1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
|
function y = G729( x, Fs, Nw, Nsh )
%G729 Summary of this function goes here
% Detailed explanation goes here
% frameSize = Nw-Nsh;
VADPar = InitVADPar(Fs,Nw,Nsh);
numFrame = floor((length(x)-Nw)/Nsh+1);
y = zeros(size(x));
x_start = 1;
x_end = Nsh; % the window shift logic is already supported by G729 implementation
for j=1:numFrame
[v,VADPar] = VAD(x(x_start:x_end), VADPar);
y(x_start:x_end) = v;
x_start = x_start+Nsh;
x_end = min(x_end+Nsh,length(x));
end
%
% function [y, new_params] = G729( x, params, Fs, window, frameSize )
%
% % frameSize = window-winshift;
% if isempty(params)
% if nargin < 5, error('not enough arguments'); end
% params = InitVADPar(Fs,window,frameSize);
% end
% % numFrame = floor(length(x)/frameSize);
%
% % y = zeros(size(x));
%
% % x_start = 1;
% % x_end = frameSize; % the window shift logic is already supported by G729 implementation
%
% % for j=1:numFrame
% [y, new_params] = VAD(x, params);
% % y(x_start:x_end) = v;
% % x_start = x_start+frameSize;
% % x_end = x_end +frameSize;
% % end
function VADPar = InitVADPar(Fs,Nw,Nsh)
% initialize constant parameters
VADPar.M = 10; % LP order
VADPar.NP = 12; % autocorrelation order
VADPar.N0 = 128; % number of frames for long-term min energy calculation
VADPar.Ni = 32; % number of frames for initialization of running averages
VADPar.INIT_COUNT = 20;
% HPFilt is a HPF that is used to preprocess the signal applied to the VAD.
% 140 Hz cutoff, unity gain near 200 Hz, falling to 0.971 at high freq.
VADPar.HPFilt.b = [ 0.92727435, -1.8544941, 0.92727435 ];
VADPar.HPFilt.a = [ 1, -1.9059465, 0.91140240 ];
VADPar.HPFilt.Mem = [];
VADPar.N = Nw; % window size
VADPar.LA = 40; % Look-ahead
VADPar.NF = Nsh; % Frame size
LWmem = VADPar.N - VADPar.NF;
VADPar.Wmem = zeros(LWmem, 1);
LA = VADPar.LA;
LB = VADPar.N - VADPar.LA;
VADPar.Window = [0.54 - 0.46*cos(2*pi*(0:LB-1)'/(2*LB-1));
cos(2*pi*(0:LA-1)'/(4*LA-1))];
% LP analysis, lag window applied to autocorrelation coefficients
% Fs = 8000;
BWExp = 60; % 60 Hz bandwidth expansion, Gaussian window
w0 = 2 * pi * BWExp / Fs;
NP = VADPar.NP;
Wn = 1.0001; % White noise compensation (diagonal loading)
VADPar.LagWindow = [Wn; exp(-0.5 * (w0 * (1:NP)').^2)] / Wn;
% Correlation for a lowpass filter (3 dB point on the power spectrum is
% at about 2 kHz)
VADPar.LBF_CORR = ...
[ 0.24017939691329, 0.21398822343783, 0.14767692339633, ...
0.07018811903116, 0.00980856433051,-0.02015934721195, ...
-0.02388269958005,-0.01480076155002,-0.00503292155509, ...
0.00012141366508, 0.00119354245231, 0.00065908718613, ...
0.00015015782285]';
% initialize variable parameters
VADPar.FrmCount = 0;
VADPar.FrmEn = Inf * ones(1,VADPar.N0);
VADPar.MeanLSF = zeros(VADPar.M, 1);
VADPar.MeanSE = 0;
VADPar.MeanSLE = 0;
VADPar.MeanE = 0;
VADPar.MeanSZC = 0;
VADPar.count_sil = 0;
VADPar.count_inert = 0; % modified for AppendixII
VADPar.count_update = 0;
VADPar.count_ext = 0;
VADPar.less_count = 0;
VADPar.flag = 1;
VADPar.PrevMarkers = [1, 1];
VADPar.PrevEnergy = 0;
VADPar.Prev_MinE = Inf;
VADPar.Next_MinE = Inf;
VADPar.MinE_buffer = Inf * ones(1, VADPar.N0/8);
return
function [Ivd, VADPar, v_flag] = VAD (x_new, VADPar)
% The Matlab routine implements the Voice Activity Detector (VAD) for
% the ITU-T G.729 coder. The VAD is specified in G.729B (annex B to
% G.729) to accompany G.729A the low complexity version of the G.729 coder.
