 404 bmax = 1.2 * b;
450 b = gmax_bin/mult; /* determine search interval */ 405 if (bmin < min_bin) bmin = min_bin;
451 bmin = 0.8*b; 406
452 bmax = 1.2*b; 407 /* lower threshold to favour previous frames pitch estimate,
453 if (bmin < min_bin) 408 this is a form of pitch tracking */
454 bmin = min_bin; 409
455 410 if ((prev_f0_bin > bmin) && (prev_f0_bin < bmax))
456 /* lower threshold to favour previous frames pitch estimate, 411 thresh = CNLP * 0.5 * gmax;
457 this is a form of pitch tracking */ 412 else
458 413 thresh = CNLP * gmax;
459 if ((prev_f0_bin > bmin) && (prev_f0_bin < bmax)) 414
460 thresh = CNLP*0.5*gmax; 415 lmax = 0;
461 else 416 lmax_bin = bmin;
462 thresh = CNLP*gmax; 417 for (b = bmin; b <= bmax; b++) /* look for maximum in interval */
463 418 if (Fw[b].real > lmax) {
464 lmax = 0; 419 lmax = Fw[b].real;
465 lmax_bin = bmin; 420 lmax_bin = b;
466 for (b=bmin; b<=bmax; b++) /* look for maximum in interval */ 421 }
467 if (Fw[b].real > lmax) { 422
468 lmax = Fw[b].real; 423 if (lmax > thresh)
469 lmax_bin = b; 424 if ((lmax > Fw[lmax_bin - 1].real) && (lmax > Fw[lmax_bin + 1].real)) {
470 } 425 cmax_bin = lmax_bin;
471 426 }
472 if (lmax > thresh) 427
473 if ((lmax > Fw[lmax_bin-1].real) && (lmax > Fw[lmax_bin+1].real)) { 428 mult++;
474 cmax_bin = lmax_bin; 429 }
475 } 430
476 431 best_f0 = (float)cmax_bin * SAMPLE_RATE / (PE_FFT_SIZE * DEC);
477 mult++; 432
478 } 433 return best_f0;
479
480 best_f0 = (float)cmax_bin*SAMPLE_RATE/(PE_FFT_SIZE*DEC);
481
482 return best_f0;
483} 434}
484 435
485/*---------------------------------------------------------------------------*\ 436/*---------------------------------------------------------------------------*\
@@ -502,20 +453,18 @@ float post_process_sub_multiples(COMP Fw[],
502 453
503\*---------------------------------------------------------------------------*/ 454\*---------------------------------------------------------------------------*/
504 455
505static void fdmdv_16_to_8(float out8k[], float in16k[], int n) 456static void fdmdv_16_to_8(float out8k[], float in16k[], int n) {
506{ 457 float acc;
507 float acc; 458 int i, j, k;
508 int i,j,k;
509 459
510 for(i=0, k=0; k<n; i+=FDMDV_OS, k++) { 460 for (i = 0, k = 0; k < n; i += FDMDV_OS, k++) {
511 acc = 0.0; 461 acc = 0.0;
512 for(j=0; j<FDMDV_OS_TAPS_16K; j++) 462 for (j = 0; j < FDMDV_OS_TAPS_16K; j++)
513 acc += fdmdv_os_filter[j]*in16k[i-j]; 463 acc += fdmdv_os_filter[j] * in16k[i - j];
514 out8k[k] = acc; 464 out8k[k] = acc;
515 } 465 }
516 466
517 /* update filter memory */ 467 /* update filter memory */
518 468
519 for(i=-FDMDV_OS_TAPS_16K; i<0; i++) 469 for (i = -FDMDV_OS_TAPS_16K; i < 0; i++) in16k[i] = in16k[i + n * FDMDV_OS];
520 in16k[i] = in16k[i + n*FDMDV_OS];
521} 470}
diff --git a/nlp.h b/nlp.h
index 9e427c2..4b56c4f 100644
--- a/nlp.h
+++ b/nlp.h
@@ -29,10 +29,11 @@
29#define __NLP__ 29#define __NLP__
30 30
31#include "comp.h" 31#include "comp.h"
32#include "defines.h"
32 33
33void *nlp_create(C2CONST *c2const); 34void *nlp_create(C2CONST *c2const);
34void nlp_destroy(void *nlp_state); 35void nlp_destroy(void *nlp_state);
35float nlp(void *nlp_state, float Sn[], int n, 36float nlp(void *nlp_state, float Sn[], int n, float *pitch_samples, COMP Sw[],
36 float *pitch_samples, COMP Sw[], float W[], float *prev_f0); 37 float W[], float *prev_f0);
37 38
38#endif 39#endif
diff --git a/os.h b/os.h
index 14b4713..261f664 100644
--- a/os.h
+++ b/os.h
@@ -1,53 +1,35 @@
1/* Generate using fir1(47,1/2) in Octave */ 1/* Generate using fir1(47,1/2) in Octave */
2 2
3static const float fdmdv_os_filter[]= { 3static const float fdmdv_os_filter[] = {
4 -0.0008215855034550382, 4 -0.0008215855034550382, -0.0007833023901802921, 0.001075563790768233,
5 -0.0007833023901802921, 5 0.001199092367787555, -0.001765309502928316, -0.002055372115328064,
6 0.001075563790768233, 6 0.002986877604154257, 0.003462567920638414, -0.004856570111126334,
7 0.001199092367787555, 7 -0.005563143845031497, 0.007533613299748122, 0.008563932468880897,
8 -0.001765309502928316, 8 -0.01126857129039911, -0.01280782411693687, 0.01651443896361847,
9 -0.002055372115328064, 9 0.01894875110322284, -0.02421604439474981, -0.02845107338464062,
10 0.002986877604154257, 10 0.03672973563400258, 0.04542046150312214, -0.06189165826716491,
11 0.003462567920638414, 11 -0.08721876380763803, 0.1496157094199961, 0.4497962274137046,
12 -0.004856570111126334, 12 0.4497962274137046, 0.1496157094199961, -0.08721876380763803,
13 -0.005563143845031497, 13 -0.0618916582671649, 0.04542046150312216, 0.03672973563400257,
14 0.007533613299748122, 14 -0.02845107338464062, -0.02421604439474984, 0.01894875110322284,
15 0.008563932468880897, 15 0.01651443896361848, -0.01280782411693687, -0.0112685712903991,
16 -0.01126857129039911, 16 0.008563932468880899, 0.007533613299748123, -0.005563143845031501,
17 -0.01280782411693687, 17 -0.004856570111126346, 0.003462567920638419, 0.002986877604154259,
18 0.01651443896361847, 18 -0.002055372115328063, -0.001765309502928318, 0.001199092367787557,
19 0.01894875110322284, 19 0.001075563790768233, -0.0007833023901802925, -0.0008215855034550383};
20 -0.02421604439474981,
21 -0.02845107338464062,
22 0.03672973563400258,
23 0.04542046150312214,
24 -0.06189165826716491,
25 -0.08721876380763803,
26 0.1496157094199961,
27 0.4497962274137046,
28 0.4497962274137046,
29 0.1496157094199961,
30 -0.08721876380763803,
31 -0.0618916582671649,
32 0.04542046150312216,
33 0.03672973563400257,
34 -0.02845107338464062,
35 -0.02421604439474984,
36 0.01894875110322284,
37 0.01651443896361848,
38 -0.01280782411693687,
39 -0.0112685712903991,
40 0.008563932468880899,
41 0.007533613299748123,
42 -0.005563143845031501,
43 -0.004856570111126346,
44 0.003462567920638419,
45 0.002986877604154259,
46 -0.002055372115328063,
47 -0.001765309502928318,
48 0.001199092367787557,
49 0.001075563790768233,
50 -0.0007833023901802925,
51 -0.0008215855034550383
52};
53 20
21/* Generate using fir1(47,1/6) in Octave */
22
23static const float fdmdv_os_filter48[] = {
24 -3.55606818e-04, -8.98615286e-04, -1.40119781e-03, -1.71713852e-03,
25 -1.56471179e-03, -6.28128960e-04, 1.24522223e-03, 3.83138676e-03,
26 6.41309478e-03, 7.85893186e-03, 6.93514929e-03, 2.79361991e-03,
27 -4.51051400e-03, -1.36671853e-02, -2.21034939e-02, -2.64084653e-02,
28 -2.31425052e-02, -9.84218694e-03, 1.40648474e-02, 4.67316298e-02,
29 8.39615986e-02, 1.19925275e-01, 1.48381174e-01, 1.64097819e-01,
30 1.64097819e-01, 1.48381174e-01, 1.19925275e-01, 8.39615986e-02,
31 4.67316298e-02, 1.40648474e-02, -9.84218694e-03, -2.31425052e-02,
32 -2.64084653e-02, -2.21034939e-02, -1.36671853e-02, -4.51051400e-03,
33 2.79361991e-03, 6.93514929e-03, 7.85893186e-03, 6.41309478e-03,
34 3.83138676e-03, 1.24522223e-03, -6.28128960e-04, -1.56471179e-03,
35 -1.71713852e-03, -1.40119781e-03, -8.98615286e-04, -3.55606818e-04};
diff --git a/pack.c b/pack.c
index 1c07230..3181223 100644
--- a/pack.c
+++ b/pack.c
@@ -15,19 +15,20 @@
15 along with this program; if not, see <http://www.gnu.org/licenses/>. 15 along with this program; if not, see <http://www.gnu.org/licenses/>.
16*/ 16*/
17 17
18#include <stdio.h>
19
18#include "defines.h" 20#include "defines.h"
19#include "quantise.h" 21#include "quantise.h"
20#include <stdio.h>
21 22
22/* Compile-time constants */ 23/* Compile-time constants */
23/* Size of unsigned char in bits. Assumes 8 bits-per-char. */ 24/* Size of unsigned char in bits. Assumes 8 bits-per-char. */
24static const unsigned int WordSize = 8; 25static const unsigned int WordSize = 8;
25 26
26/* Mask to pick the bit component out of bitIndex. */ 27/* Mask to pick the bit component out of bitIndex. */
27static const unsigned int IndexMask = 0x7; 28static const unsigned int IndexMask = 0x7;
28 29
29/* Used to pick the word component out of bitIndex. */ 30/* Used to pick the word component out of bitIndex. */
30static const unsigned int ShiftRight = 3; 31static const unsigned int ShiftRight = 3;
31 32
32/** Pack a bit field into a bit string, encoding the field in Gray code. 33/** Pack a bit field into a bit string, encoding the field in Gray code.
33 * 34 *
@@ -44,86 +45,76 @@ static const unsigned int ShiftRight = 3;
44 * compatibility with the rest of the code, indices are always expected to 45 * compatibility with the rest of the code, indices are always expected to
45 * be >= 0. 46 * be >= 0.
46 */ 47 */
47void 48void pack(unsigned char* bitArray, /* The output bit string. */
48pack( 49 unsigned int* bitIndex, /* Index into the string in BITS, not bytes.*/
49 unsigned char * bitArray, /* The output bit string. */ 50 int field, /* The bit field to be packed. */
50 unsigned int * bitIndex, /* Index into the string in BITS, not bytes.*/ 51 unsigned int fieldWidth /* Width of the field in BITS, not bytes. */
51 int field, /* The bit field to be packed. */ 52) {
52 unsigned int fieldWidth/* Width of the field in BITS, not bytes. */ 53 pack_natural_or_gray(bitArray, bitIndex, field, fieldWidth, 1);
53 )
54{
55 pack_natural_or_gray(bitArray, bitIndex, field, fieldWidth, 1);
56} 54}
57 55
58void 56void pack_natural_or_gray(
59pack_natural_or_gray( 57 unsigned char* bitArray, /* The output bit string. */
60 unsigned char * bitArray, /* The output bit string. */ 58 unsigned int* bitIndex, /* Index into the string in BITS, not bytes.*/
61 unsigned int * bitIndex, /* Index into the string in BITS, not bytes.*/ 59 int field, /* The bit field to be packed. */
62 int field, /* The bit field to be packed. */ 60 unsigned int fieldWidth, /* Width of the field in BITS, not bytes. */
63 unsigned int fieldWidth,/* Width of the field in BITS, not bytes. */ 61 unsigned int gray /* non-zero for gray coding */
64 unsigned int gray /* non-zero for gray coding */ 62) {
65 )
66{
67 if (gray) { 63 if (gray) {
68 /* Convert the field to Gray code */ 64 /* Convert the field to Gray code */
69 field = (field >> 1) ^ field; 65 field = (field >> 1) ^ field;
70 } 66 }
71 67
72 do { 68 do {
73 unsigned int bI = *bitIndex; 69 unsigned int bI = *bitIndex;
74 unsigned int bitsLeft = WordSize - (bI & IndexMask); 70 unsigned int bitsLeft = WordSize - (bI & IndexMask);
75 unsigned int sliceWidth = 71 unsigned int sliceWidth = bitsLeft < fieldWidth ? bitsLeft : fieldWidth;
76 bitsLeft < fieldWidth ? bitsLeft : fieldWidth; 72 unsigned int wordIndex = bI >> ShiftRight;
77 unsigned int wordIndex = bI >> ShiftRight;
78 73
79 bitArray[wordIndex] |= 74 bitArray[wordIndex] |= ((unsigned char)((field >> (fieldWidth - sliceWidth))
80 ((unsigned char)((field >> (fieldWidth - sliceWidth)) 75 << (bitsLeft - sliceWidth)));
81 << (bitsLeft - sliceWidth)));
82 76
83 *bitIndex = bI + sliceWidth; 77 *bitIndex = bI + sliceWidth;
84 fieldWidth -= sliceWidth; 78 fieldWidth -= sliceWidth;
85 } while ( fieldWidth != 0 ); 79 } while (fieldWidth != 0);
86} 80}
87 81
88/** Unpack a field from a bit string, converting from Gray code to binary. 82/** Unpack a field from a bit string, converting from Gray code to binary.
89 * 83 *
90 */ 84 */
91int 85int unpack(
92unpack( 86 const unsigned char* bitArray, /* The input bit string. */
93 const unsigned char * bitArray, /* The input bit string. */ 87 unsigned int* bitIndex, /* Index into the string in BITS, not bytes.*/
94 unsigned int * bitIndex, /* Index into the string in BITS, not bytes.*/ 88 unsigned int fieldWidth /* Width of the field in BITS, not bytes. */
95 unsigned int fieldWidth/* Width of the field in BITS, not bytes. */ 89) {
96 ) 90 return unpack_natural_or_gray(bitArray, bitIndex, fieldWidth, 1);
97{
98 return unpack_natural_or_gray(bitArray, bitIndex, fieldWidth, 1);
99} 91}
100 92
101/** Unpack a field from a bit string, to binary, optionally using 93/** Unpack a field from a bit string, to binary, optionally using
102 * natural or Gray code. 94 * natural or Gray code.
103 * 95 *
104 */ 96 */
105int 97int unpack_natural_or_gray(
106unpack_natural_or_gray( 98 const unsigned char* bitArray, /* The input bit string. */
107 const unsigned char * bitArray, /* The input bit string. */ 99 unsigned int* bitIndex, /* Index into the string in BITS, not bytes.*/
108 unsigned int * bitIndex, /* Index into the string in BITS, not bytes.*/ 100 unsigned int fieldWidth, /* Width of the field in BITS, not bytes. */
109 unsigned int fieldWidth,/* Width of the field in BITS, not bytes. */ 101 unsigned int gray /* non-zero for Gray coding */
110 unsigned int gray /* non-zero for Gray coding */ 102) {
111 ) 103 unsigned int field = 0;
112{ 104 unsigned int t;
113 unsigned int field = 0;
114 unsigned int t;
115 105
116 do { 106 do {
117 unsigned int bI = *bitIndex; 107 unsigned int bI = *bitIndex;
118 unsigned int bitsLeft = WordSize - (bI & IndexMask); 108 unsigned int bitsLeft = WordSize - (bI & IndexMask);
119 unsigned int sliceWidth = 109 unsigned int sliceWidth = bitsLeft < fieldWidth ? bitsLeft : fieldWidth;
120 bitsLeft < fieldWidth ? bitsLeft : fieldWidth;
121 110
122 field |= (((bitArray[bI >> ShiftRight] >> (bitsLeft - sliceWidth)) & ((1 << sliceWidth) - 1)) << (fieldWidth - sliceWidth)); 111 field |= (((bitArray[bI >> ShiftRight] >> (bitsLeft - sliceWidth)) &
112 ((1 << sliceWidth) - 1))
113 << (fieldWidth - sliceWidth));
123 114
124 *bitIndex = bI + sliceWidth; 115 *bitIndex = bI + sliceWidth;
125 fieldWidth -= sliceWidth; 116 fieldWidth -= sliceWidth;
126 } while ( fieldWidth != 0 ); 117 } while (fieldWidth != 0);
127 118
128 if (gray) { 119 if (gray) {
129 /* Convert from Gray code to binary. Works for maximum 8-bit fields. */ 120 /* Convert from Gray code to binary. Works for maximum 8-bit fields. */
@@ -131,8 +122,7 @@ unpack_natural_or_gray(
131 t ^= (t >> 4); 122 t ^= (t >> 4);
132 t ^= (t >> 2); 123 t ^= (t >> 2);
133 t ^= (t >> 1); 124 t ^= (t >> 1);
134 } 125 } else {
135 else {
136 t = field; 126 t = field;
137 } 127 }
138 128
diff --git a/phase.c b/phase.c
index e486613..dec8793 100644
--- a/phase.c
+++ b/phase.c
@@ -25,18 +25,19 @@
25 along with this program; if not,see <http://www.gnu.org/licenses/>. 25 along with this program; if not,see <http://www.gnu.org/licenses/>.
26*/ 26*/
27 27
28#include "defines.h"
29#include "phase.h" 28#include "phase.h"
30#include "kiss_fft.h"
31#include "comp.h"
32#include "comp_prim.h"
33#include "sine.h"
34 29
35#include <assert.h> 30#include <assert.h>
36#include <ctype.h> 31#include <ctype.h>
37#include <math.h> 32#include <math.h>
38#include <string.h>
39#include <stdlib.h> 33#include <stdlib.h>
34#include <string.h>
35
36#include "comp.h"
37#include "comp_prim.h"
38#include "defines.h"
39#include "kiss_fft.h"
40#include "sine.h"
40 41
41/*---------------------------------------------------------------------------*\ 42/*---------------------------------------------------------------------------*\
42 43
@@ -47,25 +48,23 @@
47 48
48\*---------------------------------------------------------------------------*/ 49\*---------------------------------------------------------------------------*/
49 50
50void sample_phase(MODEL *model, 51void sample_phase(MODEL *model, COMP H[],
51 COMP H[], 52 COMP A[] /* LPC analysis filter in freq domain */
52 COMP A[] /* LPC analysis filter in freq domain */ 53) {
53 ) 54 int m, b;
54{ 55 float r;
55 int m, b;
56 float r;
57 56
58 r = TWO_PI/(FFT_ENC); 57 r = TWO_PI / (FFT_ENC);
59 58
60 /* Sample phase at harmonics */ 59 /* Sample phase at harmonics */
61 60
62 for(m=1; m<=model->L; m++) { 61 for (m = 1; m <= model->L; m++) {
63 b = (int)(m*model->Wo/r + 0.5); 62 b = (int)(m * model->Wo / r + 0.5);
64 H[m] = cconj(A[b]); /* synth filter 1/A is opposite phase to analysis filter */ 63 H[m] =
65 } 64 cconj(A[b]); /* synth filter 1/A is opposite phase to analysis filter */
65 }
66} 66}
67 67
68
69/*---------------------------------------------------------------------------*\ 68/*---------------------------------------------------------------------------*\
70 69
71 phase_synth_zero_order() 70 phase_synth_zero_order()
@@ -158,64 +157,56 @@ void sample_phase(MODEL *model,
158\*---------------------------------------------------------------------------*/ 157\*---------------------------------------------------------------------------*/
159 158
160void phase_synth_zero_order( 159void phase_synth_zero_order(
161 int n_samp, 160 int n_samp, MODEL *model,
162 MODEL *model, 161 float *ex_phase, /* excitation phase of fundamental */
163 float *ex_phase, /* excitation phase of fundamental */ 162 COMP H[] /* L synthesis filter freq domain samples */
164 COMP H[] /* L synthesis filter freq domain samples */ 163
165 164) {
166) 165 int m;
167{ 166 float new_phi;
168 int m; 167 COMP Ex[MAX_AMP + 1]; /* excitation samples */
169 float new_phi; 168 COMP A_[MAX_AMP + 1]; /* synthesised harmonic samples */
170 COMP Ex[MAX_AMP+1]; /* excitation samples */ 169
171 COMP A_[MAX_AMP+1]; /* synthesised harmonic samples */ 170 /*
172 171 Update excitation fundamental phase track, this sets the position
173 /* 172 of each pitch pulse during voiced speech. After much experiment
174 Update excitation fundamental phase track, this sets the position 173 I found that using just this frame's Wo improved quality for UV
175 of each pitch pulse during voiced speech. After much experiment 174 sounds compared to interpolating two frames Wo like this:
176 I found that using just this frame's Wo improved quality for UV 175
177 sounds compared to interpolating two frames Wo like this: 176 ex_phase[0] += (*prev_Wo+model->Wo)*N_SAMP/2;
178 177 */
179 ex_phase[0] += (*prev_Wo+model->Wo)*N_SAMP/2; 178
180 */ 179 ex_phase[0] += (model->Wo) * n_samp;
181 180 ex_phase[0] -= TWO_PI * floorf(ex_phase[0] / TWO_PI + 0.5);
182 ex_phase[0] += (model->Wo)*n_samp; 181
183 ex_phase[0] -= TWO_PI*floorf(ex_phase[0]/TWO_PI + 0.5); 182 for (m = 1; m <= model->L; m++) {
184 183 /* generate excitation */
185 for(m=1; m<=model->L; m++) { 184
186 185 if (model->voiced) {
187 /* generate excitation */ 186 Ex[m].real = cosf(ex_phase[0] * m);
188 187 Ex[m].imag = sinf(ex_phase[0] * m);
189 if (model->voiced) { 188 } else {
190 189 /* When a few samples were tested I found that LPC filter
191 Ex[m].real = cosf(ex_phase[0]*m); 190 phase is not needed in the unvoiced case, but no harm in
192 Ex[m].imag = sinf(ex_phase[0]*m); 191 keeping it.
193 } 192 */
194 else { 193 float phi = TWO_PI * (float)codec2_rand() / CODEC2_RAND_MAX;
195 194 Ex[m].real = cosf(phi);
196 /* When a few samples were tested I found that LPC filter 195 Ex[m].imag = sinf(phi);
197 phase is not needed in the unvoiced case, but no harm in 196 }
198 keeping it.
