Andrew Top | 8b6b16e | 2018-07-25 17:44:41 -0700 | [diff] [blame] | 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 | ***********************************************************************/ |
| 27 | |
| 28 | #ifdef HAVE_CONFIG_H |
| 29 | #include "config.h" |
| 30 | #endif |
| 31 | |
| 32 | #include "main.h" |
| 33 | #include "stack_alloc.h" |
| 34 | |
| 35 | /* Silk VAD noise level estimation */ |
| 36 | # if !defined(OPUS_X86_MAY_HAVE_SSE4_1) |
| 37 | static OPUS_INLINE void silk_VAD_GetNoiseLevels( |
| 38 | const opus_int32 pX[ VAD_N_BANDS ], /* I subband energies */ |
| 39 | silk_VAD_state *psSilk_VAD /* I/O Pointer to Silk VAD state */ |
| 40 | ); |
| 41 | #endif |
| 42 | |
| 43 | /**********************************/ |
| 44 | /* Initialization of the Silk VAD */ |
| 45 | /**********************************/ |
| 46 | opus_int silk_VAD_Init( /* O Return value, 0 if success */ |
| 47 | silk_VAD_state *psSilk_VAD /* I/O Pointer to Silk VAD state */ |
| 48 | ) |
| 49 | { |
| 50 | opus_int b, ret = 0; |
| 51 | |
| 52 | /* reset state memory */ |
| 53 | silk_memset( psSilk_VAD, 0, sizeof( silk_VAD_state ) ); |
| 54 | |
| 55 | /* init noise levels */ |
| 56 | /* Initialize array with approx pink noise levels (psd proportional to inverse of frequency) */ |
| 57 | for( b = 0; b < VAD_N_BANDS; b++ ) { |
| 58 | psSilk_VAD->NoiseLevelBias[ b ] = silk_max_32( silk_DIV32_16( VAD_NOISE_LEVELS_BIAS, b + 1 ), 1 ); |
| 59 | } |
| 60 | |
| 61 | /* Initialize state */ |
| 62 | for( b = 0; b < VAD_N_BANDS; b++ ) { |
| 63 | psSilk_VAD->NL[ b ] = silk_MUL( 100, psSilk_VAD->NoiseLevelBias[ b ] ); |
| 64 | psSilk_VAD->inv_NL[ b ] = silk_DIV32( silk_int32_MAX, psSilk_VAD->NL[ b ] ); |
| 65 | } |
| 66 | psSilk_VAD->counter = 15; |
| 67 | |
| 68 | /* init smoothed energy-to-noise ratio*/ |
| 69 | for( b = 0; b < VAD_N_BANDS; b++ ) { |
| 70 | psSilk_VAD->NrgRatioSmth_Q8[ b ] = 100 * 256; /* 100 * 256 --> 20 dB SNR */ |
| 71 | } |
| 72 | |
| 73 | return( ret ); |
| 74 | } |
| 75 | |
| 76 | /* Weighting factors for tilt measure */ |
| 77 | static const opus_int32 tiltWeights[ VAD_N_BANDS ] = { 30000, 6000, -12000, -12000 }; |
| 78 | |
| 79 | /***************************************/ |
| 80 | /* Get the speech activity level in Q8 */ |
| 81 | /***************************************/ |
| 82 | opus_int silk_VAD_GetSA_Q8_c( /* O Return value, 0 if success */ |
| 83 | silk_encoder_state *psEncC, /* I/O Encoder state */ |
| 84 | const opus_int16 pIn[] /* I PCM input */ |
| 85 | ) |
| 86 | { |
| 87 | opus_int SA_Q15, pSNR_dB_Q7, input_tilt; |
| 88 | opus_int decimated_framelength1, decimated_framelength2; |
| 89 | opus_int decimated_framelength; |
| 90 | opus_int dec_subframe_length, dec_subframe_offset, SNR_Q7, i, b, s; |
| 91 | opus_int32 sumSquared, smooth_coef_Q16; |
| 92 | opus_int16 HPstateTmp; |
| 93 | VARDECL( opus_int16, X ); |