% There is a modification to the VAD given in Appendix II (G.729II).
%
% The reference code for G.729A, G.729B, and G.729II uses fixed point
% arithmetic. However, G.729C+ includes reference code in floating point
% for both the coder and the VAD. This Matlab routine in double precision
% floating point borrows the relevant parts from the Annex C+ floating
% point code, but retains the decision logic of Appendix II. A switch is
% available to disable the Appendix II modifications.
% The VAD uses the preprocessed speech (highpass filtered) and the linear
% predictive parameters from the coder. The Matlab code here is standalone
% and so includes the preprocessing and the LP analysis.
% Tests on this VAD show a match to the G.729C+ VAD decisions (with the
% Appendix II option turned off).
% P. Kabal 2008-04-03
% Ivd - VAD flag, 0 no speech, 1 speech
% VADPar - Updated parameter structure
% v_flag - one during hangover (only for VAD_APPENDIX_II = 0)
VAD_APPENDIX_II = 1;
% Constants
N = VADPar.N; % window size
N0 = VADPar.N0; % number of frames used for long-term minimum energy calculation
Ni = VADPar.Ni; % number of frames used for initialization of running averages
INIT_COUNT = VADPar.INIT_COUNT;
NOISE = 0;
VOICE = 1;
v_flag = 0;
VADPar.FrmCount = VADPar.FrmCount + 1;
frm_count = VADPar.FrmCount;
% Filter new data (HP filter)
[x_new_hp, VADPar.HPFilt.Mem] = filter(VADPar.HPFilt.b, VADPar.HPFilt.a, ...
32768 * x_new, VADPar.HPFilt.Mem);
% Append new filtered data to filter memory
xwin = [VADPar.Wmem; x_new_hp];
% LPC analysis
[r, LSF, rc2] = VADLPAnalysis(xwin, VADPar);
% Full band energy
Ef = 10*log10(r(1) / N);
% Low band energy
Elow = r(1) * VADPar.LBF_CORR(1) ...
+ 2 * sum(r(2:end) .* VADPar.LBF_CORR(2:end));
El = 10*log10(Elow / N);
% Compute SD
SD = sum((LSF-VADPar.MeanLSF).^2);
% Normalized zero-crossing rate (in current frame)
ist = VADPar.N - VADPar.LA - VADPar.NF + 1; % Current frame start
ifn = ist + VADPar.NF - 1; % Current frame end
ZC = zcr(xwin(ist:ifn+1));
% The next steps involve finding the minimum energy in the N0 frames.
% The original code in G.729 is very convoluted. The Matlab code below
% mimics the operation with a simpler structure.
% - To reduce computations, the minimum energy for blocks of 8 samples
% is determined. These values are stored in a buffer of length N0/8.
% The buffer is updated whenever the frame count is a multiple of 8.
% Starting at the beginning, the minimum of the frames 1-8 is stored
% into the buffer in frame 8, the minimum of the frames 9-16 is stored
% into the buffer at frame 16, etc.
% - Prev_Min is the minimum of the values stored in the buffer, effectively
% the minimum of N0 energy values.
% - Next_Min is the minimum used to determine the minimum of the next
% 8 samples.
% - MinE is min(Prev_Min, Next_Min).
% - Note that that for frame count equal to a multiple of 8, Next_Min is
% updated and MinE is updated before updating the buffer. This means
% that MinE is calculated over N0+8 values. MinE is effectively
% calculated over a varying window length (N0+1 to N0+8). It is
% nonincreasing while the window length increases.
% - The value of Min will not be used until frame N0.