199 */
200 float phi = TWO_PI*(float)codec2_rand()/CODEC2_RAND_MAX;
201 Ex[m].real = cosf(phi);
202 Ex[m].imag = sinf(phi);
203 }
204
205 /* filter using LPC filter */
206 197
207 A_[m].real = H[m].real*Ex[m].real - H[m].imag*Ex[m].imag; 198 /* filter using LPC filter */
208 A_[m].imag = H[m].imag*Ex[m].real + H[m].real*Ex[m].imag;
209 199
210 /* modify sinusoidal phase */ 200 A_[m].real = H[m].real * Ex[m].real - H[m].imag * Ex[m].imag;
201 A_[m].imag = H[m].imag * Ex[m].real + H[m].real * Ex[m].imag;
211 202
212 new_phi = atan2f(A_[m].imag, A_[m].real+1E-12); 203 /* modify sinusoidal phase */
213 model->phi[m] = new_phi;
214 }
215 204
205 new_phi = atan2f(A_[m].imag, A_[m].real + 1E-12);
206 model->phi[m] = new_phi;
207 }
216} 208}
217 209
218
219/*---------------------------------------------------------------------------*\ 210/*---------------------------------------------------------------------------*\
220 211
221 FUNCTION....: mag_to_phase 212 FUNCTION....: mag_to_phase
@@ -230,60 +221,55 @@ void phase_synth_zero_order(
230 221
231\*---------------------------------------------------------------------------*/ 222\*---------------------------------------------------------------------------*/
232 223
233void mag_to_phase(float phase[], /* Nfft/2+1 output phase samples in radians */ 224void mag_to_phase(
234 float Gdbfk[], /* Nfft/2+1 postive freq amplitudes samples in dB */ 225 float phase[], /* Nfft/2+1 output phase samples in radians */
235 int Nfft, 226 float Gdbfk[], /* Nfft/2+1 positive freq amplitudes samples in dB */
236 codec2_fft_cfg fft_fwd_cfg, 227 int Nfft, codec2_fft_cfg fft_fwd_cfg, codec2_fft_cfg fft_inv_cfg) {
237 codec2_fft_cfg fft_inv_cfg 228 COMP Sdb[Nfft], c[Nfft], cf[Nfft], Cf[Nfft];
238 ) 229 int Ns = Nfft / 2 + 1;
239{ 230 int i;
240 COMP Sdb[Nfft], c[Nfft], cf[Nfft], Cf[Nfft]; 231
241 int Ns = Nfft/2+1; 232 /* install negative frequency components, 1/Nfft takes into
242 int i; 233 account kiss fft lack of scaling on ifft */
243 234
244 /* install negative frequency components, 1/Nfft takes into 235 Sdb[0].real = Gdbfk[0];
245 account kiss fft lack of scaling on ifft */ 236 Sdb[0].imag = 0.0;
246 237 for (i = 1; i < Ns; i++) {
247 Sdb[0].real = Gdbfk[0]; 238 Sdb[i].real = Sdb[Nfft - i].real = Gdbfk[i];
248 Sdb[0].imag = 0.0; 239 Sdb[i].imag = Sdb[Nfft - i].imag = 0.0;
249 for(i=1; i<Ns; i++) { 240 }
250 Sdb[i].real = Sdb[Nfft-i].real = Gdbfk[i]; 241
251 Sdb[i].imag = Sdb[Nfft-i].imag = 0.0; 242 /* compute real cepstrum from log magnitude spectrum */
252 } 243
253 244 codec2_fft(fft_inv_cfg, Sdb, c);
254 /* compute real cepstrum from log magnitude spectrum */ 245 for (i = 0; i < Nfft; i++) {
255 246 c[i].real /= (float)Nfft;
256 codec2_fft(fft_inv_cfg, Sdb, c); 247 c[i].imag /= (float)Nfft;
257 for(i=0; i<Nfft; i++) { 248 }
258 c[i].real /= (float)Nfft; 249
259 c[i].imag /= (float)Nfft; 250 /* Fold cepstrum to reflect non-min-phase zeros inside unit circle */
260 } 251
261 252 cf[0] = c[0];
262 /* Fold cepstrum to reflect non-min-phase zeros inside unit circle */ 253 for (i = 1; i < Ns - 1; i++) {
263 254 cf[i] = cadd(c[i], c[Nfft - i]);
264 cf[0] = c[0]; 255 }
265 for(i=1; i<Ns-1; i++) { 256 cf[Ns - 1] = c[Ns - 1];
266 cf[i] = cadd(c[i],c[Nfft-i]); 257 for (i = Ns; i < Nfft; i++) {
267 } 258 cf[i].real = 0.0;
268 cf[Ns-1] = c[Ns-1]; 259 cf[i].imag = 0.0;
269 for(i=Ns; i<Nfft; i++) { 260 }
270 cf[i].real = 0.0; 261
271 cf[i].imag = 0.0; 262 /* Cf = dB_magnitude + j * minimum_phase */
272 } 263
273 264 codec2_fft(fft_fwd_cfg, cf, Cf);
274 /* Cf = dB_magnitude + j * minimum_phase */ 265
275 266 /* The maths says we are meant to be using log(x), not 20*log10(x),
276 codec2_fft(fft_fwd_cfg, cf, Cf); 267 so we need to scale the phase to account for this:
277 268 log(x) = 20*log10(x)/scale */
278 /* The maths says we are meant to be using log(x), not 20*log10(x), 269
279 so we need to scale the phase to account for this: 270 float scale = (20.0 / logf(10.0));
280 log(x) = 20*log10(x)/scale */ 271
281 272 for (i = 0; i < Ns; i++) {
282 float scale = (20.0/logf(10.0)); 273 phase[i] = Cf[i].imag / scale;
283 274 }
284 for(i=0; i<Ns; i++) {
285 phase[i] = Cf[i].imag/scale;
286 }
287
288
289} 275}
diff --git a/phase.h b/phase.h
index 92909f0..b7f0413 100644
--- a/phase.h
+++ b/phase.h
@@ -32,8 +32,10 @@
32#include "comp.h" 32#include "comp.h"
33 33
34void sample_phase(MODEL *model, COMP filter_phase[], COMP A[]); 34void sample_phase(MODEL *model, COMP filter_phase[], COMP A[]);
35void phase_synth_zero_order(int n_samp, MODEL *model, float *ex_phase, COMP filter_phase[]); 35void phase_synth_zero_order(int n_samp, MODEL *model, float *ex_phase,
36 COMP filter_phase[]);
36 37
37void mag_to_phase(float phase[], float Gdbfk[], int Nfft, codec2_fft_cfg fwd_cfg, codec2_fft_cfg inv_cfg); 38void mag_to_phase(float phase[], float Gdbfk[], int Nfft,
39 codec2_fft_cfg fwd_cfg, codec2_fft_cfg inv_cfg);
38 40
39#endif 41#endif
diff --git a/postfilter.c b/postfilter.c
index 6542c7c..e46d488 100644
--- a/postfilter.c
+++ b/postfilter.c
@@ -27,16 +27,17 @@
27 along with this program; if not, see <http://www.gnu.org/licenses/>. 27 along with this program; if not, see <http://www.gnu.org/licenses/>.
28*/ 28*/
29 29
30#include "postfilter.h"
31
30#include <assert.h> 32#include <assert.h>
31#include <stdlib.h>
32#include <stdio.h>
33#include <math.h> 33#include <math.h>
34#include <stdio.h>
35#include <stdlib.h>
34 36
35#include "defines.h"
36#include "comp.h" 37#include "comp.h"
38#include "defines.h"
37#include "dump.h" 39#include "dump.h"
38#include "sine.h" 40#include "sine.h"
39#include "postfilter.h"
40 41
41/*---------------------------------------------------------------------------*\ 42/*---------------------------------------------------------------------------*\
42 43
@@ -44,13 +45,14 @@
44 45
45\*---------------------------------------------------------------------------*/ 46\*---------------------------------------------------------------------------*/
46 47
47#define BG_THRESH 40.0 /* only consider low levels signals for bg_est */ 48#define BG_THRESH 40.0 /* only consider low levels signals for bg_est */
48#define BG_BETA 0.1 /* averaging filter constant */ 49#define BG_BETA 0.1 /* averaging filter constant */
49#define BG_MARGIN 6.0 /* harmonics this far above BG noise are 50#define BG_MARGIN \
50 randomised. Helped make bg noise less 51 6.0 /* harmonics this far above BG noise are \
51 spikey (impulsive) for mmt1, but speech was 52 randomised. Helped make bg noise less \
52 perhaps a little rougher. 53 spikey (impulsive) for mmt1, but speech was \
53 */ 54 perhaps a little rougher. \
55 */
54 56
55/*---------------------------------------------------------------------------*\ 57/*---------------------------------------------------------------------------*\
56 58
@@ -82,7 +84,7 @@
82 1/ If someone says "aaaaaaaahhhhhhhhh" will background estimator track 84 1/ If someone says "aaaaaaaahhhhhhhhh" will background estimator track
83 up to speech level? This would be a bad thing. 85 up to speech level? This would be a bad thing.
84 86
85 2/ If background noise suddenly dissapears from the source speech does 87 2/ If background noise suddenly disappears from the source speech does
86 estimate drop quickly? What is noise suddenly re-appears? 88 estimate drop quickly? What is noise suddenly re-appears?
87 89
88 3/ Background noise with a non-flat sepctrum. Current algorithm just 90 3/ Background noise with a non-flat sepctrum. Current algorithm just
@@ -98,45 +100,39 @@
98 100
99\*---------------------------------------------------------------------------*/ 101\*---------------------------------------------------------------------------*/
100 102
101void postfilter( 103void postfilter(MODEL *model, float *bg_est) {
102 MODEL *model, 104 int m, uv;
103 float *bg_est
104)
105{
106 int m, uv;
107 float e, thresh; 105 float e, thresh;
108 106
109 /* determine average energy across spectrum */ 107 /* determine average energy across spectrum */
110 108
111 e = 1E-12; 109 e = 1E-12;
112 for(m=1; m<=model->L; m++) 110 for (m = 1; m <= model->L; m++) e += model->A[m] * model->A[m];
113 e += model->A[m]*model->A[m];
114 111
115 assert(e > 0.0); 112 assert(e > 0.0);
116 e = 10.0*log10f(e/model->L); 113 e = 10.0 * log10f(e / model->L);
117 114
118 /* If beneath threhold, update bg estimate. The idea 115 /* If beneath threshold, update bg estimate. The idea
119 of the threshold is to prevent updating during high level 116 of the threshold is to prevent updating during high level
120 speech. */ 117 speech. */
121 118
122 if ((e < BG_THRESH) && !model->voiced) 119 if ((e < BG_THRESH) && !model->voiced)
123 *bg_est = *bg_est*(1.0 - BG_BETA) + e*BG_BETA; 120 *bg_est = *bg_est * (1.0 - BG_BETA) + e * BG_BETA;
124 121
125 /* now mess with phases during voiced frames to make any harmonics 122 /* now mess with phases during voiced frames to make any harmonics
126 less then our background estimate unvoiced. 123 less then our background estimate unvoiced.
127 */ 124 */
128 125
129 uv = 0; 126 uv = 0;
130 thresh = POW10F((*bg_est + BG_MARGIN)/20.0); 127 thresh = POW10F((*bg_est + BG_MARGIN) / 20.0);
131 if (model->voiced) 128 if (model->voiced)
132 for(m=1; m<=model->L; m++) 129 for (m = 1; m <= model->L; m++)
133 if (model->A[m] < thresh) { 130 if (model->A[m] < thresh) {
134 model->phi[m] = (TWO_PI/CODEC2_RAND_MAX)*(float)codec2_rand(); 131 model->phi[m] = (TWO_PI / CODEC2_RAND_MAX) * (float)codec2_rand();
135 uv++; 132 uv++;
136 } 133 }
137 134
138#ifdef DUMP 135#ifdef DUMP
139 dump_bg(e, *bg_est, 100.0*uv/model->L); 136 dump_bg(e, *bg_est, 100.0 * uv / model->L);
140#endif 137#endif
141
142} 138}
diff --git a/postfilter.h b/postfilter.h
index 156714e..ecf7f39 100644
--- a/postfilter.h
+++ b/postfilter.h
@@ -28,6 +28,8 @@
28#ifndef __POSTFILTER__ 28#ifndef __POSTFILTER__
29#define __POSTFILTER__ 29#define __POSTFILTER__
30 30
31#include "defines.h"
32
31void postfilter(MODEL *model, float *bg_est); 33void postfilter(MODEL *model, float *bg_est);
32 34
33#endif 35#endif
diff --git a/quantise.c b/quantise.c
index 37bf8be..da1d821 100644
--- a/quantise.c
+++ b/quantise.c
@@ -24,26 +24,27 @@
24 24
25*/ 25*/
26 26
27#include "quantise.h"
28
27#include <assert.h> 29#include <assert.h>
28#include <ctype.h> 30#include <ctype.h>
31#include <math.h>
29#include <stdio.h> 32#include <stdio.h>
30#include <stdlib.h> 33#include <stdlib.h>
31#include <string.h> 34#include <string.h>
32#include <math.h>
33 35
36#include "codec2_fft.h"
34#include "defines.h" 37#include "defines.h"
35#include "dump.h" 38#include "dump.h"
36#include "quantise.h"
37#include "lpc.h" 39#include "lpc.h"
38#include "lsp.h" 40#include "lsp.h"
39#include "codec2_fft.h"
40#include "phase.h"
41#include "mbest.h" 41#include "mbest.h"
42#include "phase.h"
42 43
43#undef PROFILE 44#undef PROFILE
44#include "machdep.h" 45#include "machdep.h"
45 46
46#define LSP_DELTA1 0.01 /* grid spacing for LSP root searches */ 47#define LSP_DELTA1 0.01 /* grid spacing for LSP root searches */
47 48
48/*---------------------------------------------------------------------------*\ 49/*---------------------------------------------------------------------------*\
49 50
@@ -52,7 +53,7 @@
52\*---------------------------------------------------------------------------*/ 53\*---------------------------------------------------------------------------*/
53 54
54float speech_to_uq_lsps(float lsp[], float ak[], float Sn[], float w[], 55float speech_to_uq_lsps(float lsp[], float ak[], float Sn[], float w[],
55 int m_pitch, int order); 56 int m_pitch, int order);
56 57
57/*---------------------------------------------------------------------------*\ 58/*---------------------------------------------------------------------------*\
58 59
@@ -60,46 +61,11 @@ float speech_to_uq_lsps(float lsp[], float ak[], float Sn[], float w[],
60 61
61\*---------------------------------------------------------------------------*/ 62\*---------------------------------------------------------------------------*/
62 63
63int lsp_bits(int i) { 64int lsp_bits(int i) { return lsp_cb[i].log2m; }
64 return lsp_cb[i].log2m;
65}
66 65
67int lspd_bits(int i) { 66int lspd_bits(int i) { return lsp_cbd[i].log2m; }
68 return lsp_cbd[i].log2m;
69}
70 67
71#ifndef CORTEX_M4 68int lsp_pred_vq_bits(int i) { return lsp_cbjmv[i].log2m; }
72int mel_bits(int i) {
73 return mel_cb[i].log2m;
74}
75
76int lspmelvq_cb_bits(int i) {
77 return lspmelvq_cb[i].log2m;
78}
79#endif
80
81#ifdef __EXPERIMENTAL__
82int lspdt_bits(int i) {
83 return lsp_cbdt[i].log2m;
84}
85#endif
86
87int lsp_pred_vq_bits(int i) {
88 return lsp_cbjvm[i].log2m;
89}
90
91/*---------------------------------------------------------------------------*\
92
93 quantise_init
94
95 Loads the entire LSP quantiser comprised of several vector quantisers
96 (codebooks).
97
98\*---------------------------------------------------------------------------*/
99
100void quantise_init()
101{
102}
103 69
104/*---------------------------------------------------------------------------*\ 70/*---------------------------------------------------------------------------*\
105 71
@@ -111,7 +77,7 @@ void quantise_init()
111 77
112\*---------------------------------------------------------------------------*/ 78\*---------------------------------------------------------------------------*/
113 79
114long quantise(const float * cb, float vec[], float w[], int k, int m, float *se) 80long quantise(const float *cb, float vec[], float w[], int k, int m, float *se)
115/* float cb[][K]; current VQ codebook */ 81/* float cb[][K]; current VQ codebook */
116/* float vec[]; vector to quantise */ 82/* float vec[]; vector to quantise */
117/* float w[]; weighting vector */ 83/* float w[]; weighting vector */
@@ -119,363 +85,123 @@ long quantise(const float * cb, float vec[], float w[], int k, int m, float *se)
119/* int m; size of codebook */ 85/* int m; size of codebook */
120/* float *se; accumulated squared error */ 86/* float *se; accumulated squared error */
121{ 87{
122 float e; /* current error */ 88 float e; /* current error */
123 long besti; /* best index so far */ 89 long besti; /* best index so far */
124 float beste; /* best error so far */ 90 float beste; /* best error so far */
125 long j; 91 long j;
126 int i; 92 int i;
127 float diff; 93 float diff;
128
129 besti = 0;
130 beste = 1E32;
131 for(j=0; j<m; j++) {
132 e = 0.0;
133 for(i=0; i<k; i++) {
134 diff = cb[j*k+i]-vec[i];
135 e += (diff*w[i] * diff*w[i]);
136 }
137 if (e < beste) {
138 beste = e;
139 besti = j;
140 }
141 }
142
143 *se += beste;
144
145 return(besti);
146}
147
148
149
150/*---------------------------------------------------------------------------*\
151
152 encode_lspds_scalar()
153
154 Scalar/VQ LSP difference quantiser.
155
156\*---------------------------------------------------------------------------*/
157 94
158void encode_lspds_scalar( 95 besti = 0;
159 int indexes[], 96 beste = 1E32;
160 float lsp[], 97 for (j = 0; j < m; j++) {
161 int order 98 e = 0.0;
162) 99 for (i = 0; i < k; i++) {
163{ 100 diff = cb[j * k + i] - vec[i];
164 int i,k,m; 101 e += (diff * w[i] * diff * w[i]);
165 float lsp_hz[order];
166 float lsp__hz[order];
167 float dlsp[order];
168 float dlsp_[order];
169 float wt[order];
170 const float *cb;
171 float se;
172
173 for(i=0; i<order; i++) {
174 wt[i] = 1.0;
175 } 102 }
176 103 if (e < beste) {
177 /* convert from radians to Hz so we can use human readable 104 beste = e;
178 frequencies */ 105 besti = j;
179
180 for(i=0; i<order; i++)
181 lsp_hz[i] = (4000.0/PI)*lsp[i];
182
183 //printf("\n");
184
185 wt[0] = 1.0;
186 for(i=0; i<order; i++) {
187
188 /* find difference from previous qunatised lsp */
189
190 if (i)
191 dlsp[i] = lsp_hz[i] - lsp__hz[i-1];
192 else
193 dlsp[0] = lsp_hz[0];
194
195 k = lsp_cbd[i].k;
196 m = lsp_cbd[i].m;
197 cb = lsp_cbd[i].cb;
198 indexes[i] = quantise(cb, &dlsp[i], wt, k, m, &se);
199 dlsp_[i] = cb[indexes[i]*k];
200
201
202 if (i)
203 lsp__hz[i] = lsp__hz[i-1] + dlsp_[i];
204 else
205 lsp__hz[0] = dlsp_[0];
206
207 //printf("%d lsp %3.2f dlsp %3.2f dlsp_ %3.2f lsp_ %3.2f\n", i, lsp_hz[i], dlsp[i], dlsp_[i], lsp__hz[i]);
208 } 106 }
107 }
209 108
210} 109 *se += beste;
211
212
213void decode_lspds_scalar(
214 float lsp_[],
215 int indexes[],
216 int order
217)
218{
219 int i,k;
220 float lsp__hz[order];
221 float dlsp_[order];
222 const float *cb;
223
224 for(i=0; i<order; i++) {
225
226 k = lsp_cbd[i].k;
227 cb = lsp_cbd[i].cb;
228 dlsp_[i] = cb[indexes[i]*k];
229
230 if (i)
231 lsp__hz[i] = lsp__hz[i-1] + dlsp_[i];
232 else
233 lsp__hz[0] = dlsp_[0];
234
235 lsp_[i] = (PI/4000.0)*lsp__hz[i];
236
237 //printf("%d dlsp_ %3.2f lsp_ %3.2f\n", i, dlsp_[i], lsp__hz[i]);
238 }
239 110
111 return (besti);
240} 112}
241 113
242#ifdef __EXPERIMENTAL__
243/*---------------------------------------------------------------------------*\ 114/*---------------------------------------------------------------------------*\
244 115
245 lspvq_quantise 116 encode_lspds_scalar()
246 117
247 Vector LSP quantiser. 118 Scalar/VQ LSP difference-in-frequency quantiser.
248 119
249\*---------------------------------------------------------------------------*/ 120\*---------------------------------------------------------------------------*/
250 121
251void lspvq_quantise( 122void encode_lspds_scalar(int indexes[], float lsp[], int order) {
252 float lsp[], 123 int i, k, m;
253 float lsp_[], 124 float lsp_hz[order];
254 int order 125 float lsp__hz[order];
255) 126 float dlsp[order];
256{ 127 float dlsp_[order];
257 int i,k,m,ncb, nlsp; 128 float wt[order];
258 float wt[order], lsp_hz[order]; 129 const float *cb;
259 const float *cb; 130 float se;
260 float se; 131
261 int index; 132 for (i = 0; i < order; i++) {
262 133 wt[i] = 1.0;
263 for(i=0; i<order; i++) { 134 }
264 wt[i] = 1.0;
265 lsp_hz[i] = 4000.0*lsp[i]/PI;
266 }
267
268 /* scalar quantise LSPs 1,2,3,4 */
269
270 /* simple uniform scalar quantisers */
271
272 for(i=0; i<4; i++) {
273 k = lsp_cb[i].k;
274 m = lsp_cb[i].m;
275 cb = lsp_cb[i].cb;
276 index = quantise(cb, &lsp_hz[i], wt, k, m, &se);
277 lsp_[i] = cb[index*k]*PI/4000.0;
278 }
279
280 //#define WGHT
281#ifdef WGHT
282 for(i=4; i<9; i++) {
283 wt[i] = 1.0/(lsp[i]-lsp[i-1]) + 1.0/(lsp[i+1]-lsp[i]);
284 //printf("wt[%d] = %f\n", i, wt[i]);
285 }
286 wt[9] = 1.0/(lsp[i]-lsp[i-1]);
287#endif
288
289 /* VQ LSPs 5,6,7,8,9,10 */
290
291 ncb = 4;
292 nlsp = 4;
293 k = lsp_cbjnd[ncb].k;
294 m = lsp_cbjnd[ncb].m;
295 cb = lsp_cbjnd[ncb].cb;
296 index = quantise(cb, &lsp_hz[nlsp], &wt[nlsp], k, m, &se);
297 for(i=4; i<order; i++) {
298 lsp_[i] = cb[index*k+i-4]*(PI/4000.0);
299 //printf("%4.f (%4.f) ", lsp_hz[i], cb[index*k+i-4]);
300 }
301}
302
303/*---------------------------------------------------------------------------*\
304
305 lspjnd_quantise
306 135
307 Experimental JND LSP quantiser. 136 /* convert from radians to Hz so we can use human readable
137 frequencies */
308 138
309\*---------------------------------------------------------------------------*/ 139 for (i = 0; i < order; i++) lsp_hz[i] = (4000.0 / PI) * lsp[i];
310 140
311void lspjnd_quantise(float lsps[], float lsps_[], int order) 141 wt[0] = 1.0;
312{ 142 for (i = 0; i < order; i++) {
313 int i,k,m; 143 /* find difference from previous quantised lsp */
314 float wt[order], lsps_hz[order];
315 const float *cb;
316 float se = 0.0;
317 int index;
318
319 for(i=0; i<order; i++) {
320 wt[i] = 1.0;
321 }
322
323 /* convert to Hz */
324 144
325 for(i=0; i<order; i++) { 145 if (i)
326 lsps_hz[i] = lsps[i]*(4000.0/PI); 146 dlsp[i] = lsp_hz[i] - lsp__hz[i - 1];
327 lsps_[i] = lsps[i]; 147 else
328 } 148 dlsp[0] = lsp_hz[0];
329 149
330 /* simple uniform scalar quantisers */ 150 k = lsp_cbd[i].k;
151 m = lsp_cbd[i].m;
152 cb = lsp_cbd[i].cb;
153 indexes[i] = quantise(cb, &dlsp[i], wt, k, m, &se);
154 dlsp_[i] = cb[indexes[i] * k];
331 155
332 for(i=0; i<4; i++) { 156 if (i)
333 k = lsp_cbjnd[i].k; 157 lsp__hz[i] = lsp__hz[i - 1] + dlsp_[i];
334 m = lsp_cbjnd[i].m; 158 else
335 cb = lsp_cbjnd[i].cb; 159 lsp__hz[0] = dlsp_[0];
336 index = quantise(cb, &lsps_hz[i], wt, k, m, &se); 160 }
337 lsps_[i] = cb[index*k]*(PI/4000.0);
338 }
339
340 /* VQ LSPs 5,6,7,8,9,10 */
341
342 k = lsp_cbjnd[4].k;
343 m = lsp_cbjnd[4].m;
344 cb = lsp_cbjnd[4].cb;
345 index = quantise(cb, &lsps_hz[4], &wt[4], k, m, &se);
346 //printf("k = %d m = %d c[0] %f cb[k] %f\n", k,m,cb[0],cb[k]);
347 //printf("index = %4d: ", index);
348 for(i=4; i<order; i++) {
349 lsps_[i] = cb[index*k+i-4]*(PI/4000.0);
350 //printf("%4.f (%4.f) ", lsps_hz[i], cb[index*k+i-4]);
351 }
352 //printf("\n");
353} 161}
354 162
355void compute_weights(const float *x, float *w, int ndim); 163void decode_lspds_scalar(float lsp_[], int indexes[], int order) {
356 164 int i, k;
357/*---------------------------------------------------------------------------*\ 165 float lsp__hz[order];
358 166 float dlsp_[order];
359 lspdt_quantise 167 const float *cb;
360
361 LSP difference in time quantiser. Split VQ, encoding LSPs 1-4 with
362 one VQ, and LSPs 5-10 with a second. Update of previous lsp memory
363 is done outside of this function to handle dT between 10 or 20ms
364 frames.
