media/libopus/silk/VAD.c@b8a032363ba2
media/libopus/silk/VAD.c
Thu, 22 Jan 2015 13:21:57 +0100
- author
- Michael Schloh von Bennewitz <michael@schloh.com>
- date
- Thu, 22 Jan 2015 13:21:57 +0100
- branch
- TOR_BUG_9701
- changeset 15
- b8a032363ba2
- permissions
- -rw-r--r--
Incorporate requested changes from Mozilla in review:
https://bugzilla.mozilla.org/show_bug.cgi?id=1123480#c6
1 /***********************************************************************
2 Copyright (c) 2006-2011, Skype Limited. All rights reserved.
3 Redistribution and use in source and binary forms, with or without
4 modification, are permitted provided that the following conditions
5 are met:
6 - Redistributions of source code must retain the above copyright notice,
7 this list of conditions and the following disclaimer.
8 - Redistributions in binary form must reproduce the above copyright
9 notice, this list of conditions and the following disclaimer in the
10 documentation and/or other materials provided with the distribution.
11 - Neither the name of Internet Society, IETF or IETF Trust, nor the
12 names of specific contributors, may be used to endorse or promote
13 products derived from this software without specific prior written
14 permission.
15 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16 AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19 LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20 CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21 SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22 INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23 CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24 ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25 POSSIBILITY OF SUCH DAMAGE.
26 ***********************************************************************/
28 #ifdef HAVE_CONFIG_H
29 #include "config.h"
30 #endif
32 #include "main.h"
33 #include "stack_alloc.h"
35 /* Silk VAD noise level estimation */
36 static OPUS_INLINE void silk_VAD_GetNoiseLevels(
37 const opus_int32 pX[ VAD_N_BANDS ], /* I subband energies */
38 silk_VAD_state *psSilk_VAD /* I/O Pointer to Silk VAD state */
39 );
41 /**********************************/
42 /* Initialization of the Silk VAD */
43 /**********************************/
44 opus_int silk_VAD_Init( /* O Return value, 0 if success */
45 silk_VAD_state *psSilk_VAD /* I/O Pointer to Silk VAD state */
46 )
47 {
48 opus_int b, ret = 0;
50 /* reset state memory */
51 silk_memset( psSilk_VAD, 0, sizeof( silk_VAD_state ) );
53 /* init noise levels */
54 /* Initialize array with approx pink noise levels (psd proportional to inverse of frequency) */
55 for( b = 0; b < VAD_N_BANDS; b++ ) {
56 psSilk_VAD->NoiseLevelBias[ b ] = silk_max_32( silk_DIV32_16( VAD_NOISE_LEVELS_BIAS, b + 1 ), 1 );
57 }
59 /* Initialize state */
60 for( b = 0; b < VAD_N_BANDS; b++ ) {
61 psSilk_VAD->NL[ b ] = silk_MUL( 100, psSilk_VAD->NoiseLevelBias[ b ] );
62 psSilk_VAD->inv_NL[ b ] = silk_DIV32( silk_int32_MAX, psSilk_VAD->NL[ b ] );
63 }
64 psSilk_VAD->counter = 15;
66 /* init smoothed energy-to-noise ratio*/
67 for( b = 0; b < VAD_N_BANDS; b++ ) {
68 psSilk_VAD->NrgRatioSmth_Q8[ b ] = 100 * 256; /* 100 * 256 --> 20 dB SNR */
69 }
71 return( ret );
72 }
74 /* Weighting factors for tilt measure */
75 static const opus_int32 tiltWeights[ VAD_N_BANDS ] = { 30000, 6000, -12000, -12000 };
77 /***************************************/
