1 /* 2 * Copyright (c) 2011 The WebRTC project authors. All Rights Reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11 /* digital_agc.c 12 * 13 */ 14 15 #include "webrtc/modules/audio_processing/agc/digital_agc.h" 16 17 #include <assert.h> 18 #include <string.h> 19 #ifdef AGC_DEBUG 20 #include <stdio.h> 21 #endif 22 23 #include "webrtc/modules/audio_processing/agc/include/gain_control.h" 24 25 // To generate the gaintable, copy&paste the following lines to a Matlab window: 26 // MaxGain = 6; MinGain = 0; CompRatio = 3; Knee = 1; 27 // zeros = 0:31; lvl = 2.^(1-zeros); 28 // A = -10*log10(lvl) * (CompRatio - 1) / CompRatio; 29 // B = MaxGain - MinGain; 30 // gains = round(2^16*10.^(0.05 * (MinGain + B * ( log(exp(-Knee*A)+exp(-Knee*B)) - log(1+exp(-Knee*B)) ) / log(1/(1+exp(Knee*B)))))); 31 // fprintf(1, '\t%i, %i, %i, %i,\n', gains); 32 // % Matlab code for plotting the gain and input/output level characteristic (copy/paste the following 3 lines): 33 // in = 10*log10(lvl); out = 20*log10(gains/65536); 34 // subplot(121); plot(in, out); axis([-30, 0, -5, 20]); grid on; xlabel('Input (dB)'); ylabel('Gain (dB)'); 35 // subplot(122); plot(in, in+out); axis([-30, 0, -30, 5]); grid on; xlabel('Input (dB)'); ylabel('Output (dB)'); 36 // zoom on; 37 38 // Generator table for y=log2(1+e^x) in Q8. 39 enum { kGenFuncTableSize = 128 }; 40 static const uint16_t kGenFuncTable[kGenFuncTableSize] = { 41 256, 485, 786, 1126, 1484, 1849, 2217, 2586, 42 2955, 3324, 3693, 4063, 4432, 4801, 5171, 5540, 43 5909, 6279, 6648, 7017, 7387, 7756, 8125, 8495, 44 8864, 9233, 9603, 9972, 10341, 10711, 11080, 11449, 45 11819, 12188, 12557, 12927, 13296, 13665, 14035, 14404, 46 14773, 15143, 15512, 15881, 16251, 16620, 16989, 17359, 47 17728, 18097, 18466, 18836, 19205, 19574, 19944, 20313, 48 20682, 21052, 21421, 21790, 22160, 22529, 22898, 23268, 49 23637, 24006, 24376, 24745, 25114, 25484, 25853, 26222, 50 26592, 26961, 27330, 27700, 28069, 28438, 28808, 29177, 51 29546, 29916, 30285, 30654, 31024, 31393, 31762, 32132, 52 32501, 32870, 33240, 33609, 33978, 34348, 34717, 35086, 53 35456, 35825, 36194, 36564, 36933, 37302, 37672, 38041, 54 38410, 38780, 39149, 39518, 39888, 40257, 40626, 40996, 55 41365, 41734, 42104, 42473, 42842, 43212, 43581, 43950, 56 44320, 44689, 45058, 45428, 45797, 46166, 46536, 46905 57 }; 58 59 static const int16_t kAvgDecayTime = 250; // frames; < 3000 60 61 int32_t WebRtcAgc_CalculateGainTable(int32_t *gainTable, // Q16 62 int16_t digCompGaindB, // Q0 63 int16_t targetLevelDbfs,// Q0 64 uint8_t limiterEnable, 65 int16_t analogTarget) // Q0 66 { 67 // This function generates the compressor gain table used in the fixed digital part. 68 uint32_t tmpU32no1, tmpU32no2, absInLevel, logApprox; 69 int32_t inLevel, limiterLvl; 70 int32_t tmp32, tmp32no1, tmp32no2, numFIX, den, y32; 71 const uint16_t kLog10 = 54426; // log2(10) in Q14 72 const uint16_t kLog10_2 = 49321; // 10*log10(2) in Q14 73 const uint16_t kLogE_1 = 23637; // log2(e) in Q14 74 uint16_t