365
366 mode action
367 ------------------
368
369 LSPDT_ALL VQ LSPs 1-4 and 5-10
370 LSPDT_LOW Just VQ LSPs 1-4, for LSPs 5-10 just copy previous
371 LSPDT_HIGH Just VQ LSPs 5-10, for LSPs 1-4 just copy previous
372
373\*---------------------------------------------------------------------------*/
374
375void lspdt_quantise(float lsps[], float lsps_[], float lsps__prev[], int mode)
376{
377 int i;
378 float wt[LPC_ORD];
379 float lsps_dt[LPC_ORD];
380#ifdef TRY_LSPDT_VQ
381 int k,m;
382 int index;
383 const float *cb;
384 float se = 0.0;
385#endif // TRY_LSPDT_VQ
386
387 //compute_weights(lsps, wt, LPC_ORD);
388 for(i=0; i<LPC_ORD; i++) {
389 wt[i] = 1.0;
390 }
391
392 //compute_weights(lsps, wt, LPC_ORD );
393 168
394 for(i=0; i<LPC_ORD; i++) { 169 for (i = 0; i < order; i++) {
395 lsps_dt[i] = lsps[i] - lsps__prev[i]; 170 k = lsp_cbd[i].k;
396 lsps_[i] = lsps__prev[i]; 171 cb = lsp_cbd[i].cb;
397 } 172 dlsp_[i] = cb[indexes[i] * k];
398 173
399 //#define TRY_LSPDT_VQ 174 if (i)
400#ifdef TRY_LSPDT_VQ 175 lsp__hz[i] = lsp__hz[i - 1] + dlsp_[i];
401 /* this actually improves speech a bit, but 40ms updates works surprsingly well.... */ 176 else
402 k = lsp_cbdt[0].k; 177 lsp__hz[0] = dlsp_[0];
403 m = lsp_cbdt[0].m;
404 cb = lsp_cbdt[0].cb;
405 index = quantise(cb, lsps_dt, wt, k, m, &se);
406 for(i=0; i<LPC_ORD; i++) {
407 lsps_[i] += cb[index*k + i];
408 }
409#endif
410 178
179 lsp_[i] = (PI / 4000.0) * lsp__hz[i];
180 }
411} 181}
412#endif
413 182
414#define MIN(a,b) ((a)<(b)?(a):(b)) 183#define MIN(a, b) ((a) < (b) ? (a) : (b))
415#define MAX_ENTRIES 16384 184#define MAX_ENTRIES 16384
416 185
417void compute_weights(const float *x, float *w, int ndim) 186void compute_weights(const float *x, float *w, int ndim) {
418{
419 int i; 187 int i;
420 w[0] = MIN(x[0], x[1]-x[0]); 188 w[0] = MIN(x[0], x[1] - x[0]);
421 for (i=1;i<ndim-1;i++) 189 for (i = 1; i < ndim - 1; i++) w[i] = MIN(x[i] - x[i - 1], x[i + 1] - x[i]);
422 w[i] = MIN(x[i]-x[i-1], x[i+1]-x[i]); 190 w[ndim - 1] = MIN(x[ndim - 1] - x[ndim - 2], PI - x[ndim - 1]);
423 w[ndim-1] = MIN(x[ndim-1]-x[ndim-2], PI-x[ndim-1]);
424
425 for (i=0;i<ndim;i++)
426 w[i] = 1./(.01+w[i]);
427 //w[0]*=3;
428 //w[1]*=2;
429}
430
431#ifdef __EXPERIMENTAL__
432/* LSP weight calculation ported from m-file function kindly submitted
433 by Anssi, OH3GDD */
434 191
435void compute_weights_anssi_mode2(const float *x, float *w, int ndim) 192 for (i = 0; i < ndim; i++) w[i] = 1. / (.01 + w[i]);
436{
437 int i;
438 float d[LPC_ORD];
439
440 assert(ndim == LPC_ORD);
441
442 for(i=0; i<LPC_ORD; i++)
443 d[i] = 1.0;
444
445 d[0] = x[1];
446 for (i=1; i<LPC_ORD-1; i++)
447 d[i] = x[i+1] - x[i-1];
448 d[LPC_ORD-1] = PI - x[8];
449 for (i=0; i<LPC_ORD; i++) {
450 if (x[i]<((400.0/4000.0)*PI))
451 w[i]=5.0/(0.01+d[i]);
452 else if (x[i]<((700.0/4000.0)*PI))
453 w[i]=4.0/(0.01+d[i]);
454 else if (x[i]<((1200.0/4000.0)*PI))
455 w[i]=3.0/(0.01+d[i]);
456 else if (x[i]<((2000.0/4000.0)*PI))
457 w[i]=2.0/(0.01+d[i]);
458 else
459 w[i]=1.0/(0.01+d[i]);
460
461 w[i]=powf(w[i]+0.3, 0.66);
462 }
463} 193}
464#endif
465 194
466int find_nearest(const float *codebook, int nb_entries, float *x, int ndim) 195int find_nearest(const float *codebook, int nb_entries, float *x, int ndim) {
467{
468 int i, j; 196 int i, j;
469 float min_dist = 1e15; 197 float min_dist = 1e15;
470 int nearest = 0; 198 int nearest = 0;
471 199
472 for (i=0;i<nb_entries;i++) 200 for (i = 0; i < nb_entries; i++) {
473 { 201 float dist = 0;
474 float dist=0; 202 for (j = 0; j < ndim; j++)
475 for (j=0;j<ndim;j++) 203 dist += (x[j] - codebook[i * ndim + j]) * (x[j] - codebook[i * ndim + j]);
476 dist += (x[j]-codebook[i*ndim+j])*(x[j]-codebook[i*ndim+j]); 204 if (dist < min_dist) {
477 if (dist<min_dist)
478 {
479 min_dist = dist; 205 min_dist = dist;
480 nearest = i; 206 nearest = i;
481 } 207 }
@@ -483,19 +209,18 @@ int find_nearest(const float *codebook, int nb_entries, float *x, int ndim)
483 return nearest; 209 return nearest;
484} 210}
485 211
486int find_nearest_weighted(const float *codebook, int nb_entries, float *x, const float *w, int ndim) 212int find_nearest_weighted(const float *codebook, int nb_entries, float *x,
487{ 213 const float *w, int ndim) {
488 int i, j; 214 int i, j;
489 float min_dist = 1e15; 215 float min_dist = 1e15;
490 int nearest = 0; 216 int nearest = 0;
491 217
492 for (i=0;i<nb_entries;i++) 218 for (i = 0; i < nb_entries; i++) {
493 { 219 float dist = 0;
494 float dist=0; 220 for (j = 0; j < ndim; j++)
495 for (j=0;j<ndim;j++) 221 dist += w[j] * (x[j] - codebook[i * ndim + j]) *
496 dist += w[j]*(x[j]-codebook[i*ndim+j])*(x[j]-codebook[i*ndim+j]); 222 (x[j] - codebook[i * ndim + j]);
497 if (dist<min_dist) 223 if (dist < min_dist) {
498 {
499 min_dist = dist; 224 min_dist = dist;
500 nearest = i; 225 nearest = i;
501 } 226 }
@@ -503,212 +228,74 @@ int find_nearest_weighted(const float *codebook, int nb_entries, float *x, const
503 return nearest; 228 return nearest;
504} 229}
505 230
506void lspjvm_quantise(float *x, float *xq, int order) 231void lspjmv_quantise(float *x, float *xq, int order) {
507{
508 int i, n1, n2, n3; 232 int i, n1, n2, n3;
509 float err[order], err2[order], err3[order]; 233 float err[order], err2[order], err3[order];
510 float w[order], w2[order], w3[order]; 234 float w[order], w2[order], w3[order];
511 const float *codebook1 = lsp_cbjvm[0].cb; 235 const float *codebook1 = lsp_cbjmv[0].cb;
512 const float *codebook2 = lsp_cbjvm[1].cb; 236 const float *codebook2 = lsp_cbjmv[1].cb;
513 const float *codebook3 = lsp_cbjvm[2].cb; 237 const float *codebook3 = lsp_cbjmv[2].cb;
514 238
515 w[0] = MIN(x[0], x[1]-x[0]); 239 w[0] = MIN(x[0], x[1] - x[0]);
516 for (i=1;i<order-1;i++) 240 for (i = 1; i < order - 1; i++) w[i] = MIN(x[i] - x[i - 1], x[i + 1] - x[i]);
517 w[i] = MIN(x[i]-x[i-1], x[i+1]-x[i]); 241 w[order - 1] = MIN(x[order - 1] - x[order - 2], PI - x[order - 1]);
518 w[order-1] = MIN(x[order-1]-x[order-2], PI-x[order-1]);
519 242
520 compute_weights(x, w, order); 243 compute_weights(x, w, order);
521 244
522 n1 = find_nearest(codebook1, lsp_cbjvm[0].m, x, order); 245 n1 = find_nearest(codebook1, lsp_cbjmv[0].m, x, order);
523 246
524 for (i=0;i<order;i++) 247 for (i = 0; i < order; i++) {
525 { 248 xq[i] = codebook1[order * n1 + i];
526 xq[i] = codebook1[order*n1+i];
527 err[i] = x[i] - xq[i]; 249 err[i] = x[i] - xq[i];
528 } 250 }
529 for (i=0;i<order/2;i++) 251 for (i = 0; i < order / 2; i++) {
530 { 252 err2[i] = err[2 * i];
531 err2[i] = err[2*i]; 253 err3[i] = err[2 * i + 1];
532 err3[i] = err[2*i+1]; 254 w2[i] = w[2 * i];
533 w2[i] = w[2*i]; 255 w3[i] = w[2 * i + 1];
534 w3[i] = w[2*i+1];
535 } 256 }
536 n2 = find_nearest_weighted(codebook2, lsp_cbjvm[1].m, err2, w2, order/2); 257 n2 = find_nearest_weighted(codebook2, lsp_cbjmv[1].m, err2, w2, order / 2);
537 n3 = find_nearest_weighted(codebook3, lsp_cbjvm[2].m, err3, w3, order/2); 258 n3 = find_nearest_weighted(codebook3, lsp_cbjmv[2].m, err3, w3, order / 2);
538 259
539 for (i=0;i<order/2;i++) 260 for (i = 0; i < order / 2; i++) {
540 { 261 xq[2 * i] += codebook2[order * n2 / 2 + i];
541 xq[2*i] += codebook2[order*n2/2+i]; 262 xq[2 * i + 1] += codebook3[order * n3 / 2 + i];
542 xq[2*i+1] += codebook3[order*n3/2+i];
543 } 263 }
544} 264}
545 265
266int check_lsp_order(float lsp[], int order) {
267 int i;
268 float tmp;
269 int swaps = 0;
270
271 for (i = 1; i < order; i++)
272 if (lsp[i] < lsp[i - 1]) {
273 // fprintf(stderr, "swap %d\n",i);
274 swaps++;
275 tmp = lsp[i - 1];
276 lsp[i - 1] = lsp[i] - 0.1;
277 lsp[i] = tmp + 0.1;
278 i = 1; /* start check again, as swap may have caused out of order */
279 }
546 280
547#ifndef CORTEX_M4 281 return swaps;
548/* simple (non mbest) 6th order LSP MEL VQ quantiser. Returns MSE of result */
549
550float lspmelvq_quantise(float *x, float *xq, int order)
551{
552 int i, n1, n2, n3;
553 float err[order];
554 const float *codebook1 = lspmelvq_cb[0].cb;
555 const float *codebook2 = lspmelvq_cb[1].cb;
556 const float *codebook3 = lspmelvq_cb[2].cb;
557 float tmp[order];
558 float mse;
559
560 assert(order == lspmelvq_cb[0].k);
561
562 n1 = find_nearest(codebook1, lspmelvq_cb[0].m, x, order);
563
564 for (i=0; i<order; i++) {
565 tmp[i] = codebook1[order*n1+i];
566 err[i] = x[i] - tmp[i];
567 }
568
569 n2 = find_nearest(codebook2, lspmelvq_cb[1].m, err, order);
570
571 for (i=0; i<order; i++) {
572 tmp[i] += codebook2[order*n2+i];
573 err[i] = x[i] - tmp[i];
574 }
575
576 n3 = find_nearest(codebook3, lspmelvq_cb[2].m, err, order);
577
578 mse = 0.0;
579 for (i=0; i<order; i++) {
580 tmp[i] += codebook3[order*n3+i];
581 err[i] = x[i] - tmp[i];
582 mse += err[i]*err[i];
583 }
584
585 for (i=0; i<order; i++) {
586 xq[i] = tmp[i];
587 }
588
589 return mse;
590}
591
592/* 3 stage VQ LSP quantiser useing mbest search. Design and guidance kindly submitted by Anssi, OH3GDD */
593
594float lspmelvq_mbest_encode(int *indexes, float *x, float *xq, int ndim, int mbest_entries)
595{
596 int i, j, n1, n2, n3;
597 const float *codebook1 = lspmelvq_cb[0].cb;
598 const float *codebook2 = lspmelvq_cb[1].cb;
599 const float *codebook3 = lspmelvq_cb[2].cb;
600 struct MBEST *mbest_stage1, *mbest_stage2, *mbest_stage3;
601 float target[ndim];
602 float w[ndim];
603 int index[MBEST_STAGES];
604 float mse, tmp;
605
606 for(i=0; i<ndim; i++)
607 w[i] = 1.0;
608
609 mbest_stage1 = mbest_create(mbest_entries);
610 mbest_stage2 = mbest_create(mbest_entries);
611 mbest_stage3 = mbest_create(mbest_entries);
612 for(i=0; i<MBEST_STAGES; i++)
613 index[i] = 0;
614
615 /* Stage 1 */
616
617 mbest_search(codebook1, x, w, ndim, lspmelvq_cb[0].m, mbest_stage1, index);
618 MBEST_PRINT("Stage 1:", mbest_stage1);
619
620 /* Stage 2 */
621
622 for (j=0; j<mbest_entries; j++) {
623 index[1] = n1 = mbest_stage1->list[j].index[0];
624 for(i=0; i<ndim; i++)
625 target[i] = x[i] - codebook1[ndim*n1+i];
626 mbest_search(codebook2, target, w, ndim, lspmelvq_cb[1].m, mbest_stage2, index);
627 }
628 MBEST_PRINT("Stage 2:", mbest_stage2);
629
630 /* Stage 3 */
631
632 for (j=0; j<mbest_entries; j++) {
633 index[2] = n1 = mbest_stage2->list[j].index[1];
634 index[1] = n2 = mbest_stage2->list[j].index[0];
635 for(i=0; i<ndim; i++)
636 target[i] = x[i] - codebook1[ndim*n1+i] - codebook2[ndim*n2+i];
637 mbest_search(codebook3, target, w, ndim, lspmelvq_cb[2].m, mbest_stage3, index);
638 }
639 MBEST_PRINT("Stage 3:", mbest_stage3);
640
641 n1 = mbest_stage3->list[0].index[2];
642 n2 = mbest_stage3->list[0].index[1];
643 n3 = mbest_stage3->list[0].index[0];
644 mse = 0.0;
645 for (i=0;i<ndim;i++) {
646 tmp = codebook1[ndim*n1+i] + codebook2[ndim*n2+i] + codebook3[ndim*n3+i];
647 mse += (x[i]-tmp)*(x[i]-tmp);
648 xq[i] = tmp;
649 }
650
651 mbest_destroy(mbest_stage1);
652 mbest_destroy(mbest_stage2);
653 mbest_destroy(mbest_stage3);
654
655 indexes[0] = n1; indexes[1] = n2; indexes[2] = n3;
656
657 return mse;
658}
659
660
661void lspmelvq_decode(int *indexes, float *xq, int ndim)
662{
663 int i, n1, n2, n3;
664 const float *codebook1 = lspmelvq_cb[0].cb;
665 const float *codebook2 = lspmelvq_cb[1].cb;
666 const float *codebook3 = lspmelvq_cb[2].cb;
667
668 n1 = indexes[0]; n2 = indexes[1]; n3 = indexes[2];
669 for (i=0;i<ndim;i++) {
670 xq[i] = codebook1[ndim*n1+i] + codebook2[ndim*n2+i] + codebook3[ndim*n3+i];
671 }
672}
673#endif
674
675
676int check_lsp_order(float lsp[], int order)
677{
678 int i;
679 float tmp;
680 int swaps = 0;
681
682 for(i=1; i<order; i++)
683 if (lsp[i] < lsp[i-1]) {
684 //fprintf(stderr, "swap %d\n",i);
685 swaps++;
686 tmp = lsp[i-1];
687 lsp[i-1] = lsp[i]-0.1;
688 lsp[i] = tmp+0.1;
689 i = 1; /* start check again, as swap may have caused out of order */
690 }
691
692 return swaps;
693} 282}
694 283
695void force_min_lsp_dist(float lsp[], int order) 284void force_min_lsp_dist(float lsp[], int order) {
696{ 285 int i;
697 int i;
698 286
699 for(i=1; i<order; i++) 287 for (i = 1; i < order; i++)
700 if ((lsp[i]-lsp[i-1]) < 0.01) { 288 if ((lsp[i] - lsp[i - 1]) < 0.01) {
701 lsp[i] += 0.01; 289 lsp[i] += 0.01;
702 } 290 }
703} 291}
704 292
705
706/*---------------------------------------------------------------------------*\ 293/*---------------------------------------------------------------------------*\
707 294
708 lpc_post_filter() 295 lpc_post_filter()
709 296
710 Applies a post filter to the LPC synthesis filter power spectrum 297 Applies a post filter to the LPC synthesis filter power spectrum
711 Pw, which supresses the inter-formant energy. 298 Pw, which suppresses the inter-formant energy.
712 299
713 The algorithm is from p267 (Section 8.6) of "Digital Speech", 300 The algorithm is from p267 (Section 8.6) of "Digital Speech",
714 edited by A.M. Kondoz, 1994 published by Wiley and Sons. Chapter 8 301 edited by A.M. Kondoz, 1994 published by Wiley and Sons. Chapter 8
@@ -722,7 +309,7 @@ void force_min_lsp_dist(float lsp[], int order)
722 it should be possible to implement this more efficiently in the 309 it should be possible to implement this more efficiently in the
723 time domain. Just not sure how to handle relative time delays 310 time domain. Just not sure how to handle relative time delays
724 between the synthesis stage and updating these coeffs. A smaller 311 between the synthesis stage and updating these coeffs. A smaller