| 94 | opus_int32 Xnrg[ VAD_N_BANDS ]; |
| 95 | opus_int32 NrgToNoiseRatio_Q8[ VAD_N_BANDS ]; |
| 96 | opus_int32 speech_nrg, x_tmp; |
| 97 | opus_int X_offset[ VAD_N_BANDS ]; |
| 98 | opus_int ret = 0; |
| 99 | silk_VAD_state *psSilk_VAD = &psEncC->sVAD; |
| 100 | SAVE_STACK; |
| 101 | |
| 102 | /* Safety checks */ |
| 103 | silk_assert( VAD_N_BANDS == 4 ); |
| 104 | celt_assert( MAX_FRAME_LENGTH >= psEncC->frame_length ); |
| 105 | celt_assert( psEncC->frame_length <= 512 ); |
| 106 | celt_assert( psEncC->frame_length == 8 * silk_RSHIFT( psEncC->frame_length, 3 ) ); |
| 107 | |
| 108 | /***********************/ |
| 109 | /* Filter and Decimate */ |
| 110 | /***********************/ |
| 111 | decimated_framelength1 = silk_RSHIFT( psEncC->frame_length, 1 ); |
| 112 | decimated_framelength2 = silk_RSHIFT( psEncC->frame_length, 2 ); |
| 113 | decimated_framelength = silk_RSHIFT( psEncC->frame_length, 3 ); |
| 114 | /* Decimate into 4 bands: |
| 115 | 0 L 3L L 3L 5L |
| 116 | - -- - -- -- |
| 117 | 8 8 2 4 4 |
| 118 | |
| 119 | [0-1 kHz| temp. |1-2 kHz| 2-4 kHz | 4-8 kHz | |
| 120 | |
| 121 | They're arranged to allow the minimal ( frame_length / 4 ) extra |
| 122 | scratch space during the downsampling process */ |
| 123 | X_offset[ 0 ] = 0; |
| 124 | X_offset[ 1 ] = decimated_framelength + decimated_framelength2; |
| 125 | X_offset[ 2 ] = X_offset[ 1 ] + decimated_framelength; |
| 126 | X_offset[ 3 ] = X_offset[ 2 ] + decimated_framelength2; |
| 127 | ALLOC( X, X_offset[ 3 ] + decimated_framelength1, opus_int16 ); |
| 128 | |
| 129 | /* 0-8 kHz to 0-4 kHz and 4-8 kHz */ |
| 130 | silk_ana_filt_bank_1( pIn, &psSilk_VAD->AnaState[ 0 ], |
| 131 | X, &X[ X_offset[ 3 ] ], psEncC->frame_length ); |
| 132 | |
| 133 | /* 0-4 kHz to 0-2 kHz and 2-4 kHz */ |
| 134 | silk_ana_filt_bank_1( X, &psSilk_VAD->AnaState1[ 0 ], |
| 135 | X, &X[ X_offset[ 2 ] ], decimated_framelength1 ); |
| 136 | |
| 137 | /* 0-2 kHz to 0-1 kHz and 1-2 kHz */ |
| 138 | silk_ana_filt_bank_1( X, &psSilk_VAD->AnaState2[ 0 ], |
| 139 | X, &X[ X_offset[ 1 ] ], decimated_framelength2 ); |
| 140 | |
| 141 | /*********************************************/ |
| 142 | /* HP filter on lowest band (differentiator) */ |
| 143 | /*********************************************/ |
| 144 | X[ decimated_framelength - 1 ] = silk_RSHIFT( X[ decimated_framelength - 1 ], 1 ); |
| 145 | HPstateTmp = X[ decimated_framelength - 1 ]; |
| 146 | for( i = decimated_framelength - 1; i > 0; i-- ) { |
| 147 | X[ i - 1 ] = silk_RSHIFT( X[ i - 1 ], 1 ); |
| 148 | X[ i ] -= X[ i - 1 ]; |
| 149 | } |
| 150 | X[ 0 ] -= psSilk_VAD->HPstate; |
| 151 | psSilk_VAD->HPstate = HPstateTmp; |
| 152 | |
| 153 | /*************************************/ |
| 154 | /* Calculate the energy in each band */ |
| 155 | /*************************************/ |
| 156 | for( b = 0; b < VAD_N_BANDS; b++ ) { |