78 /* Get the speech activity level in Q8 */
79 /***************************************/
80 opus_int silk_VAD_GetSA_Q8( /* O Return value, 0 if success */
81 silk_encoder_state *psEncC, /* I/O Encoder state */
82 const opus_int16 pIn[] /* I PCM input */
83 )
84 {
85 opus_int SA_Q15, pSNR_dB_Q7, input_tilt;
86 opus_int decimated_framelength1, decimated_framelength2;
87 opus_int decimated_framelength;
88 opus_int dec_subframe_length, dec_subframe_offset, SNR_Q7, i, b, s;
89 opus_int32 sumSquared, smooth_coef_Q16;
90 opus_int16 HPstateTmp;
91 VARDECL( opus_int16, X );
92 opus_int32 Xnrg[ VAD_N_BANDS ];
93 opus_int32 NrgToNoiseRatio_Q8[ VAD_N_BANDS ];
94 opus_int32 speech_nrg, x_tmp;
95 opus_int X_offset[ VAD_N_BANDS ];
96 opus_int ret = 0;
97 silk_VAD_state *psSilk_VAD = &psEncC->sVAD;
98 SAVE_STACK;
100 /* Safety checks */
101 silk_assert( VAD_N_BANDS == 4 );
102 silk_assert( MAX_FRAME_LENGTH >= psEncC->frame_length );
103 silk_assert( psEncC->frame_length <= 512 );
104 silk_assert( psEncC->frame_length == 8 * silk_RSHIFT( psEncC->frame_length, 3 ) );
106 /***********************/
107 /* Filter and Decimate */
108 /***********************/
109 decimated_framelength1 = silk_RSHIFT( psEncC->frame_length, 1 );
110 decimated_framelength2 = silk_RSHIFT( psEncC->frame_length, 2 );
111 decimated_framelength = silk_RSHIFT( psEncC->frame_length, 3 );
112 /* Decimate into 4 bands:
113 0 L 3L L 3L 5L
114 - -- - -- --
115 8 8 2 4 4
117 [0-1 kHz| temp. |1-2 kHz| 2-4 kHz | 4-8 kHz |
119 They're arranged to allow the minimal ( frame_length / 4 ) extra
120 scratch space during the downsampling process */
121 X_offset[ 0 ] = 0;
122 X_offset[ 1 ] = decimated_framelength + decimated_framelength2;
123 X_offset[ 2 ] = X_offset[ 1 ] + decimated_framelength;
124 X_offset[ 3 ] = X_offset[ 2 ] + decimated_framelength2;
125 ALLOC( X, X_offset[ 3 ] + decimated_framelength1, opus_int16 );
127 /* 0-8 kHz to 0-4 kHz and 4-8 kHz */
128 silk_ana_filt_bank_1( pIn, &psSilk_VAD->AnaState[ 0 ],
129 X, &X[ X_offset[ 3 ] ], psEncC->frame_length );
131 /* 0-4 kHz to 0-2 kHz and 2-4 kHz */
132 silk_ana_filt_bank_1( X, &psSilk_VAD->AnaState1[ 0 ],
133 X, &X[ X_offset[ 2 ] ], decimated_framelength1 );
135 /* 0-2 kHz to 0-1 kHz and 1-2 kHz */
136 silk_ana_filt_bank_1( X, &psSilk_VAD->AnaState2[ 0 ],
137 X, &X[ X_offset[ 1 ] ], decimated_framelength2 );
139 /*********************************************/
140 /* HP filter on lowest band (differentiator) */
141 /*********************************************/
142 X[ decimated_framelength - 1 ] = silk_RSHIFT( X[ decimated_framelength - 1 ], 1 );
143 HPstateTmp = X[ decimated_framelength - 1 ];
144 for( i = decimated_framelength - 1; i > 0; i-- ) {
145 X[ i - 1 ] = silk_RSHIFT( X[ i - 1 ], 1 );
146 X[ i ] -= X[ i - 1 ];
147 }
148 X[ 0 ] -= psSilk_VAD->HPstate;
149 psSilk_VAD->HPstate = HPstateTmp;
151 /*************************************/
152 /* Calculate the energy in each band */
153 /*************************************/
154 for( b = 0; b < VAD_N_BANDS; b++ ) {
155 /* Find the decimated framelength in the non-uniformly divided bands */
156 decimated_framelength = silk_RSHIFT( psEncC->frame_length, silk_min_int( VAD_N_BANDS - b, VAD_N_BANDS - 1 ) );
158 /* Split length into subframe lengths */
159 dec_subframe_length = silk_RSHIFT( decimated_framelength, VAD_INTERNAL_SUBFRAMES_LOG2 );
160 dec_subframe_offset = 0;
162 /* Compute energy per sub-frame */