725 FFT size might also be accetable to save CPU. 312 FFT size might also be acceptable to save CPU.
726 313
727 TODO: 314 TODO:
728 [ ] sync var names between Octave and C version 315 [ ] sync var names between Octave and C version
@@ -733,104 +320,97 @@ void force_min_lsp_dist(float lsp[], int order)
733\*---------------------------------------------------------------------------*/ 320\*---------------------------------------------------------------------------*/
734 321
735void lpc_post_filter(codec2_fftr_cfg fftr_fwd_cfg, float Pw[], float ak[], 322void lpc_post_filter(codec2_fftr_cfg fftr_fwd_cfg, float Pw[], float ak[],
736 int order, int dump, float beta, float gamma, int bass_boost, float E) 323 int order, int dump, float beta, float gamma,
737{ 324 int bass_boost, float E) {
738 int i; 325 int i;
739 float x[FFT_ENC]; /* input to FFTs */ 326 float x[FFT_ENC]; /* input to FFTs */
740 COMP Ww[FFT_ENC/2+1]; /* weighting spectrum */ 327 COMP Ww[FFT_ENC / 2 + 1]; /* weighting spectrum */
741 float Rw[FFT_ENC/2+1]; /* R = WA */ 328 float Rw[FFT_ENC / 2 + 1]; /* R = WA */
742 float e_before, e_after, gain; 329 float e_before, e_after, gain;
743 float Pfw; 330 float Pfw;
744 float max_Rw, min_Rw; 331 float max_Rw, min_Rw;
745 float coeff; 332 float coeff;
746 PROFILE_VAR(tstart, tfft1, taw, tfft2, tww, tr); 333 PROFILE_VAR(tstart, tfft1, taw, tfft2, tww, tr);
747
748 PROFILE_SAMPLE(tstart);
749
750 /* Determine weighting filter spectrum W(exp(jw)) ---------------*/
751
752 for(i=0; i<FFT_ENC; i++) {
753 x[i] = 0.0;
754 }
755 334
756 x[0] = ak[0]; 335 PROFILE_SAMPLE(tstart);
757 coeff = gamma;
758 for(i=1; i<=order; i++) {
759 x[i] = ak[i] * coeff;
760 coeff *= gamma;
761 }
762 codec2_fftr(fftr_fwd_cfg, x, Ww);
763 336
764 PROFILE_SAMPLE_AND_LOG(tfft2, taw, " fft2"); 337 /* Determine weighting filter spectrum W(exp(jw)) ---------------*/
765 338
766 for(i=0; i<FFT_ENC/2; i++) { 339 for (i = 0; i < FFT_ENC; i++) {
767 Ww[i].real = Ww[i].real*Ww[i].real + Ww[i].imag*Ww[i].imag; 340 x[i] = 0.0;
768 } 341 }
769 342
770 PROFILE_SAMPLE_AND_LOG(tww, tfft2, " Ww"); 343 x[0] = ak[0];
344 coeff = gamma;
345 for (i = 1; i <= order; i++) {
346 x[i] = ak[i] * coeff;
347 coeff *= gamma;
348 }
349 codec2_fftr(fftr_fwd_cfg, x, Ww);
771 350
772 /* Determined combined filter R = WA ---------------------------*/ 351 PROFILE_SAMPLE_AND_LOG(tfft2, taw, " fft2");
773 352
774 max_Rw = 0.0; min_Rw = 1E32; 353 for (i = 0; i < FFT_ENC / 2; i++) {
775 for(i=0; i<FFT_ENC/2; i++) { 354 Ww[i].real = Ww[i].real * Ww[i].real + Ww[i].imag * Ww[i].imag;
776 Rw[i] = sqrtf(Ww[i].real * Pw[i]); 355 }
777 if (Rw[i] > max_Rw)
778 max_Rw = Rw[i];
779 if (Rw[i] < min_Rw)
780 min_Rw = Rw[i];
781 356
782 } 357 PROFILE_SAMPLE_AND_LOG(tww, tfft2, " Ww");
783 358
784 PROFILE_SAMPLE_AND_LOG(tr, tww, " R"); 359 /* Determined combined filter R = WA ---------------------------*/
785 360
786 #ifdef DUMP 361 max_Rw = 0.0;
787 if (dump) 362 min_Rw = 1E32;
788 dump_Rw(Rw); 363 for (i = 0; i < FFT_ENC / 2; i++) {
789 #endif 364 Rw[i] = sqrtf(Ww[i].real * Pw[i]);
365 if (Rw[i] > max_Rw) max_Rw = Rw[i];
366 if (Rw[i] < min_Rw) min_Rw = Rw[i];
367 }
790 368
791 /* create post filter mag spectrum and apply ------------------*/ 369 PROFILE_SAMPLE_AND_LOG(tr, tww, " R");
792 370
793 /* measure energy before post filtering */ 371#ifdef DUMP
372 if (dump) dump_Rw(Rw);
373#endif
794 374
795 e_before = 1E-4; 375 /* create post filter mag spectrum and apply ------------------*/
796 for(i=0; i<FFT_ENC/2; i++)
797 e_before += Pw[i];
798 376
799 /* apply post filter and measure energy */ 377 /* measure energy before post filtering */
800 378
801 #ifdef DUMP 379 e_before = 1E-4;
802 if (dump) 380 for (i = 0; i < FFT_ENC / 2; i++) e_before += Pw[i];
803 dump_Pwb(Pw);
804 #endif
805 381
382 /* apply post filter and measure energy */
806 383
807 e_after = 1E-4; 384#ifdef DUMP
808 for(i=0; i<FFT_ENC/2; i++) { 385 if (dump) dump_Pwb(Pw);
809 Pfw = powf(Rw[i], beta); 386#endif
810 Pw[i] *= Pfw * Pfw;
811 e_after += Pw[i];
812 }
813 gain = e_before/e_after;
814 387
815 /* apply gain factor to normalise energy, and LPC Energy */ 388 e_after = 1E-4;
389 for (i = 0; i < FFT_ENC / 2; i++) {
390 Pfw = powf(Rw[i], beta);
391 Pw[i] *= Pfw * Pfw;
392 e_after += Pw[i];
393 }
394 gain = e_before / e_after;
816 395
817 gain *= E; 396 /* apply gain factor to normalise energy, and LPC Energy */
818 for(i=0; i<FFT_ENC/2; i++) { 397
819 Pw[i] *= gain; 398 gain *= E;
820 } 399 for (i = 0; i < FFT_ENC / 2; i++) {
400 Pw[i] *= gain;
401 }
821 402
822 if (bass_boost) { 403 if (bass_boost) {
823 /* add 3dB to first 1 kHz to account for LP effect of PF */ 404 /* add 3dB to first 1 kHz to account for LP effect of PF */
824 405
825 for(i=0; i<FFT_ENC/8; i++) { 406 for (i = 0; i < FFT_ENC / 8; i++) {
826 Pw[i] *= 1.4*1.4; 407 Pw[i] *= 1.4 * 1.4;
827 }
828 } 408 }
409 }
829 410
830 PROFILE_SAMPLE_AND_LOG2(tr, " filt"); 411 PROFILE_SAMPLE_AND_LOG2(tr, " filt");
831} 412}
832 413
833
834/*---------------------------------------------------------------------------*\ 414/*---------------------------------------------------------------------------*\
835 415
836 aks_to_M2() 416 aks_to_M2()
@@ -841,134 +421,125 @@ void lpc_post_filter(codec2_fftr_cfg fftr_fwd_cfg, float Pw[], float ak[],
841 421
842\*---------------------------------------------------------------------------*/ 422\*---------------------------------------------------------------------------*/
843 423
844void aks_to_M2( 424void aks_to_M2(codec2_fftr_cfg fftr_fwd_cfg, float ak[], /* LPC's */
845 codec2_fftr_cfg fftr_fwd_cfg, 425 int order,
846 float ak[], /* LPC's */ 426 MODEL *model, /* sinusoidal model parameters for this frame */
847 int order, 427 float E, /* energy term */
848 MODEL *model, /* sinusoidal model parameters for this frame */ 428 float *snr, /* signal to noise ratio for this frame in dB */
849 float E, /* energy term */ 429 int dump, /* true to dump sample to dump file */
850 float *snr, /* signal to noise ratio for this frame in dB */ 430 int sim_pf, /* true to simulate a post filter */
851 int dump, /* true to dump sample to dump file */ 431 int pf, /* true to enable actual LPC post filter */
852 int sim_pf, /* true to simulate a post filter */ 432 int bass_boost, /* enable LPC filter 0-1kHz 3dB boost */
853 int pf, /* true to enable actual LPC post filter */ 433 float beta, float gamma, /* LPC post filter parameters */
854 int bass_boost, /* enable LPC filter 0-1kHz 3dB boost */ 434 COMP Aw[] /* output power spectrum */
855 float beta, 435) {
856 float gamma, /* LPC post filter parameters */ 436 int i, m; /* loop variables */
857 COMP Aw[] /* output power spectrum */ 437 int am, bm; /* limits of current band */
858) 438 float r; /* no. rads/bin */
859{ 439 float Em; /* energy in band */
860 int i,m; /* loop variables */ 440 float Am; /* spectral amplitude sample */
861 int am,bm; /* limits of current band */
862 float r; /* no. rads/bin */
863 float Em; /* energy in band */
864 float Am; /* spectral amplitude sample */
865 float signal, noise; 441 float signal, noise;
866 PROFILE_VAR(tstart, tfft, tpw, tpf); 442 PROFILE_VAR(tstart, tfft, tpw, tpf);
867 443
868 PROFILE_SAMPLE(tstart); 444 PROFILE_SAMPLE(tstart);
869 445
870 r = TWO_PI/(FFT_ENC); 446 r = TWO_PI / (FFT_ENC);
871 447
872 /* Determine DFT of A(exp(jw)) --------------------------------------------*/ 448 /* Determine DFT of A(exp(jw)) --------------------------------------------*/
873 { 449 {
874 float a[FFT_ENC]; /* input to FFT for power spectrum */ 450 float a[FFT_ENC]; /* input to FFT for power spectrum */
875 451
876 for(i=0; i<FFT_ENC; i++) { 452 for (i = 0; i < FFT_ENC; i++) {
877 a[i] = 0.0; 453 a[i] = 0.0;
878 } 454 }
879 455
880 for(i=0; i<=order; i++) 456 for (i = 0; i <= order; i++) a[i] = ak[i];
881 a[i] = ak[i]; 457 codec2_fftr(fftr_fwd_cfg, a, Aw);
882 codec2_fftr(fftr_fwd_cfg, a, Aw);
883 } 458 }
884 PROFILE_SAMPLE_AND_LOG(tfft, tstart, " fft"); 459 PROFILE_SAMPLE_AND_LOG(tfft, tstart, " fft");
885 460
886 /* Determine power spectrum P(w) = E/(A(exp(jw))^2 ------------------------*/ 461 /* Determine power spectrum P(w) = E/(A(exp(jw))^2 ------------------------*/
887 462
888 float Pw[FFT_ENC/2]; 463 float Pw[FFT_ENC / 2];
889 464
890#ifndef FDV_ARM_MATH 465#ifndef FDV_ARM_MATH
891 for(i=0; i<FFT_ENC/2; i++) { 466 for (i = 0; i < FFT_ENC / 2; i++) {
892 Pw[i] = 1.0/(Aw[i].real*Aw[i].real + Aw[i].imag*Aw[i].imag + 1E-6); 467 Pw[i] = 1.0 / (Aw[i].real * Aw[i].real + Aw[i].imag * Aw[i].imag + 1E-6);
893 } 468 }
894#else 469#else
895 // this difference may seem strange, but the gcc for STM32F4 generates almost 5 times 470 // this difference may seem strange, but the gcc for STM32F4 generates almost
896 // faster code with the two loops: 1120 ms -> 242 ms 471 // 5 times faster code with the two loops: 1120 ms -> 242 ms so please leave
897 // so please leave it as is or improve further 472 // it as is or improve further since this code is called 4 times it results in
898 // since this code is called 4 times it results in almost 4ms gain (21ms -> 17ms per audio frame decode @ 1300 ) 473 // almost 4ms gain (21ms -> 17ms per audio frame decode @ 1300 )
899 474
900 for(i=0; i<FFT_ENC/2; i++) 475 for (i = 0; i < FFT_ENC / 2; i++) {
901 { 476 Pw[i] = Aw[i].real * Aw[i].real + Aw[i].imag * Aw[i].imag + 1E-6;
902 Pw[i] = Aw[i].real * Aw[i].real + Aw[i].imag * Aw[i].imag + 1E-6;
903 } 477 }
904 for(i=0; i<FFT_ENC/2; i++) { 478 for (i = 0; i < FFT_ENC / 2; i++) {
905 Pw[i] = 1.0/(Pw[i]); 479 Pw[i] = 1.0 / (Pw[i]);
906 } 480 }
907#endif 481#endif
908 482
909 PROFILE_SAMPLE_AND_LOG(tpw, tfft, " Pw"); 483 PROFILE_SAMPLE_AND_LOG(tpw, tfft, " Pw");
910 484
911 if (pf) 485 if (pf)
912 lpc_post_filter(fftr_fwd_cfg, Pw, ak, order, dump, beta, gamma, bass_boost, E); 486 lpc_post_filter(fftr_fwd_cfg, Pw, ak, order, dump, beta, gamma, bass_boost,
487 E);
913 else { 488 else {
914 for(i=0; i<FFT_ENC/2; i++) { 489 for (i = 0; i < FFT_ENC / 2; i++) {
915 Pw[i] *= E; 490 Pw[i] *= E;
916 } 491 }
917 } 492 }
918 493
919 PROFILE_SAMPLE_AND_LOG(tpf, tpw, " LPC post filter"); 494 PROFILE_SAMPLE_AND_LOG(tpf, tpw, " LPC post filter");
920 495
921 #ifdef DUMP 496#ifdef DUMP
922 if (dump) 497 if (dump) dump_Pw(Pw);
923 dump_Pw(Pw); 498#endif
924 #endif
925 499
926 /* Determine magnitudes from P(w) ----------------------------------------*/ 500 /* Determine magnitudes from P(w) ----------------------------------------*/
927 501
928 /* when used just by decoder {A} might be all zeroes so init signal 502 /* when used just by decoder {A} might be all zeroes so init signal
929 and noise to prevent log(0) errors */ 503 and noise to prevent log(0) errors */
930 504
931 signal = 1E-30; noise = 1E-32; 505 signal = 1E-30;
932 506 noise = 1E-32;
933 for(m=1; m<=model->L; m++) { 507
934 am = (int)((m - 0.5)*model->Wo/r + 0.5); 508 for (m = 1; m <= model->L; m++) {
935 bm = (int)((m + 0.5)*model->Wo/r + 0.5); 509 am = (int)((m - 0.5) * model->Wo / r + 0.5);
936 510 bm = (int)((m + 0.5) * model->Wo / r + 0.5);
937 // FIXME: With arm_rfft_fast_f32 we have to use this 511
938 // otherwise sometimes a to high bm is calculated 512 // FIXME: With arm_rfft_fast_f32 we have to use this
939 // which causes trouble later in the calculation 513 // otherwise sometimes a to high bm is calculated
940 // chain 514 // which causes trouble later in the calculation
941 // it seems for some reason model->Wo is calculated somewhat too high 515 // chain
942 if (bm>FFT_ENC/2) 516 // it seems for some reason model->Wo is calculated somewhat too high
943 { 517 if (bm > FFT_ENC / 2) {
944 bm = FFT_ENC/2; 518 bm = FFT_ENC / 2;
945 } 519 }
946 Em = 0.0; 520 Em = 0.0;
947 521
948 for(i=am; i<bm; i++) 522 for (i = am; i < bm; i++) Em += Pw[i];
949 Em += Pw[i]; 523 Am = sqrtf(Em);
950 Am = sqrtf(Em); 524
951 525 signal += model->A[m] * model->A[m];
952 signal += model->A[m]*model->A[m]; 526 noise += (model->A[m] - Am) * (model->A[m] - Am);
953 noise += (model->A[m] - Am)*(model->A[m] - Am); 527
954 528 /* This code significantly improves perf of LPC model, in
955 /* This code significantly improves perf of LPC model, in 529 particular when combined with phase0. The LPC spectrum tends
956 particular when combined with phase0. The LPC spectrum tends 530 to track just under the peaks of the spectral envelope, and
957 to track just under the peaks of the spectral envelope, and 531 just above nulls. This algorithm does the reverse to
958 just above nulls. This algorithm does the reverse to 532 compensate - raising the amplitudes of spectral peaks, while
959 compensate - raising the amplitudes of spectral peaks, while 533 attenuating the null. This enhances the formants, and
960 attenuating the null. This enhances the formants, and 534 suppresses the energy between formants. */
961 supresses the energy between formants. */ 535
962 536 if (sim_pf) {
963 if (sim_pf) { 537 if (Am > model->A[m]) Am *= 0.7;
964 if (Am > model->A[m]) 538 if (Am < model->A[m]) Am *= 1.4;
965 Am *= 0.7; 539 }
966 if (Am < model->A[m]) 540 model->A[m] = Am;
967 Am *= 1.4;
968 }
969 model->A[m] = Am;
970 } 541 }
971 *snr = 10.0*log10f(signal/noise); 542 *snr = 10.0 * log10f(signal / noise);
972 543
973 PROFILE_SAMPLE_AND_LOG2(tpf, " rec"); 544 PROFILE_SAMPLE_AND_LOG2(tpf, " rec");
974} 545}
@@ -983,19 +554,18 @@ void aks_to_M2(
983 554
984\*---------------------------------------------------------------------------*/ 555\*---------------------------------------------------------------------------*/
985 556
986int encode_Wo(C2CONST *c2const, float Wo, int bits) 557int encode_Wo(C2CONST *c2const, float Wo, int bits) {
987{ 558 int index, Wo_levels = 1 << bits;
988 int index, Wo_levels = 1<<bits; 559 float Wo_min = c2const->Wo_min;
989 float Wo_min = c2const->Wo_min; 560 float Wo_max = c2const->Wo_max;
990 float Wo_max = c2const->Wo_max; 561 float norm;
991 float norm;
992 562
993 norm = (Wo - Wo_min)/(Wo_max - Wo_min); 563 norm = (Wo - Wo_min) / (Wo_max - Wo_min);
994 index = floorf(Wo_levels * norm + 0.5); 564 index = floorf(Wo_levels * norm + 0.5);
995 if (index < 0 ) index = 0; 565 if (index < 0) index = 0;
996 if (index > (Wo_levels-1)) index = Wo_levels-1; 566 if (index > (Wo_levels - 1)) index = Wo_levels - 1;
997 567
998 return index; 568 return index;
999} 569}
1000 570
1001/*---------------------------------------------------------------------------*\ 571/*---------------------------------------------------------------------------*\
@@ -1008,18 +578,17 @@ int encode_Wo(C2CONST *c2const, float Wo, int bits)
1008 578
1009\*---------------------------------------------------------------------------*/ 579\*---------------------------------------------------------------------------*/
1010 580
1011float decode_Wo(C2CONST *c2const, int index, int bits) 581float decode_Wo(C2CONST *c2const, int index, int bits) {
1012{ 582 float Wo_min = c2const->Wo_min;
1013 float Wo_min = c2const->Wo_min; 583 float Wo_max = c2const->Wo_max;
1014 float Wo_max = c2const->Wo_max; 584 float step;
1015 float step; 585 float Wo;
1016 float Wo; 586 int Wo_levels = 1 << bits;
1017 int Wo_levels = 1<<bits;
1018 587
1019 step = (Wo_max - Wo_min)/Wo_levels; 588 step = (Wo_max - Wo_min) / Wo_levels;
1020 Wo = Wo_min + step*(index); 589 Wo = Wo_min + step * (index);
1021 590
1022 return Wo; 591 return Wo;
1023} 592}
1024 593
1025/*---------------------------------------------------------------------------*\ 594/*---------------------------------------------------------------------------*\
@@ -1032,19 +601,18 @@ float decode_Wo(C2CONST *c2const, int index, int bits)
1032 601
1033\*---------------------------------------------------------------------------*/ 602\*---------------------------------------------------------------------------*/
1034 603
1035int encode_log_Wo(C2CONST *c2const, float Wo, int bits) 604int encode_log_Wo(C2CONST *c2const, float Wo, int bits) {
1036{ 605 int index, Wo_levels = 1 << bits;
1037 int index, Wo_levels = 1<<bits; 606 float Wo_min = c2const->Wo_min;
1038 float Wo_min = c2const->Wo_min; 607 float Wo_max = c2const->Wo_max;
1039 float Wo_max = c2const->Wo_max; 608 float norm;
1040 float norm;
1041 609
1042 norm = (log10f(Wo) - log10f(Wo_min))/(log10f(Wo_max) - log10f(Wo_min)); 610 norm = (log10f(Wo) - log10f(Wo_min)) / (log10f(Wo_max) - log10f(Wo_min));
1043 index = floorf(Wo_levels * norm + 0.5); 611 index = floorf(Wo_levels * norm + 0.5);
1044 if (index < 0 ) index = 0; 612 if (index < 0) index = 0;
1045 if (index > (Wo_levels-1)) index = Wo_levels-1; 613 if (index > (Wo_levels - 1)) index = Wo_levels - 1;
1046 614
1047 return index; 615 return index;
1048} 616}
1049 617
1050/*---------------------------------------------------------------------------*\ 618/*---------------------------------------------------------------------------*\
@@ -1057,99 +625,18 @@ int encode_log_Wo(C2CONST *c2const, float Wo, int bits)
1057 625
1058\*---------------------------------------------------------------------------*/ 626\*---------------------------------------------------------------------------*/
1059 627
1060float decode_log_Wo(C2CONST *c2const, int index, int bits) 628float decode_log_Wo(C2CONST *c2const, int index, int bits) {
1061{ 629 float Wo_min = c2const->Wo_min;
1062 float Wo_min = c2const->Wo_min; 630 float Wo_max = c2const->Wo_max;
1063 float Wo_max = c2const->Wo_max; 631 float step;
1064 float step; 632 float Wo;
1065 float Wo; 633 int Wo_levels = 1 << bits;
1066 int Wo_levels = 1<<bits;
1067
1068 step = (log10f(Wo_max) - log10f(Wo_min))/Wo_levels;
1069 Wo = log10f(Wo_min) + step*(index);
1070
1071 return POW10F(Wo);
1072}
1073
1074#if 0
1075/*---------------------------------------------------------------------------*\
1076
1077 FUNCTION....: encode_Wo_dt()
1078 AUTHOR......: David Rowe
1079 DATE CREATED: 6 Nov 2011
1080
1081 Encodes Wo difference from last frame.
1082
1083\*---------------------------------------------------------------------------*/
1084
1085int encode_Wo_dt(C2CONST *c2const, float Wo, float prev_Wo)
1086{
1087 int index, mask, max_index, min_index;
1088 float Wo_min = c2const->Wo_min;
1089 float Wo_max = c2const->Wo_max;
1090 float norm;
1091
1092 norm = (Wo - prev_Wo)/(Wo_max - Wo_min);
1093 index = floorf(WO_LEVELS * norm + 0.5);
1094 //printf("ENC index: %d ", index);
1095
1096 /* hard limit */
1097
1098 max_index = (1 << (WO_DT_BITS-1)) - 1;
1099 min_index = - (max_index+1);
1100 if (index > max_index) index = max_index;
1101 if (index < min_index) index = min_index;
1102 //printf("max_index: %d min_index: %d hard index: %d ",
1103 // max_index, min_index, index);
1104
1105 /* mask so that only LSB WO_DT_BITS remain, bit WO_DT_BITS is the sign bit */
1106
1107 mask = ((1 << WO_DT_BITS) - 1);
1108 index &= mask;
1109 //printf("mask: 0x%x index: 0x%x\n", mask, index);
1110
1111 return index;
1112}
1113
1114/*---------------------------------------------------------------------------*\
1115
1116 FUNCTION....: decode_Wo_dt()
1117 AUTHOR......: David Rowe
1118 DATE CREATED: 6 Nov 2011
1119
1120 Decodes Wo using WO_DT_BITS difference from last frame.
1121
1122\*---------------------------------------------------------------------------*/
1123
1124float decode_Wo_dt(C2CONST *c2const, int index, float prev_Wo)
1125{
1126 float Wo_min = c2const->Wo_min;
1127 float Wo_max = c2const->Wo_max;
1128 float step;
1129 float Wo;
1130 int mask;
1131
1132 /* sign extend index */
1133
1134 //printf("DEC index: %d ");
1135 if (index & (1 << (WO_DT_BITS-1))) {
1136 mask = ~((1 << WO_DT_BITS) - 1);
1137 index |= mask;
1138 }
1139 //printf("DEC mask: 0x%x index: %d \n", mask, index);
1140
1141 step = (Wo_max - Wo_min)/WO_LEVELS;
1142 Wo = prev_Wo + step*(index);
1143
1144 /* bit errors can make us go out of range leading to all sorts of
1145 probs like seg faults */
1146 634
1147 if (Wo > Wo_max) Wo = Wo_max; 635 step = (log10f(Wo_max) - log10f(Wo_min)) / Wo_levels;
1148 if (Wo < Wo_min) Wo = Wo_min; 636 Wo = log10f(Wo_min) + step * (index);
1149 637
1150 return Wo; 638 return POW10F(Wo);
1151} 639}
1152#endif
1153 640
1154/*---------------------------------------------------------------------------*\ 641/*---------------------------------------------------------------------------*\
1155 642
@@ -1163,56 +650,46 @@ float decode_Wo_dt(C2CONST *c2const, int index, float prev_Wo)
1163 650
1164\*---------------------------------------------------------------------------*/ 651\*---------------------------------------------------------------------------*/
1165 652
1166float speech_to_uq_lsps(float lsp[], 653float speech_to_uq_lsps(float lsp[], float ak[], float Sn[], float w[],
1167 float ak[], 654 int m_pitch, int order) {
1168 float Sn[], 655 int i, roots;
1169 float w[], 656 float Wn[m_pitch];
1170 int m_pitch, 657 float R[order + 1];
1171 int order 658 float e, E;
1172) 659
1173{ 660 e = 0.0;
1174 int i, roots; 661 for (i = 0; i < m_pitch; i++) {
1175 float Wn[m_pitch]; 662 Wn[i] = Sn[i] * w[i];
1176 float R[order+1]; 663 e += Wn[i] * Wn[i];
1177 float e, E; 664 }
1178
1179 e = 0.0;
1180 for(i=0; i<m_pitch; i++) {
1181 Wn[i] = Sn[i]*w[i];
1182 e += Wn[i]*Wn[i];
1183 }
1184 665
1185 /* trap 0 energy case as LPC analysis will fail */ 666 /* trap 0 energy case as LPC analysis will fail */
1186 667
1187 if (e == 0.0) { 668 if (e == 0.0) {
1188 for(i=0; i<order; i++) 669 for (i = 0; i < order; i++) lsp[i] = (PI / order) * (float)i;
1189 lsp[i] = (PI/order)*(float)i; 670 return 0.0;
1190 return 0.0; 671 }
1191 }
1192 672
1193 autocorrelate(Wn, R, m_pitch, order); 673 autocorrelate(Wn, R, m_pitch, order);
1194 levinson_durbin(R, ak, order); 674 levinson_durbin(R, ak, order);
1195 675
1196 E = 0.0; 676 E = 0.0;
1197 for(i=0; i<=order; i++) 677 for (i = 0; i <= order; i++) E += ak[i] * R[i];
1198 E += ak[i]*R[i];
1199 678
1200 /* 15 Hz BW expansion as I can't hear the difference and it may help 679 /* 15 Hz BW expansion as I can't hear the difference and it may help
1201 help occasional fails in the LSP root finding. Important to do this 680 help occasional fails in the LSP root finding. Important to do this
1202 after energy calculation to avoid -ve energy values. 681 after energy calculation to avoid -ve energy values.
1203 */ 682 */
1204 683
1205 for(i=0; i<=order; i++) 684 for (i = 0; i <= order; i++) ak[i] *= powf(0.994, (float)i);
1206 ak[i] *= powf(0.994,(float)i);
1207 685
1208 roots = lpc_to_lsp(ak, order, lsp, 5, LSP_DELTA1); 686 roots = lpc_to_lsp(ak, order, lsp, 5, LSP_DELTA1);
1209 if (roots != order) { 687 if (roots != order) {
1210 /* if root finding fails use some benign LSP values instead */ 688 /* if root finding fails use some benign LSP values instead */
1211 for(i=0; i<order; i++) 689 for (i = 0; i < order; i++) lsp[i] = (PI / order) * (float)i;
1212 lsp[i] = (PI/order)*(float)i; 690 }
1213 }
1214 691
1215 return E; 692 return E;
1216} 693}
1217 694
1218/*---------------------------------------------------------------------------*\ 695/*---------------------------------------------------------------------------*\
@@ -1221,317 +698,60 @@ float speech_to_uq_lsps(float lsp[],
1221 AUTHOR......: David Rowe 698 AUTHOR......: David Rowe
1222 DATE CREATED: 22/8/2010 699 DATE CREATED: 22/8/2010
1223 700
1224 Thirty-six bit sclar LSP quantiser. From a vector of unquantised 701 Scalar LSP quantiser. From a vector of unquantised (floating point)
1225 (floating point) LSPs finds the quantised LSP indexes. 702 LSPs finds the quantised LSP indexes.