| 157 | /* Find the decimated framelength in the non-uniformly divided bands */ |
| 158 | decimated_framelength = silk_RSHIFT( psEncC->frame_length, silk_min_int( VAD_N_BANDS - b, VAD_N_BANDS - 1 ) ); |
| 159 | |
| 160 | /* Split length into subframe lengths */ |
| 161 | dec_subframe_length = silk_RSHIFT( decimated_framelength, VAD_INTERNAL_SUBFRAMES_LOG2 ); |
| 162 | dec_subframe_offset = 0; |
| 163 | |
| 164 | /* Compute energy per sub-frame */ |
| 165 | /* initialize with summed energy of last subframe */ |
| 166 | Xnrg[ b ] = psSilk_VAD->XnrgSubfr[ b ]; |
| 167 | for( s = 0; s < VAD_INTERNAL_SUBFRAMES; s++ ) { |
| 168 | sumSquared = 0; |
| 169 | for( i = 0; i < dec_subframe_length; i++ ) { |
| 170 | /* The energy will be less than dec_subframe_length * ( silk_int16_MIN / 8 ) ^ 2. */ |
| 171 | /* Therefore we can accumulate with no risk of overflow (unless dec_subframe_length > 128) */ |
| 172 | x_tmp = silk_RSHIFT( |
| 173 | X[ X_offset[ b ] + i + dec_subframe_offset ], 3 ); |
| 174 | sumSquared = silk_SMLABB( sumSquared, x_tmp, x_tmp ); |
| 175 | |
| 176 | /* Safety check */ |
| 177 | silk_assert( sumSquared >= 0 ); |
| 178 | } |
| 179 | |
| 180 | /* Add/saturate summed energy of current subframe */ |
| 181 | if( s < VAD_INTERNAL_SUBFRAMES - 1 ) { |
| 182 | Xnrg[ b ] = silk_ADD_POS_SAT32( Xnrg[ b ], sumSquared ); |
| 183 | } else { |
| 184 | /* Look-ahead subframe */ |
| 185 | Xnrg[ b ] = silk_ADD_POS_SAT32( Xnrg[ b ], silk_RSHIFT( sumSquared, 1 ) ); |
| 186 | } |
| 187 | |
| 188 | dec_subframe_offset += dec_subframe_length; |
| 189 | } |
| 190 | psSilk_VAD->XnrgSubfr[ b ] = sumSquared; |
| 191 | } |
| 192 | |
| 193 | /********************/ |
| 194 | /* Noise estimation */ |
| 195 | /********************/ |
| 196 | silk_VAD_GetNoiseLevels( &Xnrg[ 0 ], psSilk_VAD ); |
| 197 | |
| 198 | /***********************************************/ |
| 199 | /* Signal-plus-noise to noise ratio estimation */ |
| 200 | /***********************************************/ |
| 201 | sumSquared = 0; |
| 202 | input_tilt = 0; |
| 203 | for( b = 0; b < VAD_N_BANDS; b++ ) { |
| 204 | speech_nrg = Xnrg[ b ] - psSilk_VAD->NL[ b ]; |
| 205 | if( speech_nrg > 0 ) { |
| 206 | /* Divide, with sufficient resolution */ |
| 207 | if( ( Xnrg[ b ] & 0xFF800000 ) == 0 ) { |
| 208 | NrgToNoiseRatio_Q8[ b ] = silk_DIV32( silk_LSHIFT( Xnrg[ b ], 8 ), psSilk_VAD->NL[ b ] + 1 ); |
| 209 | } else { |
| 210 | NrgToNoiseRatio_Q8[ b ] = silk_DIV32( Xnrg[ b ], silk_RSHIFT( psSilk_VAD->NL[ b ], 8 ) + 1 ); |
| 211 | } |
| 212 | |
| 213 | /* Convert to log domain */ |
| 214 | SNR_Q7 = silk_lin2log( NrgToNoiseRatio_Q8[ b ] ) - 8 * 128; |
| 215 | |
| 216 | /* Sum-of-squares */ |
| 217 | sumSquared = silk_SMLABB( sumSquared, SNR_Q7, SNR_Q7 ); /* Q14 */ |
| 218 | |
| 219 | /* Tilt measure */ |
| 220 | if( speech_nrg < ( (opus_int32)1 << 20 ) ) { |
| 221 | /* Scale down SNR value for small subband speech energies */ |