constMaxGain; 75 uint16_t tmpU16, intPart, fracPart; 76 const int16_t kCompRatio = 3; 77 const int16_t kSoftLimiterLeft = 1; 78 int16_t limiterOffset = 0; // Limiter offset 79 int16_t limiterIdx, limiterLvlX; 80 int16_t constLinApprox, zeroGainLvl, maxGain, diffGain; 81 int16_t i, tmp16, tmp16no1; 82 int zeros, zerosScale; 83 84 // Constants 85 // kLogE_1 = 23637; // log2(e) in Q14 86 // kLog10 = 54426; // log2(10) in Q14 87 // kLog10_2 = 49321; // 10*log10(2) in Q14 88 89 // Calculate maximum digital gain and zero gain level 90 tmp32no1 = WEBRTC_SPL_MUL_16_16(digCompGaindB - analogTarget, kCompRatio - 1); 91 tmp16no1 = analogTarget - targetLevelDbfs; 92 tmp16no1 += WebRtcSpl_DivW32W16ResW16(tmp32no1 + (kCompRatio >> 1), kCompRatio); 93 maxGain = WEBRTC_SPL_MAX(tmp16no1, (analogTarget - targetLevelDbfs)); 94 tmp32no1 = WEBRTC_SPL_MUL_16_16(maxGain, kCompRatio); 95 zeroGainLvl = digCompGaindB; 96 zeroGainLvl -= WebRtcSpl_DivW32W16ResW16(tmp32no1 + ((kCompRatio - 1) >> 1), 97 kCompRatio - 1); 98 if ((digCompGaindB <= analogTarget) && (limiterEnable)) 99 { 100 zeroGainLvl += (analogTarget - digCompGaindB + kSoftLimiterLeft); 101 limiterOffset = 0; 102 } 103 104 // Calculate the difference between maximum gain and gain at 0dB0v: 105 // diffGain = maxGain + (compRatio-1)*zeroGainLvl/compRatio 106 // = (compRatio-1)*digCompGaindB/compRatio 107 tmp32no1 = WEBRTC_SPL_MUL_16_16(digCompGaindB, kCompRatio - 1); 108 diffGain = WebRtcSpl_DivW32W16ResW16(tmp32no1 + (kCompRatio >> 1), kCompRatio); 109 if (diffGain < 0 || diffGain >= kGenFuncTableSize) 110 { 111 assert(0); 112 return -1; 113 } 114 115 // Calculate the limiter level and index: 116 // limiterLvlX = analogTarget - limiterOffset 117 // limiterLvl = targetLevelDbfs + limiterOffset/compRatio 118 limiterLvlX = analogTarget - limiterOffset; 119 limiterIdx = 2 120 + WebRtcSpl_DivW32W16ResW16(WEBRTC_SPL_LSHIFT_W32((int32_t)limiterLvlX, 13), 121 (kLog10_2 / 2)); 122 tmp16no1 = WebRtcSpl_DivW32W16ResW16(limiterOffset + (kCompRatio >> 1), kCompRatio); 123 limiterLvl = targetLevelDbfs + tmp16no1; 124 125 // Calculate (through table lookup): 126 // constMaxGain = log2(1+2^(log2(e)*diffGain)); (in Q8) 127 constMaxGain = kGenFuncTable[diffGain]; // in Q8 128 129 // Calculate a parameter used to approximate the fractional part of 2^x with a 130 // piecewise linear function in Q14: 131 // constLinApprox = round(3/2*(4*(3-2*sqrt(2))/(log(2)^2)-0.5)*2^14); 132 constLinApprox = 22817; // in Q14 133 134 // Calculate a denominator used in the exponential part to convert from dB to linear scale: 135 // den = 20*constMaxGain (in Q8) 136 den = WEBRTC_SPL_MUL_16_U16(20, constMaxGain); // in Q8 137 138 for (i = 0; i < 32; i++) 139 { 140 // Calculate scaled input level (compressor): 141 // inLevel = fix((-constLog10_2*(compRatio-1)*(1-i)+fix(compRatio/2))/compRatio) 142 tmp16 = (int16_t)WEBRTC_SPL_MUL_16_16(kCompRatio - 1, i - 1); // Q0 143 tmp32 = WEBRTC_SPL_MUL_16_U16(tmp16, kLog10_2) + 1; // Q14 144 inLevel = WebRtcSpl_DivW32W16(tmp32, kCompRatio); // Q14 145 146 // Calculate diffGain-inLevel, to map using the genFuncTable 147 inLevel = WEBRTC_SPL_LSHIFT_W32((int32_t)diffGain, 14) - inLevel; // Q14 148 149 // Make calculations on abs(inLevel) and compensate for the sign afterwards. 