1226
1227\*---------------------------------------------------------------------------*/
1228
1229void encode_lsps_scalar(int indexes[], float lsp[], int order)
1230{
1231 int i,k,m;
1232 float wt[1];
1233 float lsp_hz[order];
1234 const float * cb;
1235 float se;
1236
1237 /* convert from radians to Hz so we can use human readable
1238 frequencies */
1239
1240 for(i=0; i<order; i++)
1241 lsp_hz[i] = (4000.0/PI)*lsp[i];
1242
1243 /* scalar quantisers */
1244
1245 wt[0] = 1.0;
1246 for(i=0; i<order; i++) {
1247 k = lsp_cb[i].k;
1248 m = lsp_cb[i].m;
1249 cb = lsp_cb[i].cb;
1250 indexes[i] = quantise(cb, &lsp_hz[i], wt, k, m, &se);
1251 }
1252}
1253
1254/*---------------------------------------------------------------------------*\
1255
1256 FUNCTION....: decode_lsps_scalar()
1257 AUTHOR......: David Rowe
1258 DATE CREATED: 22/8/2010
1259
1260 From a vector of quantised LSP indexes, returns the quantised
1261 (floating point) LSPs.
1262
1263\*---------------------------------------------------------------------------*/
1264
1265void decode_lsps_scalar(float lsp[], int indexes[], int order)
1266{
1267 int i,k;
1268 float lsp_hz[order];
1269 const float * cb;
1270
1271 for(i=0; i<order; i++) {
1272 k = lsp_cb[i].k;
1273 cb = lsp_cb[i].cb;
1274 lsp_hz[i] = cb[indexes[i]*k];
1275 }
1276
1277 /* convert back to radians */
1278
1279 for(i=0; i<order; i++)
1280 lsp[i] = (PI/4000.0)*lsp_hz[i];
1281}
1282
1283
1284/*---------------------------------------------------------------------------*\
1285
1286 FUNCTION....: encode_mels_scalar()
1287 AUTHOR......: David Rowe
1288 DATE CREATED: April 2015
1289
1290 Low bit rate mel coeff encoder.
1291 703
1292\*---------------------------------------------------------------------------*/ 704\*---------------------------------------------------------------------------*/
1293 705
1294void encode_mels_scalar(int indexes[], float mels[], int order) 706void encode_lsps_scalar(int indexes[], float lsp[], int order) {
1295{ 707 int i, k, m;
1296 int i,m; 708 float wt[1];
1297 float wt[1]; 709 float lsp_hz[order];
1298 const float * cb; 710 const float *cb;
1299 float se, mel_, dmel; 711 float se;
1300
1301 /* scalar quantisers */
1302
1303 wt[0] = 1.0;
1304 for(i=0; i<order; i++) {
1305 m = mel_cb[i].m;
1306 cb = mel_cb[i].cb;
1307 if (i%2) {
1308 /* on odd mels quantise difference */
1309 mel_ = mel_cb[i-1].cb[indexes[i-1]];
1310 dmel = mels[i] - mel_;
1311 indexes[i] = quantise(cb, &dmel, wt, 1, m, &se);
1312 //printf("%d mel: %f mel_: %f dmel: %f index: %d\n", i, mels[i], mel_, dmel, indexes[i]);
1313 }
1314 else {
1315 indexes[i] = quantise(cb, &mels[i], wt, 1, m, &se);
1316 //printf("%d mel: %f dmel: %f index: %d\n", i, mels[i], 0.0, indexes[i]);
1317 }
1318 712
1319 } 713 /* convert from radians to Hz so we can use human readable
1320} 714 frequencies */
1321 715
716 for (i = 0; i < order; i++) lsp_hz[i] = (4000.0 / PI) * lsp[i];
1322 717
1323/*---------------------------------------------------------------------------*\ 718 /* scalar quantisers */
1324 719
1325 FUNCTION....: decode_mels_scalar() 720 wt[0] = 1.0;
1326 AUTHOR......: David Rowe 721 for (i = 0; i < order; i++) {
1327 DATE CREATED: April 2015 722 k = lsp_cb[i].k;
1328 723 m = lsp_cb[i].m;
1329 From a vector of quantised mel indexes, returns the quantised 724 cb = lsp_cb[i].cb;
1330 (floating point) mels. 725 indexes[i] = quantise(cb, &lsp_hz[i], wt, k, m, &se);
1331 726 }
1332\*---------------------------------------------------------------------------*/
1333
1334void decode_mels_scalar(float mels[], int indexes[], int order)
1335{
1336 int i;
1337 const float * cb;
1338
1339 for(i=0; i<order; i++) {
1340 cb = mel_cb[i].cb;
1341 if (i%2) {
1342 /* on odd mels quantise difference */
1343 mels[i] = mels[i-1] + cb[indexes[i]];
1344 }
1345 else
1346 mels[i] = cb[indexes[i]];
1347 }
1348
1349}
1350
1351
1352#ifdef __EXPERIMENTAL__
1353
1354/*---------------------------------------------------------------------------*\
1355
1356 FUNCTION....: encode_lsps_diff_freq_vq()
1357 AUTHOR......: David Rowe
1358 DATE CREATED: 15 November 2011
1359
1360 Twenty-five bit LSP quantiser. LSPs 1-4 are quantised with scalar
1361 LSP differences (in frequency, i.e difference from the previous
1362 LSP). LSPs 5-10 are quantised with a VQ trained generated using
1363 vqtrainjnd.c
1364
1365\*---------------------------------------------------------------------------*/
1366
1367void encode_lsps_diff_freq_vq(int indexes[], float lsp[], int order)
1368{
1369 int i,k,m;
1370 float lsp_hz[order];
1371 float lsp__hz[order];
1372 float dlsp[order];
1373 float dlsp_[order];
1374 float wt[order];
1375 const float * cb;
1376 float se;
1377
1378 for(i=0; i<order; i++) {
1379 wt[i] = 1.0;
1380 }
1381
1382 /* convert from radians to Hz so we can use human readable
1383 frequencies */
1384
1385 for(i=0; i<order; i++)
1386 lsp_hz[i] = (4000.0/PI)*lsp[i];
1387
1388 /* scalar quantisers for LSP differences 1..4 */
1389
1390 wt[0] = 1.0;
1391 for(i=0; i<4; i++) {
1392 if (i)
1393 dlsp[i] = lsp_hz[i] - lsp__hz[i-1];
1394 else
1395 dlsp[0] = lsp_hz[0];
1396
1397 k = lsp_cbd[i].k;
1398 m = lsp_cbd[i].m;
1399 cb = lsp_cbd[i].cb;
1400 indexes[i] = quantise(cb, &dlsp[i], wt, k, m, &se);
1401 dlsp_[i] = cb[indexes[i]*k];
1402
1403 if (i)
1404 lsp__hz[i] = lsp__hz[i-1] + dlsp_[i];
1405 else
1406 lsp__hz[0] = dlsp_[0];
1407 }
1408
1409 /* VQ LSPs 5,6,7,8,9,10 */
1410
1411 k = lsp_cbjnd[4].k;
1412 m = lsp_cbjnd[4].m;
1413 cb = lsp_cbjnd[4].cb;
1414 indexes[4] = quantise(cb, &lsp_hz[4], &wt[4], k, m, &se);
1415} 727}
1416 728
1417
1418/*---------------------------------------------------------------------------*\ 729/*---------------------------------------------------------------------------*\
1419 730
1420 FUNCTION....: decode_lsps_diff_freq_vq() 731 FUNCTION....: decode_lsps_scalar()
1421 AUTHOR......: David Rowe 732 AUTHOR......: David Rowe
1422 DATE CREATED: 15 Nov 2011 733 DATE CREATED: 22/8/2010
1423 734
1424 From a vector of quantised LSP indexes, returns the quantised 735 From a vector of quantised LSP indexes, returns the quantised
1425 (floating point) LSPs. 736 (floating point) LSPs.
1426 737
1427\*---------------------------------------------------------------------------*/ 738\*---------------------------------------------------------------------------*/
1428 739
1429void decode_lsps_diff_freq_vq(float lsp_[], int indexes[], int order) 740void decode_lsps_scalar(float lsp[], int indexes[], int order) {
1430{ 741 int i, k;
1431 int i,k,m; 742 float lsp_hz[order];
1432 float dlsp_[order]; 743 const float *cb;
1433 float lsp__hz[order];
1434 const float * cb;
1435
1436 /* scalar LSP differences */
1437
1438 for(i=0; i<4; i++) {
1439 cb = lsp_cbd[i].cb;
1440 dlsp_[i] = cb[indexes[i]];
1441 if (i)
1442 lsp__hz[i] = lsp__hz[i-1] + dlsp_[i];
1443 else
1444 lsp__hz[0] = dlsp_[0];
1445 }
1446
1447 /* VQ */
1448
1449 k = lsp_cbjnd[4].k;
1450 m = lsp_cbjnd[4].m;
1451 cb = lsp_cbjnd[4].cb;
1452 for(i=4; i<order; i++)
1453 lsp__hz[i] = cb[indexes[4]*k+i-4];
1454
1455 /* convert back to radians */
1456
1457 for(i=0; i<order; i++)
1458 lsp_[i] = (PI/4000.0)*lsp__hz[i];
1459}
1460
1461
1462/*---------------------------------------------------------------------------*\
1463
1464 FUNCTION....: encode_lsps_diff_time()
1465 AUTHOR......: David Rowe
1466 DATE CREATED: 12 Sep 2012
1467
1468 Encode difference from preious frames's LSPs using
1469 3,3,2,2,2,2,1,1,1,1 scalar quantisers (18 bits total).
1470
1471\*---------------------------------------------------------------------------*/
1472 744
1473void encode_lsps_diff_time(int indexes[], 745 for (i = 0; i < order; i++) {
1474 float lsps[], 746 k = lsp_cb[i].k;
1475 float lsps__prev[], 747 cb = lsp_cb[i].cb;
1476 int order) 748 lsp_hz[i] = cb[indexes[i] * k];
1477{ 749 }
1478 int i,k,m;
1479 float lsps_dt[order];
1480 float wt[LPC_MAX];
1481 const float * cb;
1482 float se;
1483
1484 /* Determine difference in time and convert from radians to Hz so
1485 we can use human readable frequencies */
1486
1487 for(i=0; i<order; i++) {
1488 lsps_dt[i] = (4000/PI)*(lsps[i] - lsps__prev[i]);
1489 }
1490
1491 /* scalar quantisers */
1492
1493 wt[0] = 1.0;
1494 for(i=0; i<order; i++) {
1495 k = lsp_cbdt[i].k;
1496 m = lsp_cbdt[i].m;
1497 cb = lsp_cbdt[i].cb;
1498 indexes[i] = quantise(cb, &lsps_dt[i], wt, k, m, &se);
1499 }
1500
1501}
1502
1503
1504/*---------------------------------------------------------------------------*\
1505
1506 FUNCTION....: decode_lsps_diff_time()
1507 AUTHOR......: David Rowe
1508 DATE CREATED: 15 Nov 2011
1509
1510 From a quantised LSP indexes, returns the quantised
1511 (floating point) LSPs.
1512
1513\*---------------------------------------------------------------------------*/
1514
1515void decode_lsps_diff_time(
1516 float lsps_[],
1517 int indexes[],
1518 float lsps__prev[],
1519 int order)
1520{
1521 int i,k,m;
1522 const float * cb;
1523
1524 for(i=0; i<order; i++)
1525 lsps_[i] = lsps__prev[i];
1526 750
1527 for(i=0; i<order; i++) { 751 /* convert back to radians */
1528 k = lsp_cbdt[i].k;
1529 cb = lsp_cbdt[i].cb;
1530 lsps_[i] += (PI/4000.0)*cb[indexes[i]*k];
1531 }
1532 752
753 for (i = 0; i < order; i++) lsp[i] = (PI / 4000.0) * lsp_hz[i];
1533} 754}
1534#endif
1535 755
1536/*---------------------------------------------------------------------------*\ 756/*---------------------------------------------------------------------------*\
1537 757
@@ -1543,45 +763,40 @@ void decode_lsps_diff_time(
1543 763
1544\*---------------------------------------------------------------------------*/ 764\*---------------------------------------------------------------------------*/
1545 765
1546void encode_lsps_vq(int *indexes, float *x, float *xq, int order) 766void encode_lsps_vq(int *indexes, float *x, float *xq, int order) {
1547{
1548 int i, n1, n2, n3; 767 int i, n1, n2, n3;
1549 float err[order], err2[order], err3[order]; 768 float err[order], err2[order], err3[order];
1550 float w[order], w2[order], w3[order]; 769 float w[order], w2[order], w3[order];
1551 const float *codebook1 = lsp_cbjvm[0].cb; 770 const float *codebook1 = lsp_cbjmv[0].cb;
1552 const float *codebook2 = lsp_cbjvm[1].cb; 771 const float *codebook2 = lsp_cbjmv[1].cb;
1553 const float *codebook3 = lsp_cbjvm[2].cb; 772 const float *codebook3 = lsp_cbjmv[2].cb;
1554 773
1555 w[0] = MIN(x[0], x[1]-x[0]); 774 w[0] = MIN(x[0], x[1] - x[0]);
1556 for (i=1;i<order-1;i++) 775 for (i = 1; i < order - 1; i++) w[i] = MIN(x[i] - x[i - 1], x[i + 1] - x[i]);
1557 w[i] = MIN(x[i]-x[i-1], x[i+1]-x[i]); 776 w[order - 1] = MIN(x[order - 1] - x[order - 2], PI - x[order - 1]);
1558 w[order-1] = MIN(x[order-1]-x[order-2], PI-x[order-1]);
1559 777
1560 compute_weights(x, w, order); 778 compute_weights(x, w, order);
1561 779
1562 n1 = find_nearest(codebook1, lsp_cbjvm[0].m, x, order); 780 n1 = find_nearest(codebook1, lsp_cbjmv[0].m, x, order);
1563 781
1564 for (i=0;i<order;i++) 782 for (i = 0; i < order; i++) {
1565 { 783 xq[i] = codebook1[order * n1 + i];
1566 xq[i] = codebook1[order*n1+i];
1567 err[i] = x[i] - xq[i]; 784 err[i] = x[i] - xq[i];
1568 } 785 }
1569 for (i=0;i<order/2;i++) 786 for (i = 0; i < order / 2; i++) {
1570 { 787 err2[i] = err[2 * i];
1571 err2[i] = err[2*i]; 788 err3[i] = err[2 * i + 1];
1572 err3[i] = err[2*i+1]; 789 w2[i] = w[2 * i];
1573 w2[i] = w[2*i]; 790 w3[i] = w[2 * i + 1];
1574 w3[i] = w[2*i+1];
1575 } 791 }
1576 n2 = find_nearest_weighted(codebook2, lsp_cbjvm[1].m, err2, w2, order/2); 792 n2 = find_nearest_weighted(codebook2, lsp_cbjmv[1].m, err2, w2, order / 2);
1577 n3 = find_nearest_weighted(codebook3, lsp_cbjvm[2].m, err3, w3, order/2); 793 n3 = find_nearest_weighted(codebook3, lsp_cbjmv[2].m, err3, w3, order / 2);
1578 794
1579 indexes[0] = n1; 795 indexes[0] = n1;
1580 indexes[1] = n2; 796 indexes[1] = n2;
1581 indexes[2] = n3; 797 indexes[2] = n3;
1582} 798}
1583 799
1584
1585/*---------------------------------------------------------------------------*\ 800/*---------------------------------------------------------------------------*\
1586 801
1587 FUNCTION....: decode_lsps_vq() 802 FUNCTION....: decode_lsps_vq()
@@ -1590,31 +805,28 @@ void encode_lsps_vq(int *indexes, float *x, float *xq, int order)
1590 805
1591\*---------------------------------------------------------------------------*/ 806\*---------------------------------------------------------------------------*/
1592 807
1593void decode_lsps_vq(int *indexes, float *xq, int order, int stages) 808void decode_lsps_vq(int *indexes, float *xq, int order, int stages) {
1594{
1595 int i, n1, n2, n3; 809 int i, n1, n2, n3;
1596 const float *codebook1 = lsp_cbjvm[0].cb; 810 const float *codebook1 = lsp_cbjmv[0].cb;
1597 const float *codebook2 = lsp_cbjvm[1].cb; 811 const float *codebook2 = lsp_cbjmv[1].cb;
1598 const float *codebook3 = lsp_cbjvm[2].cb; 812 const float *codebook3 = lsp_cbjmv[2].cb;
1599 813
1600 n1 = indexes[0]; 814 n1 = indexes[0];
1601 n2 = indexes[1]; 815 n2 = indexes[1];
1602 n3 = indexes[2]; 816 n3 = indexes[2];
1603 817
1604 for (i=0;i<order;i++) { 818 for (i = 0; i < order; i++) {
1605 xq[i] = codebook1[order*n1+i]; 819 xq[i] = codebook1[order * n1 + i];
1606 } 820 }
1607 821
1608 if (stages != 1) { 822 if (stages != 1) {
1609 for (i=0;i<order/2;i++) { 823 for (i = 0; i < order / 2; i++) {
1610 xq[2*i] += codebook2[order*n2/2+i]; 824 xq[2 * i] += codebook2[order * n2 / 2 + i];
1611 xq[2*i+1] += codebook3[order*n3/2+i]; 825 xq[2 * i + 1] += codebook3[order * n3 / 2 + i];
1612 } 826 }
1613 } 827 }
1614
1615} 828}
1616 829
1617
1618/*---------------------------------------------------------------------------*\ 830/*---------------------------------------------------------------------------*\
1619 831
1620 FUNCTION....: bw_expand_lsps() 832 FUNCTION....: bw_expand_lsps()
@@ -1628,124 +840,45 @@ void decode_lsps_vq(int *indexes, float *xq, int order, int stages)
1628 840
1629\*---------------------------------------------------------------------------*/ 841\*---------------------------------------------------------------------------*/
1630 842
1631void bw_expand_lsps(float lsp[], int order, float min_sep_low, float min_sep_high) 843void bw_expand_lsps(float lsp[], int order, float min_sep_low,
1632{ 844 float min_sep_high) {
1633 int i; 845 int i;
1634
1635 for(i=1; i<4; i++) {
1636
1637 if ((lsp[i] - lsp[i-1]) < min_sep_low*(PI/4000.0))
1638 lsp[i] = lsp[i-1] + min_sep_low*(PI/4000.0);
1639
1640 }
1641
1642 /* As quantiser gaps increased, larger BW expansion was required
1643 to prevent twinkly noises. This may need more experiment for
1644 different quanstisers.
1645 */
1646
1647 for(i=4; i<order; i++) {
1648 if (lsp[i] - lsp[i-1] < min_sep_high*(PI/4000.0))
1649 lsp[i] = lsp[i-1] + min_sep_high*(PI/4000.0);
1650 }
1651}
1652
1653void bw_expand_lsps2(float lsp[],
1654 int order
1655)
1656{
1657 int i;
1658
1659 for(i=1; i<4; i++) {
1660
1661 if ((lsp[i] - lsp[i-1]) < 100.0*(PI/4000.0))
1662 lsp[i] = lsp[i-1] + 100.0*(PI/4000.0);
1663 846
1664 } 847 for (i = 1; i < 4; i++) {
848 if ((lsp[i] - lsp[i - 1]) < min_sep_low * (PI / 4000.0))
849 lsp[i] = lsp[i - 1] + min_sep_low * (PI / 4000.0);
850 }
1665 851
1666 /* As quantiser gaps increased, larger BW expansion was required 852 /* As quantiser gaps increased, larger BW expansion was required
1667 to prevent twinkly noises. This may need more experiment for 853 to prevent twinkly noises. This may need more experiment for
1668 different quanstisers. 854 different quanstisers.
1669 */ 855 */
1670 856
1671 for(i=4; i<order; i++) { 857 for (i = 4; i < order; i++) {
1672 if (lsp[i] - lsp[i-1] < 200.0*(PI/4000.0)) 858 if (lsp[i] - lsp[i - 1] < min_sep_high * (PI / 4000.0))
1673 lsp[i] = lsp[i-1] + 200.0*(PI/4000.0); 859 lsp[i] = lsp[i - 1] + min_sep_high * (PI / 4000.0);
1674 } 860 }
1675} 861}
1676 862
1677/*---------------------------------------------------------------------------*\ 863void bw_expand_lsps2(float lsp[], int order) {
1678 864 int i;
1679 FUNCTION....: locate_lsps_jnd_steps()
1680 AUTHOR......: David Rowe
1681 DATE CREATED: 27/10/2011
1682
1683 Applies a form of Bandwidth Expansion (BW) to a vector of LSPs.
1684 Listening tests have determined that "quantising" the position of
1685 each LSP to the non-linear steps below introduces a "just noticable
1686 difference" in the synthesised speech.
1687
1688 This operation can be used before quantisation to limit the input
1689 data to the quantiser to a number of discrete steps.
1690
1691 This operation can also be used during quantisation as a form of
1692 hysteresis in the calculation of quantiser error. For example if
1693 the quantiser target of lsp1 is 500 Hz, candidate vectors with lsp1
1694 of 515 and 495 Hz sound effectively the same.
1695
1696\*---------------------------------------------------------------------------*/
1697
1698void locate_lsps_jnd_steps(float lsps[], int order)
1699{
1700 int i;
1701 float lsp_hz, step;
1702
1703 assert(order == 10);
1704
1705 /* quantise to 25Hz steps */
1706
1707 step = 25;
1708 for(i=0; i<2; i++) {
1709 lsp_hz = lsps[i]*4000.0/PI;
1710 lsp_hz = floorf(lsp_hz/step + 0.5)*step;
1711 lsps[i] = lsp_hz*PI/4000.0;
1712 if (i) {
1713 if (lsps[i] == lsps[i-1])
1714 lsps[i] += step*PI/4000.0;
1715
1716 }
1717 }
1718
1719 /* quantise to 50Hz steps */
1720
1721 step = 50;
1722 for(i=2; i<4; i++) {
1723 lsp_hz = lsps[i]*4000.0/PI;
1724 lsp_hz = floorf(lsp_hz/step + 0.5)*step;
1725 lsps[i] = lsp_hz*PI/4000.0;
1726 if (i) {
1727 if (lsps[i] == lsps[i-1])
1728 lsps[i] += step*PI/4000.0;
1729
1730 }
1731 }
1732 865
1733 /* quantise to 100Hz steps */ 866 for (i = 1; i < 4; i++) {
867 if ((lsp[i] - lsp[i - 1]) < 100.0 * (PI / 4000.0))
868 lsp[i] = lsp[i - 1] + 100.0 * (PI / 4000.0);
869 }
1734 870
1735 step = 100; 871 /* As quantiser gaps increased, larger BW expansion was required
1736 for(i=4; i<10; i++) { 872 to prevent twinkly noises. This may need more experiment for
1737 lsp_hz = lsps[i]*4000.0/PI; 873 different quanstisers.