163 /* initialize with summed energy of last subframe */
164 Xnrg[ b ] = psSilk_VAD->XnrgSubfr[ b ];
165 for( s = 0; s < VAD_INTERNAL_SUBFRAMES; s++ ) {
166 sumSquared = 0;
167 for( i = 0; i < dec_subframe_length; i++ ) {
168 /* The energy will be less than dec_subframe_length * ( silk_int16_MIN / 8 ) ^ 2. */
169 /* Therefore we can accumulate with no risk of overflow (unless dec_subframe_length > 128) */
170 x_tmp = silk_RSHIFT(
171 X[ X_offset[ b ] + i + dec_subframe_offset ], 3 );
172 sumSquared = silk_SMLABB( sumSquared, x_tmp, x_tmp );
174 /* Safety check */
175 silk_assert( sumSquared >= 0 );
176 }
178 /* Add/saturate summed energy of current subframe */
179 if( s < VAD_INTERNAL_SUBFRAMES - 1 ) {
180 Xnrg[ b ] = silk_ADD_POS_SAT32( Xnrg[ b ], sumSquared );
181 } else {
182 /* Look-ahead subframe */
183 Xnrg[ b ] = silk_ADD_POS_SAT32( Xnrg[ b ], silk_RSHIFT( sumSquared, 1 ) );
184 }
186 dec_subframe_offset += dec_subframe_length;
187 }
188 psSilk_VAD->XnrgSubfr[ b ] = sumSquared;
189 }
191 /********************/
192 /* Noise estimation */
193 /********************/
194 silk_VAD_GetNoiseLevels( &Xnrg[ 0 ], psSilk_VAD );
196 /***********************************************/
197 /* Signal-plus-noise to noise ratio estimation */
198 /***********************************************/
199 sumSquared = 0;
200 input_tilt = 0;
201 for( b = 0; b < VAD_N_BANDS; b++ ) {
202 speech_nrg = Xnrg[ b ] - psSilk_VAD->NL[ b ];
203 if( speech_nrg > 0 ) {
204 /* Divide, with sufficient resolution */
205 if( ( Xnrg[ b ] & 0xFF800000 ) == 0 ) {
206 NrgToNoiseRatio_Q8[ b ] = silk_DIV32( silk_LSHIFT( Xnrg[ b ], 8 ), psSilk_VAD->NL[ b ] + 1 );
207 } else {
208 NrgToNoiseRatio_Q8[ b ] = silk_DIV32( Xnrg[ b ], silk_RSHIFT( psSilk_VAD->NL[ b ], 8 ) + 1 );
209 }
211 /* Convert to log domain */
212 SNR_Q7 = silk_lin2log( NrgToNoiseRatio_Q8[ b ] ) - 8 * 128;
214 /* Sum-of-squares */
215 sumSquared = silk_SMLABB( sumSquared, SNR_Q7, SNR_Q7 ); /* Q14 */
217 /* Tilt measure */
218 if( speech_nrg < ( (opus_int32)1 << 20 ) ) {
219 /* Scale down SNR value for small subband speech energies */
220 SNR_Q7 = silk_SMULWB( silk_LSHIFT( silk_SQRT_APPROX( speech_nrg ), 6 ), SNR_Q7 );
221 }
222 input_tilt = silk_SMLAWB( input_tilt, tiltWeights[ b ], SNR_Q7 );
223 } else {
224 NrgToNoiseRatio_Q8[ b ] = 256;
225 }
226 }
228 /* Mean-of-squares */
229 sumSquared = silk_DIV32_16( sumSquared, VAD_N_BANDS ); /* Q14 */
231 /* Root-mean-square approximation, scale to dBs, and write to output pointer */
232 pSNR_dB_Q7 = (opus_int16)( 3 * silk_SQRT_APPROX( sumSquared ) ); /* Q7 */
234 /*********************************/
235 /* Speech Probability Estimation */
236 /*********************************/
237 SA_Q15 = silk_sigm_Q15( silk_SMULWB( VAD_SNR_FACTOR_Q16, pSNR_dB_Q7 ) - VAD_NEGATIVE_OFFSET_Q5 );
239 /**************************/
240 /* Frequency Tilt Measure */
241 /**************************/
242 psEncC->input_tilt_Q15 = silk_LSHIFT( silk_sigm_Q15( input_tilt ) - 16384, 1 );
244 /**************************************************/
245 /* Scale the sigmoid output based on power levels */
246 /**************************************************/
247 speech_nrg = 0;
248 for( b = 0; b < VAD_N_BANDS; b++ ) {
249 /* Accumulate signal-without-noise energies, higher frequency bands have more weight */
250 speech_nrg += ( b + 1 ) * silk_RSHIFT( Xnrg[ b ] - psSilk_VAD->NL[ b ], 4 );
251 }
253 /* Power scaling */
254 if( speech_nrg <= 0 ) {