% Long-term minimum energy
VADPar.Next_MinE = min(Ef, VADPar.Next_MinE);
MinE = min(VADPar.Prev_MinE, VADPar.Next_MinE);
if (mod(frm_count, 8) == 0)
VADPar.MinE_buffer = [VADPar.MinE_buffer(2:end), VADPar.Next_MinE];
VADPar.Prev_MinE = min(VADPar.MinE_buffer);
VADPar.Next_MinE = Inf;
end
% Initialization of running averages
if (frm_count <= Ni)
if (Ef < 21)
VADPar.less_count = VADPar.less_count + 1;
marker = NOISE;
else
marker = VOICE;
NEp = (frm_count - 1) - VADPar.less_count;
NE = NEp + 1;
VADPar.MeanE = (VADPar.MeanE * NEp + Ef) / NE;
VADPar.MeanSZC = (VADPar.MeanSZC * NEp + ZC) / NE;
VADPar.MeanLSF = (VADPar.MeanLSF * NEp + LSF) / NE;
end
end
if (frm_count >= Ni)
if (frm_count == Ni)
if (VAD_APPENDIX_II)
if (VADPar.less_count >= Ni) % modified for Appendix II
VADPar.FrmCount = 0;
frm_count = VADPar.FrmCount;
VADPar.less_count = 0;
end
end
VADPar.MeanSE = VADPar.MeanE - 10;
VADPar.MeanSLE = VADPar.MeanE - 12;
end
dSE = VADPar.MeanSE - Ef;
dSLE = VADPar.MeanSLE - El;
dSZC = VADPar.MeanSZC - ZC;
if (Ef < 21)
marker = NOISE;
else
marker = MakeDec(dSLE, dSE, SD, dSZC);
end
if (VAD_APPENDIX_II)
if (marker == VOICE) % modified for Appendix II
VADPar.count_inert = 0;
end
if (marker == NOISE && VADPar.count_inert < 6)
VADPar.count_inert = VADPar.count_inert + 1;
marker = VOICE;
end
else
v_flag = 0;
end
% Voice activity decision smoothing: Step 1
if (VADPar.PrevMarkers(1) == VOICE && marker == NOISE ...
&& Ef > VADPar.MeanSE + 2 && Ef > 21)
marker = VOICE;
if (~VAD_APPENDIX_II)
v_flag = 1;
end
end
% Voice activity decision smoothing: Step 2
if (VADPar.flag == 1)
if (VADPar.PrevMarkers(2) == VOICE ...
&& VADPar.PrevMarkers(1) == VOICE ...
&& marker == NOISE ...
&& abs(Ef - VADPar.PrevEnergy) <= 3)
VADPar.count_ext = VADPar.count_ext + 1;
marker = VOICE;
if(~ VAD_APPENDIX_II)
v_flag = 1;
end
if (VADPar.count_ext <= 4)
VADPar.flag = 1;
else
VADPar.flag = 0;
VADPar.count_ext = 0;
end
end
else
VADPar.flag = 1;
end
% For unvoiced case, count_sil is incremented
if (marker == NOISE)
VADPar.count_sil = VADPar.count_sil + 1;
end
% Voice activity decision smoothing: Step 3
if (marker == VOICE && VADPar.count_sil > 10 ...
&& Ef - VADPar.PrevEnergy <= 3)
marker = NOISE;
VADPar.count_sil = 0;
if (VAD_APPENDIX_II)
VADPar.count_inert = 6; % modified for AppendixII
end
end
if (marker == VOICE)
VADPar.count_sil = 0;
end
% Voice activity decision smoothing: Step 4
if (~VAD_APPENDIX_II)
if (Ef < VADPar.MeanSE + 3 && VADPar.FrmCount > N0 ...
&& v_flag == 0 && rc2 < 0.6)
marker = NOISE;
end
end
if (VAD_APPENDIX_II)
TestC = (Ef < VADPar.MeanSE + 3 && rc2 < 0.75); % Appendix II
else
TestC = (Ef < VADPar.MeanSE + 3 && rc2 < 0.75 && SD < 0.002532959);
end
if (TestC)
VADPar.count_update = VADPar.count_update + 1;
% Modify update speed coefficients
if (VADPar.count_update < INIT_COUNT)
COEF = 0.75;
COEFZC = 0.8;
COEFSD = 0.6;
elseif (VADPar.count_update < INIT_COUNT + 10)
COEF = 0.95;
COEFZC = 0.92;
COEFSD = 0.65;
elseif (VADPar.count_update < INIT_COUNT + 20)
COEF = 0.97;
COEFZC = 0.94;
COEFSD = 0.70;
elseif (VADPar.count_update < INIT_COUNT + 30)
COEF = 0.99;
COEFZC = 0.96;
COEFSD = 0.75;
elseif (VADPar.count_update < INIT_COUNT + 40)
COEF = 0.995;
COEFZC = 0.99;
COEFSD = 0.75;
else
COEF = 0.995;
COEFZC = 0.998;
COEFSD = 0.75;
end
% Update mean LSF, SE, SLE, SZC
VADPar.MeanLSF = COEFSD * VADPar.MeanLSF + (1-COEFSD) * LSF;
VADPar.MeanSE = COEF * VADPar.MeanSE + (1-COEF) * Ef;
VADPar.MeanSLE = COEF * VADPar.MeanSLE + (1-COEF) * El;
VADPar.MeanSZC = COEFZC * VADPar.MeanSZC + (1-COEFZC) * ZC;
end
if (frm_count > N0 && ...