1738 lsp_hz = floorf(lsp_hz/step + 0.5)*step; 874 */
1739 lsps[i] = lsp_hz*PI/4000.0;
1740 if (i) {
1741 if (lsps[i] == lsps[i-1])
1742 lsps[i] += step*PI/4000.0;
1743 875
1744 } 876 for (i = 4; i < order; i++) {
1745 } 877 if (lsp[i] - lsp[i - 1] < 200.0 * (PI / 4000.0))
878 lsp[i] = lsp[i - 1] + 200.0 * (PI / 4000.0);
879 }
1746} 880}
1747 881
1748
1749/*---------------------------------------------------------------------------*\ 882/*---------------------------------------------------------------------------*\
1750 883
1751 FUNCTION....: apply_lpc_correction() 884 FUNCTION....: apply_lpc_correction()
@@ -1757,11 +890,10 @@ void locate_lsps_jnd_steps(float lsps[], int order)
1757 890
1758\*---------------------------------------------------------------------------*/ 891\*---------------------------------------------------------------------------*/
1759 892
1760void apply_lpc_correction(MODEL *model) 893void apply_lpc_correction(MODEL *model) {
1761{ 894 if (model->Wo < (PI * 150.0 / 4000)) {
1762 if (model->Wo < (PI*150.0/4000)) { 895 model->A[1] *= 0.032;
1763 model->A[1] *= 0.032; 896 }
1764 }
1765} 897}
1766 898
1767/*---------------------------------------------------------------------------*\ 899/*---------------------------------------------------------------------------*\
@@ -1774,124 +906,81 @@ void apply_lpc_correction(MODEL *model)
1774 906
1775\*---------------------------------------------------------------------------*/ 907\*---------------------------------------------------------------------------*/
1776 908
1777int encode_energy(float e, int bits) 909int encode_energy(float e, int bits) {
1778{ 910 int index, e_levels = 1 << bits;
1779 int index, e_levels = 1<<bits; 911 float e_min = E_MIN_DB;
1780 float e_min = E_MIN_DB; 912 float e_max = E_MAX_DB;
1781 float e_max = E_MAX_DB; 913 float norm;
1782 float norm;
1783
1784 e = 10.0*log10f(e);
1785 norm = (e - e_min)/(e_max - e_min);
1786 index = floorf(e_levels * norm + 0.5);
1787 if (index < 0 ) index = 0;
1788 if (index > (e_levels-1)) index = e_levels-1;
1789
1790 return index;
1791}
1792
1793/*---------------------------------------------------------------------------*\
1794
1795 FUNCTION....: decode_energy()
1796 AUTHOR......: David Rowe
1797 DATE CREATED: 22/8/2010
1798
1799 Decodes energy using a E_LEVELS quantiser.
1800
1801\*---------------------------------------------------------------------------*/
1802
1803float decode_energy(int index, int bits)
1804{
1805 float e_min = E_MIN_DB;
1806 float e_max = E_MAX_DB;
1807 float step;
1808 float e;
1809 int e_levels = 1<<bits;
1810 914
1811 step = (e_max - e_min)/e_levels; 915 e = 10.0 * log10f(e);
1812 e = e_min + step*(index); 916 norm = (e - e_min) / (e_max - e_min);
1813 e = POW10F(e/10.0); 917 index = floorf(e_levels * norm + 0.5);
918 if (index < 0) index = 0;
919 if (index > (e_levels - 1)) index = e_levels - 1;
1814 920
1815 return e; 921 return index;
1816} 922}
1817 923
1818#ifdef NOT_USED
1819/*---------------------------------------------------------------------------*\ 924/*---------------------------------------------------------------------------*\
1820 925
1821 FUNCTION....: decode_amplitudes() 926 FUNCTION....: decode_energy()
1822 AUTHOR......: David Rowe 927 AUTHOR......: David Rowe
1823 DATE CREATED: 22/8/2010 928 DATE CREATED: 22/8/2010
1824 929
1825 Given the amplitude quantiser indexes recovers the harmonic 930 Decodes energy using a E_LEVELS quantiser.
1826 amplitudes.
1827 931
1828\*---------------------------------------------------------------------------*/ 932\*---------------------------------------------------------------------------*/
1829 933
1830float decode_amplitudes(codec2_fft_cfg fft_fwd_cfg, 934float decode_energy(int index, int bits) {
1831 MODEL *model, 935 float e_min = E_MIN_DB;
1832 float ak[], 936 float e_max = E_MAX_DB;
1833 int lsp_indexes[], 937 float step;
1834 int energy_index, 938 float e;
1835 float lsps[], 939 int e_levels = 1 << bits;
1836 float *e
1837)
1838{
1839 float snr;
1840 940
1841 decode_lsps_scalar(lsps, lsp_indexes, LPC_ORD); 941 step = (e_max - e_min) / e_levels;
1842 bw_expand_lsps(lsps, LPC_ORD); 942 e = e_min + step * (index);
1843 lsp_to_lpc(lsps, ak, LPC_ORD); 943 e = POW10F(e / 10.0);
1844 *e = decode_energy(energy_index);
1845 aks_to_M2(ak, LPC_ORD, model, *e, &snr, 1, 0, 0, 1);
1846 apply_lpc_correction(model);
1847 944
1848 return snr; 945 return e;
1849} 946}
1850#endif
1851 947
1852static float ge_coeff[2] = {0.8, 0.9}; 948static float ge_coeff[2] = {0.8, 0.9};
1853 949
1854void compute_weights2(const float *x, const float *xp, float *w) 950void compute_weights2(const float *x, const float *xp, float *w) {
1855{
1856 w[0] = 30; 951 w[0] = 30;
1857 w[1] = 1; 952 w[1] = 1;
1858 if (x[1]<0) 953 if (x[1] < 0) {
1859 { 954 w[0] *= .6;
1860 w[0] *= .6; 955 w[1] *= .3;
1861 w[1] *= .3;
1862 } 956 }
1863 if (x[1]<-10) 957 if (x[1] < -10) {
1864 { 958 w[0] *= .3;
1865 w[0] *= .3; 959 w[1] *= .3;
1866 w[1] *= .3;
1867 } 960 }
1868 /* Higher weight if pitch is stable */ 961 /* Higher weight if pitch is stable */
1869 if (fabsf(x[0]-xp[0])<.2) 962 if (fabsf(x[0] - xp[0]) < .2) {
1870 { 963 w[0] *= 2;
1871 w[0] *= 2; 964 w[1] *= 1.5;
1872 w[1] *= 1.5; 965 } else if (fabsf(x[0] - xp[0]) > .5) /* Lower if not stable */
1873 } else if (fabsf(x[0]-xp[0])>.5) /* Lower if not stable */
1874 { 966 {
1875 w[0] *= .5; 967 w[0] *= .5;
1876 } 968 }
1877 969
1878 /* Lower weight for low energy */ 970 /* Lower weight for low energy */
1879 if (x[1] < xp[1]-10) 971 if (x[1] < xp[1] - 10) {
1880 { 972 w[1] *= .5;
1881 w[1] *= .5;
1882 } 973 }
1883 if (x[1] < xp[1]-20) 974 if (x[1] < xp[1] - 20) {
1884 { 975 w[1] *= .5;
1885 w[1] *= .5;
1886 } 976 }
1887 977
1888 //w[0] = 30; 978 // w[0] = 30;
1889 //w[1] = 1; 979 // w[1] = 1;
1890 980
1891 /* Square the weights because it's applied on the squared error */ 981 /* Square the weights because it's applied on the squared error */
1892 w[0] *= w[0]; 982 w[0] *= w[0];
1893 w[1] *= w[1]; 983 w[1] *= w[1];
1894
1895} 984}
1896 985
1897/*---------------------------------------------------------------------------*\ 986/*---------------------------------------------------------------------------*\
@@ -1906,7 +995,7 @@ void compute_weights2(const float *x, const float *xp, float *w)
1906 both the pitch and energy tend to only change by small amounts 995 both the pitch and energy tend to only change by small amounts
1907 during voiced speech, however it is important that these changes be 996 during voiced speech, however it is important that these changes be
1908 coded carefully. During unvoiced speech they both change a lot but 997 coded carefully. During unvoiced speech they both change a lot but
1909 the ear is less sensitve to errors so coarser quantisation is OK. 998 the ear is less sensitive to errors so coarser quantisation is OK.
1910 999
1911 The ear is sensitive to log energy and loq pitch so we quantise in 1000 The ear is sensitive to log energy and loq pitch so we quantise in
1912 these domains. That way the error measure used to quantise the 1001 these domains. That way the error measure used to quantise the
@@ -1916,15 +1005,14 @@ void compute_weights2(const float *x, const float *xp, float *w)
1916 1005
1917\*---------------------------------------------------------------------------*/ 1006\*---------------------------------------------------------------------------*/
1918 1007
1919void quantise_WoE(C2CONST *c2const, MODEL *model, float *e, float xq[]) 1008void quantise_WoE(C2CONST *c2const, MODEL *model, float *e, float xq[]) {
1920{ 1009 int i, n1;
1921 int i, n1; 1010 float x[2];
1922 float x[2]; 1011 float err[2];
1923 float err[2]; 1012 float w[2];
1924 float w[2];
1925 const float *codebook1 = ge_cb[0].cb; 1013 const float *codebook1 = ge_cb[0].cb;
1926 int nb_entries = ge_cb[0].m; 1014 int nb_entries = ge_cb[0].m;
1927 int ndim = ge_cb[0].k; 1015 int ndim = ge_cb[0].k;
1928 float Wo_min = c2const->Wo_min; 1016 float Wo_min = c2const->Wo_min;
1929 float Wo_max = c2const->Wo_max; 1017 float Wo_max = c2const->Wo_max;
1930 float Fs = c2const->Fs; 1018 float Fs = c2const->Fs;
@@ -1933,18 +1021,16 @@ void quantise_WoE(C2CONST *c2const, MODEL *model, float *e, float xq[])
1933 1021
1934 assert(Fs == 8000); 1022 assert(Fs == 8000);
1935 1023
1936 x[0] = log10f((model->Wo/PI)*4000.0/50.0)/log10f(2); 1024 x[0] = log10f((model->Wo / PI) * 4000.0 / 50.0) / log10f(2);
1937 x[1] = 10.0*log10f(1e-4 + *e); 1025 x[1] = 10.0 * log10f(1e-4 + *e);
1938 1026
1939 compute_weights2(x, xq, w); 1027 compute_weights2(x, xq, w);
1940 for (i=0;i<ndim;i++) 1028 for (i = 0; i < ndim; i++) err[i] = x[i] - ge_coeff[i] * xq[i];
1941 err[i] = x[i]-ge_coeff[i]*xq[i];
1942 n1 = find_nearest_weighted(codebook1, nb_entries, err, w, ndim); 1029 n1 = find_nearest_weighted(codebook1, nb_entries, err, w, ndim);
1943 1030
1944 for (i=0;i<ndim;i++) 1031 for (i = 0; i < ndim; i++) {
1945 { 1032 xq[i] = ge_coeff[i] * xq[i] + codebook1[ndim * n1 + i];
1946 xq[i] = ge_coeff[i]*xq[i] + codebook1[ndim*n1+i]; 1033 err[i] -= codebook1[ndim * n1 + i];
1947 err[i] -= codebook1[ndim*n1+i];
1948 } 1034 }
1949 1035
1950 /* 1036 /*
@@ -1953,7 +1039,7 @@ void quantise_WoE(C2CONST *c2const, MODEL *model, float *e, float xq[])
1953 Wo = (2^x)*(PI*50)/4000; 1039 Wo = (2^x)*(PI*50)/4000;
1954 */ 1040 */
1955 1041
1956 model->Wo = powf(2.0, xq[0])*(PI*50.0)/4000.0; 1042 model->Wo = powf(2.0, xq[0]) * (PI * 50.0) / 4000.0;
1957 1043
1958 /* bit errors can make us go out of range leading to all sorts of 1044 /* bit errors can make us go out of range leading to all sorts of
1959 probs like seg faults */ 1045 probs like seg faults */
@@ -1961,9 +1047,9 @@ void quantise_WoE(C2CONST *c2const, MODEL *model, float *e, float xq[])
1961 if (model->Wo > Wo_max) model->Wo = Wo_max; 1047 if (model->Wo > Wo_max) model->Wo = Wo_max;
1962 if (model->Wo < Wo_min) model->Wo = Wo_min; 1048 if (model->Wo < Wo_min) model->Wo = Wo_min;
1963 1049
1964 model->L = PI/model->Wo; /* if we quantise Wo re-compute L */ 1050 model->L = PI / model->Wo; /* if we quantise Wo re-compute L */
1965 1051
1966 *e = POW10F(xq[1]/10.0); 1052 *e = POW10F(xq[1] / 10.0);
1967} 1053}
1968 1054
1969/*---------------------------------------------------------------------------*\ 1055/*---------------------------------------------------------------------------*\
@@ -1977,39 +1063,36 @@ void quantise_WoE(C2CONST *c2const, MODEL *model, float *e, float xq[])
1977 1063
1978\*---------------------------------------------------------------------------*/ 1064\*---------------------------------------------------------------------------*/
1979 1065
1980int encode_WoE(MODEL *model, float e, float xq[]) 1066int encode_WoE(MODEL *model, float e, float xq[]) {
1981{ 1067 int i, n1;
1982 int i, n1; 1068 float x[2];
1983 float x[2]; 1069 float err[2];
1984 float err[2]; 1070 float w[2];
1985 float w[2];
1986 const float *codebook1 = ge_cb[0].cb; 1071 const float *codebook1 = ge_cb[0].cb;
1987 int nb_entries = ge_cb[0].m; 1072 int nb_entries = ge_cb[0].m;
1988 int ndim = ge_cb[0].k; 1073 int ndim = ge_cb[0].k;
1989 1074
1990 assert((1<<WO_E_BITS) == nb_entries); 1075 assert((1 << WO_E_BITS) == nb_entries);
1991 1076
1992 if (e < 0.0) e = 0; /* occasional small negative energies due LPC round off I guess */ 1077 if (e < 0.0)
1078 e = 0; /* occasional small negative energies due LPC round off I guess */
1993 1079
1994 x[0] = log10f((model->Wo/PI)*4000.0/50.0)/log10f(2); 1080 x[0] = log10f((model->Wo / PI) * 4000.0 / 50.0) / log10f(2);
1995 x[1] = 10.0*log10f(1e-4 + e); 1081 x[1] = 10.0 * log10f(1e-4 + e);
1996 1082
1997 compute_weights2(x, xq, w); 1083 compute_weights2(x, xq, w);
1998 for (i=0;i<ndim;i++) 1084 for (i = 0; i < ndim; i++) err[i] = x[i] - ge_coeff[i] * xq[i];
1999 err[i] = x[i]-ge_coeff[i]*xq[i];
2000 n1 = find_nearest_weighted(codebook1, nb_entries, err, w, ndim); 1085 n1 = find_nearest_weighted(codebook1, nb_entries, err, w, ndim);
2001 1086
2002 for (i=0;i<ndim;i++) 1087 for (i = 0; i < ndim; i++) {
2003 { 1088 xq[i] = ge_coeff[i] * xq[i] + codebook1[ndim * n1 + i];
2004 xq[i] = ge_coeff[i]*xq[i] + codebook1[ndim*n1+i]; 1089 err[i] -= codebook1[ndim * n1 + i];
2005 err[i] -= codebook1[ndim*n1+i];
2006 } 1090 }
2007 1091
2008 //printf("enc: %f %f (%f)(%f) \n", xq[0], xq[1], e, 10.0*log10(1e-4 + e)); 1092 // printf("enc: %f %f (%f)(%f) \n", xq[0], xq[1], e, 10.0*log10(1e-4 + e));
2009 return n1; 1093 return n1;
2010} 1094}
2011 1095
2012
2013/*---------------------------------------------------------------------------*\ 1096/*---------------------------------------------------------------------------*\
2014 1097
2015 FUNCTION....: decode_WoE() 1098 FUNCTION....: decode_WoE()
@@ -2022,21 +1105,19 @@ int encode_WoE(MODEL *model, float e, float xq[])
2022 1105
2023\*---------------------------------------------------------------------------*/ 1106\*---------------------------------------------------------------------------*/
2024 1107
2025void decode_WoE(C2CONST *c2const, MODEL *model, float *e, float xq[], int n1) 1108void decode_WoE(C2CONST *c2const, MODEL *model, float *e, float xq[], int n1) {
2026{ 1109 int i;
2027 int i;
2028 const float *codebook1 = ge_cb[0].cb; 1110 const float *codebook1 = ge_cb[0].cb;
2029 int ndim = ge_cb[0].k; 1111 int ndim = ge_cb[0].k;
2030 float Wo_min = c2const->Wo_min; 1112 float Wo_min = c2const->Wo_min;
2031 float Wo_max = c2const->Wo_max; 1113 float Wo_max = c2const->Wo_max;
2032 1114
2033 for (i=0;i<ndim;i++) 1115 for (i = 0; i < ndim; i++) {
2034 { 1116 xq[i] = ge_coeff[i] * xq[i] + codebook1[ndim * n1 + i];
2035 xq[i] = ge_coeff[i]*xq[i] + codebook1[ndim*n1+i];
2036 } 1117 }
2037 1118
2038 //printf("dec: %f %f\n", xq[0], xq[1]); 1119 // printf("dec: %f %f\n", xq[0], xq[1]);
2039 model->Wo = powf(2.0, xq[0])*(PI*50.0)/4000.0; 1120 model->Wo = powf(2.0, xq[0]) * (PI * 50.0) / 4000.0;
2040 1121
2041 /* bit errors can make us go out of range leading to all sorts of 1122 /* bit errors can make us go out of range leading to all sorts of
2042 probs like seg faults */ 1123 probs like seg faults */
@@ -2044,8 +1125,7 @@ void decode_WoE(C2CONST *c2const, MODEL *model, float *e, float xq[], int n1)
2044 if (model->Wo > Wo_max) model->Wo = Wo_max; 1125 if (model->Wo > Wo_max) model->Wo = Wo_max;
2045 if (model->Wo < Wo_min) model->Wo = Wo_min; 1126 if (model->Wo < Wo_min) model->Wo = Wo_min;
2046 1127
2047 model->L = PI/model->Wo; /* if we quantise Wo re-compute L */ 1128 model->L = PI / model->Wo; /* if we quantise Wo re-compute L */
2048 1129
2049 *e = POW10F(xq[1]/10.0); 1130 *e = POW10F(xq[1] / 10.0);
2050} 1131}
2051
diff --git a/quantise.h b/quantise.h
index 4baa87c..08d5532 100644
--- a/quantise.h
+++ b/quantise.h
@@ -29,113 +29,75 @@
29#include "codec2_fft.h" 29#include "codec2_fft.h"
30#include "comp.h" 30#include "comp.h"
31 31
32#define WO_BITS 7 32#define WO_BITS 7
33#define WO_LEVELS (1<<WO_BITS) 33#define WO_LEVELS (1 << WO_BITS)
34#define WO_DT_BITS 3 34#define WO_DT_BITS 3
35 35
36#define E_BITS 5 36#define E_BITS 5
37#define E_LEVELS (1<<E_BITS) 37#define E_LEVELS (1 << E_BITS)
38#define E_MIN_DB -10.0 38#define E_MIN_DB -10.0
39#define E_MAX_DB 40.0 39#define E_MAX_DB 40.0
40 40
41#define LSP_SCALAR_INDEXES 10 41#define LSP_SCALAR_INDEXES 10
42#define LSPD_SCALAR_INDEXES 10 42#define LSPD_SCALAR_INDEXES 10
43#define LSP_PRED_VQ_INDEXES 3 43#define LSP_PRED_VQ_INDEXES 3
44#define LSP_DIFF_FREQ_INDEXES 5
45#define LSP_DIFF_TIME_BITS 7
46 44
47#define LSPDT_ALL 0 45#define WO_E_BITS 8
48#define LSPDT_LOW 1
49#define LSPDT_HIGH 2
50
51#define WO_E_BITS 8
52 46
53#define LPCPF_GAMMA 0.5 47#define LPCPF_GAMMA 0.5
54#define LPCPF_BETA 0.2 48#define LPCPF_BETA 0.2
55 49
56void quantise_init();
57float lpc_model_amplitudes(float Sn[], float w[], MODEL *model, int order, 50float lpc_model_amplitudes(float Sn[], float w[], MODEL *model, int order,
58 int lsp,float ak[]); 51 int lsp, float ak[]);
59void aks_to_M2(codec2_fftr_cfg fftr_fwd_cfg, float ak[], int order, MODEL *model, 52void aks_to_M2(codec2_fftr_cfg fftr_fwd_cfg, float ak[], int order,
60 float E, float *snr, int dump, int sim_pf, 53 MODEL *model, float E, float *snr, int dump, int sim_pf, int pf,
61 int pf, int bass_boost, float beta, float gamma, COMP Aw[]); 54 int bass_boost, float beta, float gamma, COMP Aw[]);
62 55
63int encode_Wo(C2CONST *c2const, float Wo, int bits); 56int encode_Wo(C2CONST *c2const, float Wo, int bits);
64float decode_Wo(C2CONST *c2const, int index, int bits); 57float decode_Wo(C2CONST *c2const, int index, int bits);
65int encode_log_Wo(C2CONST *c2const, float Wo, int bits); 58int encode_log_Wo(C2CONST *c2const, float Wo, int bits);
66float decode_log_Wo(C2CONST *c2const, int index, int bits); 59float decode_log_Wo(C2CONST *c2const, int index, int bits);
67#if 0 60void encode_lsps_scalar(int indexes[], float lsp[], int order);
68int encode_Wo_dt(C2CONST *c2const, float Wo, float prev_Wo); 61void decode_lsps_scalar(float lsp[], int indexes[], int order);
69float decode_Wo_dt(C2CONST *c2const, int index, float prev_Wo); 62void encode_lspds_scalar(int indexes[], float lsp[], int order);
70#endif 63void decode_lspds_scalar(float lsp[], int indexes[], int order);
71void encode_lsps_scalar(int indexes[], float lsp[], int order);
72void decode_lsps_scalar(float lsp[], int indexes[], int order);
73void encode_lspds_scalar(int indexes[], float lsp[], int order);
74void decode_lspds_scalar(float lsp[], int indexes[], int order);
75void encode_lsps_diff_freq_vq(int indexes[], float lsp[], int order);
76void decode_lsps_diff_freq_vq(float lsp_[], int indexes[], int order);
77void encode_lsps_diff_time(int indexes[],
78 float lsp[],
79 float lsp__prev[],
80 int order);
81void decode_lsps_diff_time(float lsp_[],
82 int indexes[],
83 float lsp__prev[],
84 int order);
85 64
86void encode_lsps_vq(int *indexes, float *x, float *xq, int order); 65void encode_lsps_vq(int *indexes, float *x, float *xq, int order);
87void decode_lsps_vq(int *indexes, float *xq, int order, int stages); 66void decode_lsps_vq(int *indexes, float *xq, int order, int stages);
88 67
89long quantise(const float * cb, float vec[], float w[], int k, int m, float *se); 68long quantise(const float *cb, float vec[], float w[], int k, int m, float *se);
90void lspvq_quantise(float lsp[], float lsp_[], int order); 69void lspvq_quantise(float lsp[], float lsp_[], int order);
91void lspjnd_quantise(float lsp[], float lsp_[], int order); 70void lspjmv_quantise(float lsps[], float lsps_[], int order);
92void lspdt_quantise(float lsps[], float lsps_[], float lsps__prev[], int mode);
93void lspjvm_quantise(float lsps[], float lsps_[], int order);
94void lspanssi_quantise(float lsps[], float lsps_[], int order, int mbest_entries);
95float lspmelvq_quantise(float *x, float *xq, int order);
96
97float lspmelvq_mbest_encode(int *indexes, float *x, float *xq, int ndim, int mbest_entries);
98void lspmelvq_decode(int *indexes, float *xq, int ndim);
99
100void encode_mels_scalar(int mel_indexes[], float mels[], int order);
101void decode_mels_scalar(float mels[], int mel_indexes[], int order);
102 71
103void quantise_WoE(C2CONST *c2const, MODEL *model, float *e, float xq[]); 72void quantise_WoE(C2CONST *c2const, MODEL *model, float *e, float xq[]);
104int encode_WoE(MODEL *model, float e, float xq[]); 73int encode_WoE(MODEL *model, float e, float xq[]);
105void decode_WoE(C2CONST *c2const, MODEL *model, float *e, float xq[], int n1); 74void decode_WoE(C2CONST *c2const, MODEL *model, float *e, float xq[], int n1);
106 75
107int encode_energy(float e, int bits); 76int encode_energy(float e, int bits);
108float decode_energy(int index, int bits); 77float decode_energy(int index, int bits);
109 78
110void pack(unsigned char * bits, unsigned int *nbit, int index, unsigned int index_bits); 79void pack(unsigned char *bits, unsigned int *nbit, int index,
111void pack_natural_or_gray(unsigned char * bits, unsigned int *nbit, int index, unsigned int index_bits, unsigned int gray); 80 unsigned int index_bits);
112int unpack(const unsigned char * bits, unsigned int *nbit, unsigned int index_bits); 81void pack_natural_or_gray(unsigned char *bits, unsigned int *nbit, int index,
113int unpack_natural_or_gray(const unsigned char * bits, unsigned int *nbit, unsigned int index_bits, unsigned int gray); 82 unsigned int index_bits, unsigned int gray);
83int unpack(const unsigned char *bits, unsigned int *nbit,
84 unsigned int index_bits);
85int unpack_natural_or_gray(const unsigned char *bits, unsigned int *nbit,
86 unsigned int index_bits, unsigned int gray);
114 87
115int lsp_bits(int i); 88int lsp_bits(int i);
116int lspd_bits(int i); 89int lspd_bits(int i);
117int lspdt_bits(int i);
118int lsp_pred_vq_bits(int i); 90int lsp_pred_vq_bits(int i);
119int mel_bits(int i);
120int lspmelvq_cb_bits(int i);
121 91
122void apply_lpc_correction(MODEL *model); 92void apply_lpc_correction(MODEL *model);
123float speech_to_uq_lsps(float lsp[], 93float speech_to_uq_lsps(float lsp[], float ak[], float Sn[], float w[],
124 float ak[], 94 int m_pitch, int order);
125 float Sn[],
126 float w[],
127 int m_pitch,
128 int order
129 );
130int check_lsp_order(float lsp[], int lpc_order); 95int check_lsp_order(float lsp[], int lpc_order);
131void bw_expand_lsps(float lsp[], int order, float min_sep_low, float min_sep_high); 96void bw_expand_lsps(float lsp[], int order, float min_sep_low,
97 float min_sep_high);
132void bw_expand_lsps2(float lsp[], int order); 98void bw_expand_lsps2(float lsp[], int order);
133void locate_lsps_jnd_steps(float lsp[], int order); 99void locate_lsps_jnd_steps(float lsp[], int order);
134float decode_amplitudes(MODEL *model, 100float decode_amplitudes(MODEL *model, float ak[], int lsp_indexes[],
135 float ak[], 101 int energy_index, float lsps[], float *e);
136 int lsp_indexes[],
137 int energy_index,
138 float lsps[],
139 float *e);
140 102
141#endif 103#endif
diff --git a/sine.c b/sine.c
index d912cd7..814c269 100644
--- a/sine.c
+++ b/sine.c
@@ -27,64 +27,66 @@
27 27