255 SA_Q15 = silk_RSHIFT( SA_Q15, 1 );
256 } else if( speech_nrg < 32768 ) {
257 if( psEncC->frame_length == 10 * psEncC->fs_kHz ) {
258 speech_nrg = silk_LSHIFT_SAT32( speech_nrg, 16 );
259 } else {
260 speech_nrg = silk_LSHIFT_SAT32( speech_nrg, 15 );
261 }
263 /* square-root */
264 speech_nrg = silk_SQRT_APPROX( speech_nrg );
265 SA_Q15 = silk_SMULWB( 32768 + speech_nrg, SA_Q15 );
266 }
268 /* Copy the resulting speech activity in Q8 */
269 psEncC->speech_activity_Q8 = silk_min_int( silk_RSHIFT( SA_Q15, 7 ), silk_uint8_MAX );
271 /***********************************/
272 /* Energy Level and SNR estimation */
273 /***********************************/
274 /* Smoothing coefficient */
275 smooth_coef_Q16 = silk_SMULWB( VAD_SNR_SMOOTH_COEF_Q18, silk_SMULWB( (opus_int32)SA_Q15, SA_Q15 ) );
277 if( psEncC->frame_length == 10 * psEncC->fs_kHz ) {
278 smooth_coef_Q16 >>= 1;
279 }
281 for( b = 0; b < VAD_N_BANDS; b++ ) {
282 /* compute smoothed energy-to-noise ratio per band */
283 psSilk_VAD->NrgRatioSmth_Q8[ b ] = silk_SMLAWB( psSilk_VAD->NrgRatioSmth_Q8[ b ],
284 NrgToNoiseRatio_Q8[ b ] - psSilk_VAD->NrgRatioSmth_Q8[ b ], smooth_coef_Q16 );
286 /* signal to noise ratio in dB per band */
287 SNR_Q7 = 3 * ( silk_lin2log( psSilk_VAD->NrgRatioSmth_Q8[b] ) - 8 * 128 );
288 /* quality = sigmoid( 0.25 * ( SNR_dB - 16 ) ); */
289 psEncC->input_quality_bands_Q15[ b ] = silk_sigm_Q15( silk_RSHIFT( SNR_Q7 - 16 * 128, 4 ) );
290 }
292 RESTORE_STACK;
293 return( ret );
294 }
296 /**************************/
297 /* Noise level estimation */
298 /**************************/
299 static OPUS_INLINE void silk_VAD_GetNoiseLevels(
300 const opus_int32 pX[ VAD_N_BANDS ], /* I subband energies */
301 silk_VAD_state *psSilk_VAD /* I/O Pointer to Silk VAD state */
302 )
303 {
304 opus_int k;
305 opus_int32 nl, nrg, inv_nrg;
306 opus_int coef, min_coef;
308 /* Initially faster smoothing */
309 if( psSilk_VAD->counter < 1000 ) { /* 1000 = 20 sec */
310 min_coef = silk_DIV32_16( silk_int16_MAX, silk_RSHIFT( psSilk_VAD->counter, 4 ) + 1 );
311 } else {
312 min_coef = 0;
313 }
315 for( k = 0; k < VAD_N_BANDS; k++ ) {
316 /* Get old noise level estimate for current band */
317 nl = psSilk_VAD->NL[ k ];
318 silk_assert( nl >= 0 );
320 /* Add bias */
321 nrg = silk_ADD_POS_SAT32( pX[ k ], psSilk_VAD->NoiseLevelBias[ k ] );
322 silk_assert( nrg > 0 );
324 /* Invert energies */
325 inv_nrg = silk_DIV32( silk_int32_MAX, nrg );
326 silk_assert( inv_nrg >= 0 );
328 /* Less update when subband energy is high */
329 if( nrg > silk_LSHIFT( nl, 3 ) ) {
330 coef = VAD_NOISE_LEVEL_SMOOTH_COEF_Q16 >> 3;
331 } else if( nrg < nl ) {
332 coef = VAD_NOISE_LEVEL_SMOOTH_COEF_Q16;
333 } else {
334 coef = silk_SMULWB( silk_SMULWW( inv_nrg, nl ), VAD_NOISE_LEVEL_SMOOTH_COEF_Q16 << 1 );
335 }
337 /* Initially faster smoothing */
338 coef = silk_max_int( coef, min_coef );
340 /* Smooth inverse energies */
341 psSilk_VAD->inv_NL[ k ] = silk_SMLAWB( psSilk_VAD->inv_NL[ k ], inv_nrg - psSilk_VAD->inv_NL[ k ], coef );
342 silk_assert( psSilk_VAD->inv_NL[ k ] >= 0 );
344 /* Compute noise level by inverting again */
345 nl = silk_DIV32( silk_int32_MAX, psSilk_VAD->inv_NL[ k ] );
346 silk_assert( nl >= 0 );
348 /* Limit noise levels (guarantee 7 bits of head room) */
349 nl = silk_min( nl, 0x00FFFFFF );
351 /* Store as part of state */
352 psSilk_VAD->NL[ k ] = nl;
353 }
355 /* Increment frame counter */
356 psSilk_VAD->counter++;
357 }