| 222 | SNR_Q7 = silk_SMULWB( silk_LSHIFT( silk_SQRT_APPROX( speech_nrg ), 6 ), SNR_Q7 ); |
| 223 | } |
| 224 | input_tilt = silk_SMLAWB( input_tilt, tiltWeights[ b ], SNR_Q7 ); |
| 225 | } else { |
| 226 | NrgToNoiseRatio_Q8[ b ] = 256; |
| 227 | } |
| 228 | } |
| 229 | |
| 230 | /* Mean-of-squares */ |
| 231 | sumSquared = silk_DIV32_16( sumSquared, VAD_N_BANDS ); /* Q14 */ |
| 232 | |
| 233 | /* Root-mean-square approximation, scale to dBs, and write to output pointer */ |
| 234 | pSNR_dB_Q7 = (opus_int16)( 3 * silk_SQRT_APPROX( sumSquared ) ); /* Q7 */ |
| 235 | |
| 236 | /*********************************/ |
| 237 | /* Speech Probability Estimation */ |
| 238 | /*********************************/ |
| 239 | SA_Q15 = silk_sigm_Q15( silk_SMULWB( VAD_SNR_FACTOR_Q16, pSNR_dB_Q7 ) - VAD_NEGATIVE_OFFSET_Q5 ); |
| 240 | |
| 241 | /**************************/ |
| 242 | /* Frequency Tilt Measure */ |
| 243 | /**************************/ |
| 244 | psEncC->input_tilt_Q15 = silk_LSHIFT( silk_sigm_Q15( input_tilt ) - 16384, 1 ); |
| 245 | |
| 246 | /**************************************************/ |
| 247 | /* Scale the sigmoid output based on power levels */ |
| 248 | /**************************************************/ |
| 249 | speech_nrg = 0; |
| 250 | for( b = 0; b < VAD_N_BANDS; b++ ) { |
| 251 | /* Accumulate signal-without-noise energies, higher frequency bands have more weight */ |
| 252 | speech_nrg += ( b + 1 ) * silk_RSHIFT( Xnrg[ b ] - psSilk_VAD->NL[ b ], 4 ); |
| 253 | } |
| 254 | |
| 255 | if( psEncC->frame_length == 20 * psEncC->fs_kHz ) { |
| 256 | speech_nrg = silk_RSHIFT32( speech_nrg, 1 ); |
| 257 | } |
| 258 | /* Power scaling */ |
| 259 | if( speech_nrg <= 0 ) { |
| 260 | SA_Q15 = silk_RSHIFT( SA_Q15, 1 ); |
| 261 | } else if( speech_nrg < 16384 ) { |
| 262 | speech_nrg = silk_LSHIFT32( speech_nrg, 16 ); |
| 263 | |
| 264 | /* square-root */ |
| 265 | speech_nrg = silk_SQRT_APPROX( speech_nrg ); |
| 266 | SA_Q15 = silk_SMULWB( 32768 + speech_nrg, SA_Q15 ); |
| 267 | } |
| 268 | |
| 269 | /* Copy the resulting speech activity in Q8 */ |
| 270 | psEncC->speech_activity_Q8 = silk_min_int( silk_RSHIFT( SA_Q15, 7 ), silk_uint8_MAX ); |
| 271 | |
| 272 | /***********************************/ |
| 273 | /* Energy Level and SNR estimation */ |
| 274 | /***********************************/ |
| 275 | /* Smoothing coefficient */ |
| 276 | smooth_coef_Q16 = silk_SMULWB( VAD_SNR_SMOOTH_COEF_Q18, silk_SMULWB( (opus_int32)SA_Q15, SA_Q15 ) ); |
| 277 | |
| 278 | if( psEncC->frame_length == 10 * psEncC->fs_kHz ) { |
| 279 | smooth_coef_Q16 >>= 1; |
| 280 | } |
| 281 | |
| 282 | for( b = 0; b < VAD_N_BANDS; b++ ) { |
| 283 | /* compute smoothed energy-to-noise ratio per band */ |
| 284 | psSilk_VAD->NrgRatioSmth_Q8[ b ] = silk_SMLAWB( psSilk_VAD->NrgRatioSmth_Q8[ b ], |
| 285 | NrgToNoiseRatio_Q8[ b ] - psSilk_VAD->NrgRatioSmth_Q8[ b ], smooth_coef_Q16 ); |