28/*---------------------------------------------------------------------------*\ 28/*---------------------------------------------------------------------------*\
29 29
30 INCLUDES 30 INCLUDES
31 31
32\*---------------------------------------------------------------------------*/ 32\*---------------------------------------------------------------------------*/
33 33
34#include <stdlib.h> 34#include "sine.h"
35#include <stdio.h> 35
36#include <math.h> 36#include <math.h>
37#include <stdio.h>
38#include <stdlib.h>
37 39
38#include "defines.h" 40#include "defines.h"
39#include "sine.h"
40#include "kiss_fft.h" 41#include "kiss_fft.h"
41 42
42#define HPF_BETA 0.125 43#define HPF_BETA 0.125
43 44
44/*---------------------------------------------------------------------------*\ 45/*---------------------------------------------------------------------------*\
45 46
46 HEADERS 47 HEADERS
47 48
48\*---------------------------------------------------------------------------*/ 49\*---------------------------------------------------------------------------*/
49 50
50void hs_pitch_refinement(MODEL *model, COMP Sw[], float pmin, float pmax, 51void hs_pitch_refinement(MODEL *model, COMP Sw[], float pmin, float pmax,
51 float pstep); 52 float pstep);
52 53
53/*---------------------------------------------------------------------------*\ 54/*---------------------------------------------------------------------------*\
54 55
55 FUNCTIONS 56 FUNCTIONS
56 57
57\*---------------------------------------------------------------------------*/ 58\*---------------------------------------------------------------------------*/
58 59
59C2CONST c2const_create(int Fs, float framelength_s) { 60C2CONST c2const_create(int Fs, float framelength_s) {
60 C2CONST c2const; 61 C2CONST c2const;
61 62
62 assert((Fs == 8000) || (Fs = 16000)); 63 assert((Fs == 8000) || (Fs == 16000));
63 c2const.Fs = Fs; 64 c2const.Fs = Fs;
64 c2const.n_samp = round(Fs*framelength_s); 65 c2const.n_samp = round(Fs * framelength_s);
65 c2const.max_amp = floor(Fs*P_MIN_S/2); 66 c2const.max_amp = floor(Fs * P_MAX_S / 2);
66 c2const.p_min = floor(Fs*P_MIN_S); 67 c2const.p_min = floor(Fs * P_MIN_S);
67 c2const.p_max = floor(Fs*P_MAX_S); 68 c2const.p_max = floor(Fs * P_MAX_S);
68 c2const.m_pitch = floor(Fs*M_PITCH_S); 69 c2const.m_pitch = floor(Fs * M_PITCH_S);
69 c2const.Wo_min = TWO_PI/c2const.p_max; 70 c2const.Wo_min = TWO_PI / c2const.p_max;
70 c2const.Wo_max = TWO_PI/c2const.p_min; 71 c2const.Wo_max = TWO_PI / c2const.p_min;
71 72
72 if (Fs == 8000) { 73 if (Fs == 8000) {
73 c2const.nw = 279; 74 c2const.nw = 279;
74 } else { 75 } else {
75 c2const.nw = 511; /* actually a bit shorter in time but lets us maintain constant FFT size */ 76 c2const.nw = 511; /* actually a bit shorter in time but lets us maintain
76 } 77 constant FFT size */
78 }
77 79
78 c2const.tw = Fs*TW_S; 80 c2const.tw = Fs * TW_S;
79 81
80 /* 82 /*
81 fprintf(stderr, "max_amp: %d m_pitch: %d\n", c2const.n_samp, c2const.m_pitch); 83 fprintf(stderr, "max_amp: %d m_pitch: %d\n", c2const.n_samp, c2const.m_pitch);
82 fprintf(stderr, "p_min: %d p_max: %d\n", c2const.p_min, c2const.p_max); 84 fprintf(stderr, "p_min: %d p_max: %d\n", c2const.p_min, c2const.p_max);
83 fprintf(stderr, "Wo_min: %f Wo_max: %f\n", c2const.Wo_min, c2const.Wo_max); 85 fprintf(stderr, "Wo_min: %f Wo_max: %f\n", c2const.Wo_min, c2const.Wo_max);
84 fprintf(stderr, "nw: %d tw: %d\n", c2const.nw, c2const.tw); 86 fprintf(stderr, "nw: %d tw: %d\n", c2const.nw, c2const.tw);
85 */ 87 */
86 88
87 return c2const; 89 return c2const;
88} 90}
89 91
90/*---------------------------------------------------------------------------*\ 92/*---------------------------------------------------------------------------*\
@@ -97,13 +99,13 @@ C2CONST c2const_create(int Fs, float framelength_s) {
97 99
98\*---------------------------------------------------------------------------*/ 100\*---------------------------------------------------------------------------*/
99 101
100void make_analysis_window(C2CONST *c2const, codec2_fft_cfg fft_fwd_cfg, float w[], float W[]) 102void make_analysis_window(C2CONST *c2const, codec2_fft_cfg fft_fwd_cfg,
101{ 103 float w[], float W[]) {
102 float m; 104 float m;
103 COMP wshift[FFT_ENC]; 105 COMP wshift[FFT_ENC];
104 int i,j; 106 int i, j;
105 int m_pitch = c2const->m_pitch; 107 int m_pitch = c2const->m_pitch;
106 int nw = c2const->nw; 108 int nw = c2const->nw;
107 109
108 /* 110 /*
109 Generate Hamming window centered on M-sample pitch analysis window 111 Generate Hamming window centered on M-sample pitch analysis window
@@ -117,20 +119,18 @@ void make_analysis_window(C2CONST *c2const, codec2_fft_cfg fft_fwd_cfg, float w[
117 */ 119 */
118 120
119 m = 0.0; 121 m = 0.0;
120 for(i=0; i<m_pitch/2-nw/2; i++) 122 for (i = 0; i < m_pitch / 2 - nw / 2; i++) w[i] = 0.0;
121 w[i] = 0.0; 123 for (i = m_pitch / 2 - nw / 2, j = 0; i < m_pitch / 2 + nw / 2; i++, j++) {
122 for(i=m_pitch/2-nw/2,j=0; i<m_pitch/2+nw/2; i++,j++) { 124 w[i] = 0.5 - 0.5 * cosf(TWO_PI * j / (nw - 1));
123 w[i] = 0.5 - 0.5*cosf(TWO_PI*j/(nw-1)); 125 m += w[i] * w[i];
124 m += w[i]*w[i];
125 } 126 }
126 for(i=m_pitch/2+nw/2; i<m_pitch; i++) 127 for (i = m_pitch / 2 + nw / 2; i < m_pitch; i++) w[i] = 0.0;
127 w[i] = 0.0;
128 128
129 /* Normalise - makes freq domain amplitude estimation straight 129 /* Normalise - makes freq domain amplitude estimation straight
130 forward */ 130 forward */
131 131
132 m = 1.0/sqrtf(m*FFT_ENC); 132 m = 1.0 / sqrtf(m * FFT_ENC);
133 for(i=0; i<m_pitch; i++) { 133 for (i = 0; i < m_pitch; i++) {
134 w[i] *= m; 134 w[i] *= m;
135 } 135 }
136 136
@@ -157,14 +157,13 @@ void make_analysis_window(C2CONST *c2const, codec2_fft_cfg fft_fwd_cfg, float w[
157 157
158 COMP temp[FFT_ENC]; 158 COMP temp[FFT_ENC];
159 159
160 for(i=0; i<FFT_ENC; i++) { 160 for (i = 0; i < FFT_ENC; i++) {
161 wshift[i].real = 0.0; 161 wshift[i].real = 0.0;
162 wshift[i].imag = 0.0; 162 wshift[i].imag = 0.0;
163 } 163 }
164 for(i=0; i<nw/2; i++) 164 for (i = 0; i < nw / 2; i++) wshift[i].real = w[i + m_pitch / 2];
165 wshift[i].real = w[i+m_pitch/2]; 165 for (i = FFT_ENC - nw / 2, j = m_pitch / 2 - nw / 2; i < FFT_ENC; i++, j++)
166 for(i=FFT_ENC-nw/2,j=m_pitch/2-nw/2; i<FFT_ENC; i++,j++) 166 wshift[i].real = w[j];
167 wshift[i].real = w[j];
168 167
169 codec2_fft(fft_fwd_cfg, wshift, temp); 168 codec2_fft(fft_fwd_cfg, wshift, temp);
170 169
@@ -191,12 +190,10 @@ void make_analysis_window(C2CONST *c2const, codec2_fft_cfg fft_fwd_cfg, float w[
191 190
192 */ 191 */
193 192
194 193 for (i = 0; i < FFT_ENC / 2; i++) {
195 for(i=0; i<FFT_ENC/2; i++) { 194 W[i] = temp[i + FFT_ENC / 2].real;
196 W[i] = temp[i + FFT_ENC / 2].real; 195 W[i + FFT_ENC / 2] = temp[i].real;
197 W[i + FFT_ENC / 2] = temp[i].real;
198 } 196 }
199
200} 197}
201 198
202/*---------------------------------------------------------------------------*\ 199/*---------------------------------------------------------------------------*\
@@ -211,12 +208,11 @@ void make_analysis_window(C2CONST *c2const, codec2_fft_cfg fft_fwd_cfg, float w[
211 208
212\*---------------------------------------------------------------------------*/ 209\*---------------------------------------------------------------------------*/
213 210
214float hpf(float x, float states[]) 211float hpf(float x, float states[]) {
215{ 212 states[0] = -HPF_BETA * states[0] + x - states[1];
216 states[0] = -HPF_BETA*states[0] + x - states[1]; 213 states[1] = x;
217 states[1] = x;
218 214
219 return states[0]; 215 return states[0];
220} 216}
221 217
222/*---------------------------------------------------------------------------*\ 218/*---------------------------------------------------------------------------*\
@@ -230,41 +226,43 @@ float hpf(float x, float states[])
230\*---------------------------------------------------------------------------*/ 226\*---------------------------------------------------------------------------*/
231 227
232// TODO: we can either go for a faster FFT using fftr and some stack usage 228// TODO: we can either go for a faster FFT using fftr and some stack usage
233// or we can reduce stack usage to almost zero on STM32 by switching to fft_inplace 229// or we can reduce stack usage to almost zero on STM32 by switching to
230// fft_inplace
234#if 1 231#if 1
235void dft_speech(C2CONST *c2const, codec2_fft_cfg fft_fwd_cfg, COMP Sw[], float Sn[], float w[]) 232void dft_speech(C2CONST *c2const, codec2_fft_cfg fft_fwd_cfg, COMP Sw[],
236{ 233 float Sn[], float w[]) {
237 int i; 234 int i;
238 int m_pitch = c2const->m_pitch; 235 int m_pitch = c2const->m_pitch;
239 int nw = c2const->nw; 236 int nw = c2const->nw;
240 237
241 for(i=0; i<FFT_ENC; i++) { 238 for (i = 0; i < FFT_ENC; i++) {
242 Sw[i].real = 0.0; 239 Sw[i].real = 0.0;
243 Sw[i].imag = 0.0; 240 Sw[i].imag = 0.0;
244 } 241 }
245 242
246 /* Centre analysis window on time axis, we need to arrange input 243 /* Centre analysis window on time axis, we need to arrange input
247 to FFT this way to make FFT phases correct */ 244 to FFT this way to make FFT phases correct */
248 245
249 /* move 2nd half to start of FFT input vector */ 246 /* move 2nd half to start of FFT input vector */
250 247
251 for(i=0; i<nw/2; i++) 248 for (i = 0; i < nw / 2; i++)
252 Sw[i].real = Sn[i+m_pitch/2]*w[i+m_pitch/2]; 249 Sw[i].real = Sn[i + m_pitch / 2] * w[i + m_pitch / 2];
253 250
254 /* move 1st half to end of FFT input vector */ 251 /* move 1st half to end of FFT input vector */
255 252
256 for(i=0; i<nw/2; i++) 253 for (i = 0; i < nw / 2; i++)
257 Sw[FFT_ENC-nw/2+i].real = Sn[i+m_pitch/2-nw/2]*w[i+m_pitch/2-nw/2]; 254 Sw[FFT_ENC - nw / 2 + i].real =
255 Sn[i + m_pitch / 2 - nw / 2] * w[i + m_pitch / 2 - nw / 2];
258 256
259 codec2_fft_inplace(fft_fwd_cfg, Sw); 257 codec2_fft_inplace(fft_fwd_cfg, Sw);
260} 258}
261#else 259#else
262void dft_speech(codec2_fftr_cfg fftr_fwd_cfg, COMP Sw[], float Sn[], float w[]) 260void dft_speech(codec2_fftr_cfg fftr_fwd_cfg, COMP Sw[], float Sn[],
263{ 261 float w[]) {
264 int i; 262 int i;
265 float sw[FFT_ENC]; 263 float sw[FFT_ENC];
266 264
267 for(i=0; i<FFT_ENC; i++) { 265 for (i = 0; i < FFT_ENC; i++) {
268 sw[i] = 0.0; 266 sw[i] = 0.0;
269 } 267 }
270 268
@@ -273,19 +271,18 @@ void dft_speech(codec2_fftr_cfg fftr_fwd_cfg, COMP Sw[], float Sn[], float w[])
273 271
274 /* move 2nd half to start of FFT input vector */ 272 /* move 2nd half to start of FFT input vector */
275 273
276 for(i=0; i<nw/2; i++) 274 for (i = 0; i < nw / 2; i++) sw[i] = Sn[i + m_pitch / 2] * w[i + m_pitch / 2];
277 sw[i] = Sn[i+m_pitch/2]*w[i+m_pitch/2];
278 275
279 /* move 1st half to end of FFT input vector */ 276 /* move 1st half to end of FFT input vector */
280 277
281 for(i=0; i<nw/2; i++) 278 for (i = 0; i < nw / 2; i++)
282 sw[FFT_ENC-nw/2+i] = Sn[i+m_pitch/2-nw/2]*w[i+m_pitch/2-nw/2]; 279 sw[FFT_ENC - nw / 2 + i] =
280 Sn[i + m_pitch / 2 - nw / 2] * w[i + m_pitch / 2 - nw / 2];
283 281
284 codec2_fftr(fftr_fwd_cfg, sw, Sw); 282 codec2_fftr(fftr_fwd_cfg, sw, Sw);
285} 283}
286#endif 284#endif
287 285
288
289/*---------------------------------------------------------------------------*\ 286/*---------------------------------------------------------------------------*\
290 287
291 FUNCTION....: two_stage_pitch_refinement 288 FUNCTION....: two_stage_pitch_refinement
@@ -297,38 +294,35 @@ void dft_speech(codec2_fftr_cfg fftr_fwd_cfg, COMP Sw[], float Sn[], float w[])
297 294
298\*---------------------------------------------------------------------------*/ 295\*---------------------------------------------------------------------------*/
299 296
300void two_stage_pitch_refinement(C2CONST *c2const, MODEL *model, COMP Sw[]) 297void two_stage_pitch_refinement(C2CONST *c2const, MODEL *model, COMP Sw[]) {
301{ 298 float pmin, pmax, pstep; /* pitch refinement minimum, maximum and step */
302 float pmin,pmax,pstep; /* pitch refinment minimum, maximum and step */
303 299
304 /* Coarse refinement */ 300 /* Coarse refinement */
305 301
306 pmax = TWO_PI/model->Wo + 5; 302 pmax = TWO_PI / model->Wo + 5;
307 pmin = TWO_PI/model->Wo - 5; 303 pmin = TWO_PI / model->Wo - 5;
308 pstep = 1.0; 304 pstep = 1.0;
309 hs_pitch_refinement(model,Sw,pmin,pmax,pstep); 305 hs_pitch_refinement(model, Sw, pmin, pmax, pstep);
310 306
311 /* Fine refinement */ 307 /* Fine refinement */
312 308
313 pmax = TWO_PI/model->Wo + 1; 309 pmax = TWO_PI / model->Wo + 1;
314 pmin = TWO_PI/model->Wo - 1; 310 pmin = TWO_PI / model->Wo - 1;
315 pstep = 0.25; 311 pstep = 0.25;
316 hs_pitch_refinement(model,Sw,pmin,pmax,pstep); 312 hs_pitch_refinement(model, Sw, pmin, pmax, pstep);
317 313
318 /* Limit range */ 314 /* Limit range */
319 315
320 if (model->Wo < TWO_PI/c2const->p_max) 316 if (model->Wo < TWO_PI / c2const->p_max) model->Wo = TWO_PI / c2const->p_max;
321 model->Wo = TWO_PI/c2const->p_max; 317 if (model->Wo > TWO_PI / c2const->p_min) model->Wo = TWO_PI / c2const->p_min;
322 if (model->Wo > TWO_PI/c2const->p_min)
323 model->Wo = TWO_PI/c2const->p_min;
324 318
325 model->L = floorf(PI/model->Wo); 319 model->L = floorf(PI / model->Wo);
326 320
327 /* trap occasional round off issues with floorf() */ 321 /* trap occasional round off issues with floorf() */
328 if (model->Wo*model->L >= 0.95*PI) { 322 if (model->Wo * model->L >= 0.95 * PI) {
329 model->L--; 323 model->L--;
330 } 324 }
331 assert(model->Wo*model->L < PI); 325 assert(model->Wo * model->L < PI);
332} 326}
333 327
334/*---------------------------------------------------------------------------*\ 328/*---------------------------------------------------------------------------*\
@@ -348,35 +342,39 @@ void two_stage_pitch_refinement(C2CONST *c2const, MODEL *model, COMP Sw[])
348 342
349\*---------------------------------------------------------------------------*/ 343\*---------------------------------------------------------------------------*/
350 344
351void hs_pitch_refinement(MODEL *model, COMP Sw[], float pmin, float pmax, float pstep) 345void hs_pitch_refinement(MODEL *model, COMP Sw[], float pmin, float pmax,
352{ 346 float pstep) {
353 int m; /* loop variable */ 347 int m; /* loop variable */
354 int b; /* bin for current harmonic centre */ 348 int b; /* bin for current harmonic centre */
355 float E; /* energy for current pitch*/ 349 float E; /* energy for current pitch*/
356 float Wo; /* current "test" fundamental freq. */ 350 float Wo; /* current "test" fundamental freq. */
357 float Wom; /* Wo that maximises E */ 351 float Wom; /* Wo that maximises E */
358 float Em; /* mamimum energy */ 352 float Em; /* mamimum energy */
359 float r, one_on_r; /* number of rads/bin */ 353 float r, one_on_r; /* number of rads/bin */
360 float p; /* current pitch */ 354 float p; /* current pitch */
361 355
362 /* Initialisation */ 356 /* Initialisation */
363 357
364 model->L = PI/model->Wo; /* use initial pitch est. for L */ 358 model->L = PI / model->Wo; /* use initial pitch est. for L */
365 Wom = model->Wo; 359 Wom = model->Wo;
366 Em = 0.0; 360 Em = 0.0;
367 r = TWO_PI/FFT_ENC; 361 r = TWO_PI / FFT_ENC;
368 one_on_r = 1.0/r; 362 one_on_r = 1.0 / r;
369 363
370 /* Determine harmonic sum for a range of Wo values */ 364 /* Determine harmonic sum for a range of Wo values */
371 365
372 for(p=pmin; p<=pmax; p+=pstep) { 366 for (p = pmin; p <= pmax; p += pstep) {
373 E = 0.0; 367 E = 0.0;
374 Wo = TWO_PI/p; 368 Wo = TWO_PI / p;
369
370 float bFloat = Wo * one_on_r;
371 float currentBFloat = bFloat;
375 372
376 /* Sum harmonic magnitudes */ 373 /* Sum harmonic magnitudes */
377 for(m=1; m<=model->L; m++) { 374 for (m = 1; m <= model->L; m++) {
378 b = (int)(m*Wo*one_on_r + 0.5); 375 b = (int)(currentBFloat + 0.5);
379 E += Sw[b].real*Sw[b].real + Sw[b].imag*Sw[b].imag; 376 E += Sw[b].real * Sw[b].real + Sw[b].imag * Sw[b].imag;
377 currentBFloat += bFloat;
380 } 378 }
381 /* Compare to see if this is a maximum */ 379 /* Compare to see if this is a maximum */
382 380
@@ -399,35 +397,35 @@ void hs_pitch_refinement(MODEL *model, COMP Sw[], float pmin, float pmax, float
399 397
400\*---------------------------------------------------------------------------*/ 398\*---------------------------------------------------------------------------*/
401 399
402void estimate_amplitudes(MODEL *model, COMP Sw[], float W[], int est_phase) 400void estimate_amplitudes(MODEL *model, COMP Sw[], float W[], int est_phase) {
403{ 401 int i, m; /* loop variables */
404 int i,m; /* loop variables */ 402 int am, bm; /* bounds of current harmonic */
405 int am,bm; /* bounds of current harmonic */ 403 float den; /* denominator of amplitude expression */
406 float den; /* denominator of amplitude expression */
407 404
408 float r = TWO_PI/FFT_ENC; 405 float r = TWO_PI / FFT_ENC;
409 float one_on_r = 1.0/r; 406 float one_on_r = 1.0 / r;
410 407
411 for(m=1; m<=model->L; m++) { 408 for (m = 1; m <= model->L; m++) {
412 /* Estimate ampltude of harmonic */ 409 /* Estimate ampltude of harmonic */
413 410
414 den = 0.0; 411 den = 0.0;
415 am = (int)((m - 0.5)*model->Wo*one_on_r + 0.5); 412 am = (int)((m - 0.5) * model->Wo * one_on_r + 0.5);
416 bm = (int)((m + 0.5)*model->Wo*one_on_r + 0.5); 413 bm = (int)((m + 0.5) * model->Wo * one_on_r + 0.5);
417 414
418 for(i=am; i<bm; i++) { 415 for (i = am; i < bm; i++) {
419 den += Sw[i].real*Sw[i].real + Sw[i].imag*Sw[i].imag; 416 den += Sw[i].real * Sw[i].real + Sw[i].imag * Sw[i].imag;
420 } 417 }
421 418
422 model->A[m] = sqrtf(den); 419 model->A[m] = sqrtf(den);
423 420
424 if (est_phase) { 421 if (est_phase) {
425 int b = (int)(m*model->Wo/r + 0.5); /* DFT bin of centre of current harmonic */ 422 int b = (int)(m * model->Wo / r +
423 0.5); /* DFT bin of centre of current harmonic */
426 424
427 /* Estimate phase of harmonic, this is expensive in CPU for 425 /* Estimate phase of harmonic, this is expensive in CPU for
428 embedded devicesso we make it an option */ 426 embedded devicesso we make it an option */
429 427
430 model->phi[m] = atan2f(Sw[b].imag,Sw[b].real); 428 model->phi[m] = atan2f(Sw[b].imag, Sw[b].real);
431 } 429 }
432 } 430 }
433} 431}
@@ -443,119 +441,110 @@ void estimate_amplitudes(MODEL *model, COMP Sw[], float W[], int est_phase)
443 441
444\*---------------------------------------------------------------------------*/ 442\*---------------------------------------------------------------------------*/
445 443
446float est_voicing_mbe( 444float est_voicing_mbe(C2CONST *c2const, MODEL *model, COMP Sw[], float W[]) {
447 C2CONST *c2const, 445 int l, al, bl, m; /* loop variables */
448 MODEL *model, 446 COMP Am; /* amplitude sample for this band */
449 COMP Sw[], 447 int offset; /* centers Hw[] about current harmonic */
450 float W[] 448 float den; /* denominator of Am expression */
451 ) 449 float error; /* accumulated error between original and synthesised */
452{ 450 float Wo;
453 int l,al,bl,m; /* loop variables */ 451 float sig, snr;
454 COMP Am; /* amplitude sample for this band */ 452 float elow, ehigh, eratio;
455 int offset; /* centers Hw[] about current harmonic */ 453 float sixty;
456 float den; /* denominator of Am expression */ 454 COMP Ew;
457 float error; /* accumulated error between original and synthesised */ 455 Ew.real = 0;
458 float Wo; 456 Ew.imag = 0;
459 float sig, snr; 457
460 float elow, ehigh, eratio; 458 int l_1000hz = model->L * 1000.0 / (c2const->Fs / 2);
461 float sixty; 459 sig = 1E-4;
462 COMP Ew; 460 for (l = 1; l <= l_1000hz; l++) {
463 Ew.real = 0; 461 sig += model->A[l] * model->A[l];
464 Ew.imag = 0; 462 }
465
466 int l_1000hz = model->L*1000.0/(c2const->Fs/2);
467 sig = 1E-4;
468 for(l=1; l<=l_1000hz; l++) {
469 sig += model->A[l]*model->A[l];
470 }
471 463
472 Wo = model->Wo; 464 Wo = model->Wo;
473 error = 1E-4; 465 error = 1E-4;
474 466
475 /* Just test across the harmonics in the first 1000 Hz */ 467 /* Just test across the harmonics in the first 1000 Hz */
476 468
477 for(l=1; l<=l_1000hz; l++) { 469 for (l = 1; l <= l_1000hz; l++) {
478 Am.real = 0.0; 470 Am.real = 0.0;
479 Am.imag = 0.0; 471 Am.imag = 0.0;
480 den = 0.0; 472 den = 0.0;
481 al = ceilf((l - 0.5)*Wo*FFT_ENC/TWO_PI); 473 al = ceilf((l - 0.5) * Wo * FFT_ENC / TWO_PI);
482 bl = ceilf((l + 0.5)*Wo*FFT_ENC/TWO_PI); 474 bl = ceilf((l + 0.5) * Wo * FFT_ENC / TWO_PI);
483 475
484 /* Estimate amplitude of harmonic assuming harmonic is totally voiced */ 476 /* Estimate amplitude of harmonic assuming harmonic is totally voiced */
485 477
486 offset = FFT_ENC/2 - l*Wo*FFT_ENC/TWO_PI + 0.5; 478 offset = FFT_ENC / 2 - l * Wo * FFT_ENC / TWO_PI + 0.5;
487 for(m=al; m<bl; m++) { 479 for (m = al; m < bl; m++) {
488 Am.real += Sw[m].real*W[offset+m]; 480 Am.real += Sw[m].real * W[offset + m];
489 Am.imag += Sw[m].imag*W[offset+m]; 481 Am.imag += Sw[m].imag * W[offset + m];
490 den += W[offset+m]*W[offset+m]; 482 den += W[offset + m] * W[offset + m];
491 } 483 }
492 484
493 Am.real = Am.real/den; 485 Am.real = Am.real / den;
494 Am.imag = Am.imag/den; 486 Am.imag = Am.imag / den;
495 487
496 /* Determine error between estimated harmonic and original */ 488 /* Determine error between estimated harmonic and original */
497 489
498 for(m=al; m<bl; m++) { 490 for (m = al; m < bl; m++) {
499 Ew.real = Sw[m].real - Am.real*W[offset+m]; 491 Ew.real = Sw[m].real - Am.real * W[offset + m];
500 Ew.imag = Sw[m].imag - Am.imag*W[offset+m]; 492 Ew.imag = Sw[m].imag - Am.imag * W[offset + m];
501 error += Ew.real*Ew.real; 493 error += Ew.real * Ew.real;
502 error += Ew.imag*Ew.imag; 494 error += Ew.imag * Ew.imag;
503 }
504 } 495 }
496 }
505 497
506 snr = 10.0*log10f(sig/error); 498 snr = 10.0 * log10f(sig / error);
507 if (snr > V_THRESH) 499 if (snr > V_THRESH)
508 model->voiced = 1; 500 model->voiced = 1;
509 else 501 else
510 model->voiced = 0; 502 model->voiced = 0;
511 503
512 /* post processing, helps clean up some voicing errors ------------------*/ 504 /* post processing, helps clean up some voicing errors ------------------*/
513
514 /*
515 Determine the ratio of low freqency to high frequency energy,
516 voiced speech tends to be dominated by low frequency energy,
517 unvoiced by high frequency. This measure can be used to
518 determine if we have made any gross errors.