(VADPar.MeanSE < MinE && SD < 0.002532959) ...
|| VADPar.MeanSE > MinE + 10 )
VADPar.MeanSE = MinE;
VADPar.count_update = 0;
end
end
VADPar.PrevEnergy = Ef;
VADPar.PrevMarkers = [marker, VADPar.PrevMarkers(1)];
ist = VADPar.NF + 1;
VADPar.Wmem = xwin(ist:end);
Ivd = marker;
return
% ----- ----- ----- -----
function dec = MakeDec(dSLE, dSE, SD, dSZC)
a = [0.00175, -0.004545455, -25, 20, 0, ...
8800, 0, 25, -29.09091, 0, ...
14000, 0.928571, -1.5, 0.714285];
b = [0.00085, 0.001159091, -5, -6, -4.7, ...
-12.2, 0.0009, -7.0, -4.8182, -5.3, ...
-15.5, 1.14285, -9, -2.1428571];
dec = 0;
% SD vs dSZC
if SD > a(1)*dSZC+b(1)
dec = 1;
return;
end
if SD > a(2)*dSZC+b(2)
dec = 1;
return;
end
% dSE vs dSZC
if dSE < a(3)*dSZC+b(3)
dec = 1;
return;
end
if dSE < a(4)*dSZC+b(4)
dec = 1;
return;
end
if dSE < b(5)
dec = 1;
return;
end
% dSE vs SD
if dSE < a(6)*SD+b(6)
dec = 1;
return;
end
if SD > b(7)
dec = 1;
return;
end
% dSLE vs dSZC
if dSLE < a(8)*dSZC+b(8)
dec = 1;
return;
end
if dSLE < a(9)*dSZC+b(9)
dec = 1;
return;
end
if dSLE < b(10)
dec = 1;
return;
end
% dSLE vs SD
if dSLE < a(11)*SD+b(11)
dec = 1;
return;
end
% dSLE vs dSE
if dSLE > a(12)*dSE+b(12)
dec = 1;
return
end
if dSLE < a(13)*dSE+b(13)
dec = 1;
return;
end
if dSLE < a(14)*dSE+b(14)
dec = 1;
return;
end
return
% ----- ----- ----- -----
function [zc] = zcr (x)
% Calculate normalized (per sample) zero-crossing rate
% Input is the frame plus the first sample of the next
% frame.
M = length(x) - 1;
x1 = x(1:end-1);
x2 = x(2:end);
xp = x1 .* x2;
I = (xp < 0);
%sign1 = sign(x);
%sign2 = sign([mem; x(1:M-1)]);
%
%zc = 1/(2*M)*sum(abs(sign1-sign2));
zc = sum(I) / M;
return
% -----------------------------
function [r, LSF, rc2] = VADLPAnalysis (x, VADPar)
M = VADPar.M; % LP order
NP = VADPar.NP; % autocorrelation order
% Apply window to input frame
xw = VADPar.Window .* x;
% Compute autocorrelation
r = acorr(xw, NP+1) .* VADPar.LagWindow;
% Compute normalized LSF
A = ac2poly(r(1:M+1));
LSF = poly2lsf(A) / (2 * pi); % normalized to 0 to 0.5
% Reflection coefficients
rc = ac2rc(r(1:3));
rc2 = rc(2);
return
% -----------------------------
function rxx = acorr (x, Nt)
Nx = length (x);
N = Nt;
if (Nt > Nx)
N = Nx;
end
rxx = zeros(Nt, 1);
for (i = 0:N-1)
Nv = Nx - i;
rxx(i+1) = x(1:Nv)' * x(i+1:i+Nv);
end
return |
G729.m Algorithme probleme
Répondre avec citation
Partager