| 286 | |
| 287 | /* signal to noise ratio in dB per band */ |
| 288 | SNR_Q7 = 3 * ( silk_lin2log( psSilk_VAD->NrgRatioSmth_Q8[b] ) - 8 * 128 ); |
| 289 | /* quality = sigmoid( 0.25 * ( SNR_dB - 16 ) ); */ |
| 290 | psEncC->input_quality_bands_Q15[ b ] = silk_sigm_Q15( silk_RSHIFT( SNR_Q7 - 16 * 128, 4 ) ); |
| 291 | } |
| 292 | |
| 293 | RESTORE_STACK; |
| 294 | return( ret ); |
| 295 | } |
| 296 | |
| 297 | /**************************/ |
| 298 | /* Noise level estimation */ |
| 299 | /**************************/ |
| 300 | # if !defined(OPUS_X86_MAY_HAVE_SSE4_1) |
| 301 | static OPUS_INLINE |
| 302 | #endif |
| 303 | void silk_VAD_GetNoiseLevels( |
| 304 | const opus_int32 pX[ VAD_N_BANDS ], /* I subband energies */ |
| 305 | silk_VAD_state *psSilk_VAD /* I/O Pointer to Silk VAD state */ |
| 306 | ) |
| 307 | { |
| 308 | opus_int k; |
| 309 | opus_int32 nl, nrg, inv_nrg; |
| 310 | opus_int coef, min_coef; |
| 311 | |
| 312 | /* Initially faster smoothing */ |
| 313 | if( psSilk_VAD->counter < 1000 ) { /* 1000 = 20 sec */ |
| 314 | min_coef = silk_DIV32_16( silk_int16_MAX, silk_RSHIFT( psSilk_VAD->counter, 4 ) + 1 ); |
| 315 | } else { |
| 316 | min_coef = 0; |
| 317 | } |
| 318 | |
| 319 | for( k = 0; k < VAD_N_BANDS; k++ ) { |
| 320 | /* Get old noise level estimate for current band */ |
| 321 | nl = psSilk_VAD->NL[ k ]; |
| 322 | silk_assert( nl >= 0 ); |
| 323 | |
| 324 | /* Add bias */ |
| 325 | nrg = silk_ADD_POS_SAT32( pX[ k ], psSilk_VAD->NoiseLevelBias[ k ] ); |
| 326 | silk_assert( nrg > 0 ); |
| 327 | |
| 328 | /* Invert energies */ |
| 329 | inv_nrg = silk_DIV32( silk_int32_MAX, nrg ); |
| 330 | silk_assert( inv_nrg >= 0 ); |
| 331 | |
| 332 | /* Less update when subband energy is high */ |
| 333 | if( nrg > silk_LSHIFT( nl, 3 ) ) { |
| 334 | coef = VAD_NOISE_LEVEL_SMOOTH_COEF_Q16 >> 3; |
| 335 | } else if( nrg < nl ) { |
| 336 | coef = VAD_NOISE_LEVEL_SMOOTH_COEF_Q16; |
| 337 | } else { |
| 338 | coef = silk_SMULWB( silk_SMULWW( inv_nrg, nl ), VAD_NOISE_LEVEL_SMOOTH_COEF_Q16 << 1 ); |
| 339 | } |
| 340 | |
| 341 | /* Initially faster smoothing */ |
| 342 | coef = silk_max_int( coef, min_coef ); |
| 343 | |
| 344 | /* Smooth inverse energies */ |
| 345 | psSilk_VAD->inv_NL[ k ] = silk_SMLAWB( psSilk_VAD->inv_NL[ k ], inv_nrg - psSilk_VAD->inv_NL[ k ], coef ); |
| 346 | silk_assert( psSilk_VAD->inv_NL[ k ] >= 0 ); |
| 347 | |
| 348 | /* Compute noise level by inverting again */ |
| 349 | nl = silk_DIV32( silk_int32_MAX, psSilk_VAD->inv_NL[ k ] ); |
| 350 | silk_assert( nl >= 0 ); |
| 351 | |
| 352 | /* Limit noise levels (guarantee 7 bits of head room) */ |
| 353 | nl = silk_min( nl, 0x00FFFFFF ); |
| 354 | |
| 355 | /* Store as part of state */ |
| 356 | psSilk_VAD->NL[ k ] = nl; |
| 357 | } |
| 358 | |
| 359 | /* Increment frame counter */ |
| 360 | psSilk_VAD->counter++; |
| 361 | } |