519 */
520
521 int l_2000hz = model->L*2000.0/(c2const->Fs/2);
522 int l_4000hz = model->L*4000.0/(c2const->Fs/2);
523 elow = ehigh = 1E-4;
524 for(l=1; l<=l_2000hz; l++) {
525 elow += model->A[l]*model->A[l];
526 }
527 for(l=l_2000hz; l<=l_4000hz; l++) {
528 ehigh += model->A[l]*model->A[l];
529 }
530 eratio = 10.0*log10f(elow/ehigh);
531 505
532 /* Look for Type 1 errors, strongly V speech that has been 506 /*
533 accidentally declared UV */ 507 Determine the ratio of low frequency to high frequency energy,
508 voiced speech tends to be dominated by low frequency energy,
509 unvoiced by high frequency. This measure can be used to
510 determine if we have made any gross errors.
511 */
512
513 int l_2000hz = model->L * 2000.0 / (c2const->Fs / 2);
514 int l_4000hz = model->L * 4000.0 / (c2const->Fs / 2);
515 elow = ehigh = 1E-4;
516 for (l = 1; l <= l_2000hz; l++) {
517 elow += model->A[l] * model->A[l];
518 }
519 for (l = l_2000hz; l <= l_4000hz; l++) {
520 ehigh += model->A[l] * model->A[l];
521 }
522 eratio = 10.0 * log10f(elow / ehigh);
534 523
535 if (model->voiced == 0) 524 /* Look for Type 1 errors, strongly V speech that has been
536 if (eratio > 10.0) 525 accidentally declared UV */
537 model->voiced = 1;
538 526
539 /* Look for Type 2 errors, strongly UV speech that has been 527 if (model->voiced == 0)
540 accidentally declared V */ 528 if (eratio > 10.0) model->voiced = 1;
541 529
542 if (model->voiced == 1) { 530 /* Look for Type 2 errors, strongly UV speech that has been
543 if (eratio < -10.0) 531 accidentally declared V */
544 model->voiced = 0;
545 532
546 /* A common source of Type 2 errors is the pitch estimator 533 if (model->voiced == 1) {
547 gives a low (50Hz) estimate for UV speech, which gives a 534 if (eratio < -10.0) model->voiced = 0;
548 good match with noise due to the close harmoonic spacing.
549 These errors are much more common than people with 50Hz3
550 pitch, so we have just a small eratio threshold. */
551 535
552 sixty = 60.0*TWO_PI/c2const->Fs; 536 /* A common source of Type 2 errors is the pitch estimator
553 if ((eratio < -4.0) && (model->Wo <= sixty)) 537 gives a low (50Hz) estimate for UV speech, which gives a
554 model->voiced = 0; 538 good match with noise due to the close harmoonic spacing.
555 } 539 These errors are much more common than people with 50Hz3
556 //printf(" v: %d snr: %f eratio: %3.2f %f\n",model->voiced,snr,eratio,dF0); 540 pitch, so we have just a small eratio threshold. */
557 541
558 return snr; 542 sixty = 60.0 * TWO_PI / c2const->Fs;
543 if ((eratio < -4.0) && (model->Wo <= sixty)) model->voiced = 0;
544 }
545 // printf(" v: %d snr: %f eratio: %3.2f %f\n",model->voiced,snr,eratio,dF0);
546
547 return snr;
559} 548}
560 549
561/*---------------------------------------------------------------------------*\ 550/*---------------------------------------------------------------------------*\
@@ -564,32 +553,29 @@ float est_voicing_mbe(
564 AUTHOR......: David Rowe 553 AUTHOR......: David Rowe
565 DATE CREATED: 11/5/94 554 DATE CREATED: 11/5/94
566 555
567 Init function that generates the trapezoidal (Parzen) sythesis window. 556 Init function that generates the trapezoidal (Parzen) synthesis window.
568 557
569\*---------------------------------------------------------------------------*/ 558\*---------------------------------------------------------------------------*/
570 559
571void make_synthesis_window(C2CONST *c2const, float Pn[]) 560void make_synthesis_window(C2CONST *c2const, float Pn[]) {
572{ 561 int i;
573 int i;
574 float win; 562 float win;
575 int n_samp = c2const->n_samp; 563 int n_samp = c2const->n_samp;
576 int tw = c2const->tw; 564 int tw = c2const->tw;
577 565
578 /* Generate Parzen window in time domain */ 566 /* Generate Parzen window in time domain */
579 567
580 win = 0.0; 568 win = 0.0;
581 for(i=0; i<n_samp/2-tw; i++) 569 for (i = 0; i < n_samp / 2 - tw; i++) Pn[i] = 0.0;
582 Pn[i] = 0.0;
583 win = 0.0; 570 win = 0.0;
584 for(i=n_samp/2-tw; i<n_samp/2+tw; win+=1.0/(2*tw), i++ ) 571 for (i = n_samp / 2 - tw; i < n_samp / 2 + tw; win += 1.0 / (2 * tw), i++)
585 Pn[i] = win; 572 Pn[i] = win;
586 for(i=n_samp/2+tw; i<3*n_samp/2-tw; i++) 573 for (i = n_samp / 2 + tw; i < 3 * n_samp / 2 - tw; i++) Pn[i] = 1.0;
587 Pn[i] = 1.0;
588 win = 1.0; 574 win = 1.0;
589 for(i=3*n_samp/2-tw; i<3*n_samp/2+tw; win-=1.0/(2*tw), i++) 575 for (i = 3 * n_samp / 2 - tw; i < 3 * n_samp / 2 + tw;
576 win -= 1.0 / (2 * tw), i++)
590 Pn[i] = win; 577 Pn[i] = win;
591 for(i=3*n_samp/2+tw; i<2*n_samp; i++) 578 for (i = 3 * n_samp / 2 + tw; i < 2 * n_samp; i++) Pn[i] = 0.0;
592 Pn[i] = 0.0;
593} 579}
594 580
595/*---------------------------------------------------------------------------*\ 581/*---------------------------------------------------------------------------*\
@@ -600,77 +586,72 @@ void make_synthesis_window(C2CONST *c2const, float Pn[])
600 586
601 Synthesise a speech signal in the frequency domain from the 587 Synthesise a speech signal in the frequency domain from the
602 sinusodal model parameters. Uses overlap-add with a trapezoidal 588 sinusodal model parameters. Uses overlap-add with a trapezoidal
603 window to smoothly interpolate betwen frames. 589 window to smoothly interpolate between frames.
604 590
605\*---------------------------------------------------------------------------*/ 591\*---------------------------------------------------------------------------*/
606 592
607void synthesise( 593void synthesise(int n_samp, codec2_fftr_cfg fftr_inv_cfg,
608 int n_samp, 594 float Sn_[], /* time domain synthesised signal */
609 codec2_fftr_cfg fftr_inv_cfg, 595 MODEL *model, /* ptr to model parameters for this frame */
610 float Sn_[], /* time domain synthesised signal */ 596 float Pn[], /* time domain Parzen window */
611 MODEL *model, /* ptr to model parameters for this frame */ 597 int shift /* flag used to handle transition frames */
612 float Pn[], /* time domain Parzen window */ 598) {
613 int shift /* flag used to handle transition frames */ 599 int i, l, j, b; /* loop variables */
614) 600 COMP Sw_[FFT_DEC / 2 + 1]; /* DFT of synthesised signal */
615{ 601 float sw_[FFT_DEC]; /* synthesised signal */
616 int i,l,j,b; /* loop variables */ 602
617 COMP Sw_[FFT_DEC/2+1]; /* DFT of synthesised signal */ 603 if (shift) {
618 float sw_[FFT_DEC]; /* synthesised signal */ 604 /* Update memories */
619 605 for (i = 0; i < n_samp - 1; i++) {
620 if (shift) { 606 Sn_[i] = Sn_[i + n_samp];
621 /* Update memories */
622 for(i=0; i<n_samp-1; i++) {
623 Sn_[i] = Sn_[i+n_samp];
624 }
625 Sn_[n_samp-1] = 0.0;
626 } 607 }
608 Sn_[n_samp - 1] = 0.0;
609 }
627 610
628 for(i=0; i<FFT_DEC/2+1; i++) { 611 for (i = 0; i < FFT_DEC / 2 + 1; i++) {
629 Sw_[i].real = 0.0; 612 Sw_[i].real = 0.0;
630 Sw_[i].imag = 0.0; 613 Sw_[i].imag = 0.0;
631 } 614 }
632 615
633 /* Now set up frequency domain synthesised speech */ 616 /* Now set up frequency domain synthesised speech */
634 617
635 for(l=1; l<=model->L; l++) { 618 for (l = 1; l <= model->L; l++) {
636 b = (int)(l*model->Wo*FFT_DEC/TWO_PI + 0.5); 619 b = (int)(l * model->Wo * FFT_DEC / TWO_PI + 0.5);
637 if (b > ((FFT_DEC/2)-1)) { 620 if (b > ((FFT_DEC / 2) - 1)) {
638 b = (FFT_DEC/2)-1; 621 b = (FFT_DEC / 2) - 1;
639 }
640 Sw_[b].real = model->A[l]*cosf(model->phi[l]);
641 Sw_[b].imag = model->A[l]*sinf(model->phi[l]);
642 } 622 }
623 Sw_[b].real = model->A[l] * cosf(model->phi[l]);
624 Sw_[b].imag = model->A[l] * sinf(model->phi[l]);
625 }
643 626
644 /* Perform inverse DFT */ 627 /* Perform inverse DFT */
645 628
646 codec2_fftri(fftr_inv_cfg, Sw_,sw_); 629 codec2_fftri(fftr_inv_cfg, Sw_, sw_);
647 630
648 /* Overlap add to previous samples */ 631 /* Overlap add to previous samples */
649 632
650 #ifdef USE_KISS_FFT 633#ifdef USE_KISS_FFT
651 #define FFTI_FACTOR ((float)1.0) 634#define FFTI_FACTOR ((float)1.0)
652 #else 635#else
653 #define FFTI_FACTOR ((float32_t)FFT_DEC) 636#define FFTI_FACTOR ((float32_t)FFT_DEC)
654 #endif 637#endif
655 638
656 for(i=0; i<n_samp-1; i++) { 639 for (i = 0; i < n_samp - 1; i++) {
657 Sn_[i] += sw_[FFT_DEC-n_samp+1+i]*Pn[i] * FFTI_FACTOR; 640 Sn_[i] += sw_[FFT_DEC - n_samp + 1 + i] * Pn[i] * FFTI_FACTOR;
658 } 641 }
659 642
660 if (shift) 643 if (shift)
661 for(i=n_samp-1,j=0; i<2*n_samp; i++,j++) 644 for (i = n_samp - 1, j = 0; i < 2 * n_samp; i++, j++)
662 Sn_[i] = sw_[j]*Pn[i] * FFTI_FACTOR; 645 Sn_[i] = sw_[j] * Pn[i] * FFTI_FACTOR;
663 else 646 else
664 for(i=n_samp-1,j=0; i<2*n_samp; i++,j++) 647 for (i = n_samp - 1, j = 0; i < 2 * n_samp; i++, j++)
665 Sn_[i] += sw_[j]*Pn[i] * FFTI_FACTOR; 648 Sn_[i] += sw_[j] * Pn[i] * FFTI_FACTOR;
666} 649}
667 650
668
669/* todo: this should probably be in some states rather than a static */ 651/* todo: this should probably be in some states rather than a static */
670static unsigned long next = 1; 652static unsigned long next = 1;
671 653
672int codec2_rand(void) { 654int codec2_rand(void) {
673 next = next * 1103515245 + 12345; 655 next = next * 1103515245 + 12345;
674 return((unsigned)(next/65536) % 32768); 656 return ((unsigned)(next / 65536) % 32768);
675} 657}
676
diff --git a/sine.h b/sine.h
index 21d21fb..3f709b2 100644
--- a/sine.h
+++ b/sine.h
@@ -28,20 +28,23 @@
28#ifndef __SINE__ 28#ifndef __SINE__
29#define __SINE__ 29#define __SINE__
30 30
31#include "defines.h"
32#include "comp.h"
33#include "codec2_fft.h" 31#include "codec2_fft.h"
32#include "comp.h"
33#include "defines.h"
34 34
35C2CONST c2const_create(int Fs, float framelength_ms); 35C2CONST c2const_create(int Fs, float framelength_ms);
36 36
37void make_analysis_window(C2CONST *c2const, codec2_fft_cfg fft_fwd_cfg, float w[], float W[]); 37void make_analysis_window(C2CONST *c2const, codec2_fft_cfg fft_fwd_cfg,
38 float w[], float W[]);
38float hpf(float x, float states[]); 39float hpf(float x, float states[]);
39void dft_speech(C2CONST *c2const, codec2_fft_cfg fft_fwd_cfg, COMP Sw[], float Sn[], float w[]); 40void dft_speech(C2CONST *c2const, codec2_fft_cfg fft_fwd_cfg, COMP Sw[],
41 float Sn[], float w[]);
40void two_stage_pitch_refinement(C2CONST *c2const, MODEL *model, COMP Sw[]); 42void two_stage_pitch_refinement(C2CONST *c2const, MODEL *model, COMP Sw[]);
41void estimate_amplitudes(MODEL *model, COMP Sw[], float W[], int est_phase); 43void estimate_amplitudes(MODEL *model, COMP Sw[], float W[], int est_phase);
42float est_voicing_mbe(C2CONST *c2const, MODEL *model, COMP Sw[], float W[]); 44float est_voicing_mbe(C2CONST *c2const, MODEL *model, COMP Sw[], float W[]);
43void make_synthesis_window(C2CONST *c2const, float Pn[]); 45void make_synthesis_window(C2CONST *c2const, float Pn[]);
44void synthesise(int n_samp, codec2_fftr_cfg fftr_inv_cfg, float Sn_[], MODEL *model, float Pn[], int shift); 46void synthesise(int n_samp, codec2_fftr_cfg fftr_inv_cfg, float Sn_[],
47 MODEL *model, float Pn[], int shift);
45 48
46#define CODEC2_RAND_MAX 32767 49#define CODEC2_RAND_MAX 32767
47int codec2_rand(void); 50int codec2_rand(void);
diff --git a/stripdown.sh b/stripdown.sh
index c8d8332..c91694a 100755
--- a/stripdown.sh
+++ b/stripdown.sh
@@ -3,10 +3,10 @@
3BUILDDIR=build_stripped 3BUILDDIR=build_stripped
4DESTDIR=codec2lite 4DESTDIR=codec2lite
5 5
6LIBSRC="codebook.c codebookd.c codebookge.c codebookjvm.c codebooknewamp1.c codebooknewamp1_energy.c codec2.c codec2_fft.c interp.c kiss_fft.c kiss_fftr.c lpc.c lsp.c mbest.c newamp1.c newamp2.c nlp.c pack.c phase.c postfilter.c quantise.c sine.c" 6LIBSRC="codec2.c codec2_fft.c interp.c kiss_fft.c kiss_fftr.c lpc.c lsp.c mbest.c newamp1.c nlp.c pack.c phase.c postfilter.c quantise.c sine.c"
7LIBGENSRC="codebooklspmelvq.c codebookmel.c codebooknewamp2.c codebooknewamp2_energy.c" 7LIBGENSRC="codebook.c codebookd.c codebookge.c codebooknewamp1.c codebooknewamp1_energy.c codebooknewamp2.c codebooknewamp2_energy.c codebookjmv.c"
8 8
9LIBINC="_kiss_fft_guts.h comp_prim.h lpc.h os.h bpf.h lsp.h phase.h bpfb.h defines.h machdep.h postfilter.h mbest.h quantise.h codec2_fft.h interp.h newamp1.h sine.h codec2_internal.h kiss_fft.h newamp2.h comp.h kiss_fftr.h nlp.h" 9LIBINC="_kiss_fft_guts.h comp_prim.h lpc.h os.h bpf.h lsp.h phase.h bpfb.h defines.h machdep.h postfilter.h mbest.h quantise.h codec2_fft.h interp.h newamp1.h sine.h codec2_internal.h kiss_fft.h comp.h kiss_fftr.h nlp.h"
10LIBGENINC="version.h" 10LIBGENINC="version.h"
11 11
12EMPTYINC="c2wideband.h dump.h" 12EMPTYINC="c2wideband.h dump.h"
@@ -36,7 +36,7 @@ done
36 36
37cd ${BUILDDIR} 37cd ${BUILDDIR}
38cmake .. 38cmake ..
39make tnewamp1 39cmake --build . -t codec2
40 40
41# copy generated files 41# copy generated files
42for file in ${LIBGENSRC}; do 42for file in ${LIBGENSRC}; do
@@ -55,10 +55,9 @@ CFLAGS = -Wall -Wno-strict-overflow -std=gnu11 -fPIC -g -O2 -I. -lm
55CFLAGS += -DHORUS_L2_RX -DINTERLEAVER -DRUN_TIME_TABLES -DSCRAMBLER -Dcodec2_EXPORTS 55CFLAGS += -DHORUS_L2_RX -DINTERLEAVER -DRUN_TIME_TABLES -DSCRAMBLER -Dcodec2_EXPORTS
56CFLAGS += -Wno-incompatible-pointer-types-discards-qualifiers 56CFLAGS += -Wno-incompatible-pointer-types-discards-qualifiers
57 57
58LIBSRC=codebook.c codebookd.c codebookge.c codebookmel.c codebooklspmelvq.c codebookjvm.c \ 58LIBSRC=codebook.c codebookd.c codebookge.c codebookjmv.c codebooknewamp1_energy.c \
59 codebooknewamp1.c codebooknewamp1_energy.c codebooknewamp2.c codebooknewamp2_energy.c \ 59 codebooknewamp1.c codec2.c codec2_fft.c interp.c kiss_fft.c kiss_fftr.c \
60 codec2.c codec2_fft.c interp.c kiss_fft.c kiss_fftr.c lpc.c lsp.c mbest.c \ 60 lpc.c lsp.c mbest.c newamp1.c nlp.c pack.c phase.c postfilter.c quantise.c sine.c
61 newamp1.c newamp2.c nlp.c pack.c phase.c postfilter.c quantise.c sine.c
62 61
63all: main 62all: main
64 63
@@ -87,6 +86,8 @@ int main() {
87 int off = 0; 86 int off = 0;
88 unsigned char bits[128]; 87 unsigned char bits[128];
89 88
89 fprintf(stderr, "%d samples per frame, %d bits per frame\n", nsam, nbit);
90
90 fread(input, 976692, 1, f); 91 fread(input, 976692, 1, f);
91 fclose(f); 92 fclose(f);
92 93
diff --git a/version.h b/version.h
index bae2151..459e241 100644
--- a/version.h
+++ b/version.h
@@ -29,9 +29,9 @@
29#ifndef CODEC2_HAVE_VERSION 29#ifndef CODEC2_HAVE_VERSION
30#define CODEC2_HAVE_VERSION 30#define CODEC2_HAVE_VERSION
31 31
32#define CODEC2_VERSION_MAJOR 0 32#define CODEC2_VERSION_MAJOR 1
33#define CODEC2_VERSION_MINOR 9 33#define CODEC2_VERSION_MINOR 2
34#define CODEC2_VERSION_PATCH 2 34/* #undef CODEC2_VERSION_PATCH */
35#define CODEC2_VERSION "0.9.2" 35#define CODEC2_VERSION "1.2.0"
36 36
37#endif //CODEC2_HAVE_VERSION 37#endif //CODEC2_HAVE_VERSION
generated by cgit v1.2.3 (git 2.25.1) at 2025-08-29 23:47:07 +0000