150 absInLevel = (uint32_t)WEBRTC_SPL_ABS_W32(inLevel); // Q14 151 152 // LUT with interpolation 153 intPart = (uint16_t)WEBRTC_SPL_RSHIFT_U32(absInLevel, 14); 154 fracPart = (uint16_t)(absInLevel & 0x00003FFF); // extract the fractional part 155 tmpU16 = kGenFuncTable[intPart + 1] - kGenFuncTable[intPart]; // Q8 156 tmpU32no1 = WEBRTC_SPL_UMUL_16_16(tmpU16, fracPart); // Q22 157 tmpU32no1 += WEBRTC_SPL_LSHIFT_U32((uint32_t)kGenFuncTable[intPart], 14); // Q22 158 logApprox = WEBRTC_SPL_RSHIFT_U32(tmpU32no1, 8); // Q14 159 // Compensate for negative exponent using the relation: 160 // log2(1 + 2^-x) = log2(1 + 2^x) - x 161 if (inLevel < 0) 162 { 163 zeros = WebRtcSpl_NormU32(absInLevel); 164 zerosScale = 0; 165 if (zeros < 15) 166 { 167 // Not enough space for multiplication 168 tmpU32no2 = WEBRTC_SPL_RSHIFT_U32(absInLevel, 15 - zeros); // Q(zeros-1) 169 tmpU32no2 = WEBRTC_SPL_UMUL_32_16(tmpU32no2, kLogE_1); // Q(zeros+13) 170 if (zeros < 9) 171 { 172 tmpU32no1 = WEBRTC_SPL_RSHIFT_U32(tmpU32no1, 9 - zeros); // Q(zeros+13) 173 zerosScale = 9 - zeros; 174 } else 175 { 176 tmpU32no2 = WEBRTC_SPL_RSHIFT_U32(tmpU32no2, zeros - 9); // Q22 177 } 178 } else 179 { 180 tmpU32no2 = WEBRTC_SPL_UMUL_32_16(absInLevel, kLogE_1); // Q28 181 tmpU32no2 = WEBRTC_SPL_RSHIFT_U32(tmpU32no2, 6); // Q22 182 } 183 logApprox = 0; 184 if (tmpU32no2 < tmpU32no1) 185 { 186 logApprox = WEBRTC_SPL_RSHIFT_U32(tmpU32no1 - tmpU32no2, 8 - zerosScale); //Q14 187 } 188 } 189 numFIX = WEBRTC_SPL_LSHIFT_W32(WEBRTC_SPL_MUL_16_U16(maxGain, constMaxGain), 6); // Q14 190 numFIX -= WEBRTC_SPL_MUL_32_16((int32_t)logApprox, diffGain); // Q14 191 192 // Calculate ratio 193 // Shift |numFIX| as much as possible. 194 // Ensure we avoid wrap-around in |den| as well. 195 if (numFIX > (den >> 8)) // |den| is Q8. 196 { 197 zeros = WebRtcSpl_NormW32(numFIX); 198 } else 199 { 200 zeros = WebRtcSpl_NormW32(den) + 8; 201 } 202 numFIX = WEBRTC_SPL_LSHIFT_W32(numFIX, zeros); // Q(14+zeros) 203 204 // Shift den so we end up in Qy1 205 tmp32no1 = WEBRTC_SPL_SHIFT_W32(den, zeros - 8); // Q(zeros) 206 if (numFIX < 0) 207 { 208 numFIX -= WEBRTC_SPL_RSHIFT_W32(tmp32no1, 1); 209 } else 210 { 211 numFIX += WEBRTC_SPL_RSHIFT_W32(tmp32no1, 1); 212 } 213 y32 = WEBRTC_SPL_DIV(numFIX, tmp32no1); // in Q14 214 if (limiterEnable && (i < limiterIdx)) 215 { 216 tmp32 = WEBRTC_SPL_MUL_16_U16(i - 1, kLog10_2); // Q14 217 tmp32 -= WEBRTC_SPL_LSHIFT_W32(limiterLvl, 14); // Q14 218 y32 = WebRtcSpl_DivW32W16(tmp32 + 10, 20); 219 } 220 if (y32 > 39000) 221 { 222 tmp32 = WEBRTC_SPL_MUL(y32 >> 1, kLog10) + 4096; // in Q27 223 tmp32 = WEBRTC_SPL_RSHIFT_W32(tmp32, 13); // in Q14 224 } else 225 { 226 tmp32 = WEBRTC_SPL_MUL(y32, kLog10) + 8192; // in Q28 227 tmp32 = WEBRTC_SPL_RSHIFT_W32(tmp32, 14); // in Q14 228 } 229 tmp32 += WEBRTC_SPL_LSHIFT_W32(16, 14); // in Q14 (Make sure final output is in Q16) 230 231 // Calculate power 232 if (tmp32 > 0) 233 { 234 intPart = (int16_t)WEBRTC_SPL_RSHIFT_W32(tmp32, 14); 235 fracPart = (uint16_t)(tmp32 & 0x00003FFF); // in Q14 236 if (WEBRTC_SPL_RSHIFT_W32(fracPart, 13)) 237 { 238 tmp16 = WEBRTC_SPL_LSHIFT_W16(2, 14) - constLinApprox; 239 tmp32no2 = WEBRTC_SPL_LSHIFT_W32(1, 14) - fracPart; 240 tmp32no2 = WEBRTC_SPL_MUL_32_16(tmp32no2, tmp16); 241 tmp32no2 = WEBRTC_SPL_RSHIFT_W32(tmp32no2, 13); 242 tmp32no2 = WEBRTC_SPL_LSHIFT_W32(1, 14) - tmp32no2; 243 } else 244 { 245 tmp16 = constLinApprox - WEBRTC_SPL_LSHIFT_W16(1, 14); 246 tmp32no2 = WEBRTC_SPL_MUL_32_16(fracPart, tmp16); 247 tmp32no2 = WEBRTC_SPL_RSHIFT_W32(tmp32no2, 13); 248 } 249 fracPart = (uint16_t)tmp32no2; 250 gainTable[i] = WEBRTC_SPL_LSHIFT_W32(1, intPart) 251 + WEBRTC_SPL_SHIFT_W32(fracPart, intPart - 14); 252 } else 253 { 254 gainTable[i] = 0; 255 } 256 } 257 258 return 0; 259 } 260 261 int32_t WebRtcAgc_InitDigital(DigitalAgc_t *stt, int16_t agcMode) 262 { 263 264 if (agcMode == kAgcModeFixedDigital) 265 { 266 // start at minimum to find correct gain faster 267 stt->capacitorSlow = 0; 268 } else 269 { 270 // start out with 0 dB gain 271 stt->capacitorSlow = 134217728; // (int32_t)(0.125f * 32768.0f * 32768.0f); 272 } 273 stt->capacitorFast = 0; 274 stt->gain = 65536; 275 stt->gatePrevious = 0; 276 stt->agcMode = agcMode; 277 #ifdef AGC_DEBUG 278 stt->frameCounter = 0; 279 #endif 280 281 // initialize VADs 282 WebRtcAgc_InitVad(&stt->vadNearend); 283 WebRtcAgc_InitVad(&stt->vadFarend); 284 285 return 0; 286 } 287 288 int32_t WebRtcAgc_AddFarendToDigital(DigitalAgc_t *stt, const int16_t *in_far, 289 int16_t nrSamples) 290 { 291 assert(stt != NULL); 292 // VAD for far end 293 WebRtcAgc_ProcessVad(&stt->vadFarend, in_far, nrSamples); 294 295 return 0; 296 } 297 298 int32_t WebRtcAgc_ProcessDigital(DigitalAgc_t *stt, const int16_t *in_near, 299 const int16_t *in_near_H, int16_t *out, 300 int16_t *out_H, uint32_t FS, 301 int16_t lowlevelSignal) 302 { 303 // array for gains (one value per ms, incl start & end) 304 int32_t gains[11]; 305 306 int32_t out_tmp, tmp32; 307 int32_t env[10]; 308 int32_t nrg, max_nrg; 309 int32_t cur_level; 310 int32_t gain32, delta; 311 int16_t logratio; 312 int16_t lower_thr, upper_thr; 313 int16_t zeros, zeros_fast, frac; 314 int16_t decay; 315 int16_t gate, gain_adj; 316 int16_t k, n; 317 int16_t L, L2; // samples/subframe 318 319 // determine number of samples per ms 320 if (FS == 8000) 321 { 322 L = 8; 323 L2 = 3; 324 } else if (FS == 16000) 325 { 326 L = 16; 327 L2 = 4; 328 } else if (FS == 32000) 329 { 330 L = 16; 331 L2 = 4; 332 } else 333 { 334 return -1; 335 } 336 337 // TODO(andrew): again, we don't need input and output pointers... 338 if (in_near != out) 339 { 340 // Only needed if they don't already point to the same place. 341 memcpy(out, in_near, 10 * L * sizeof(int16_t)); 342 } 343 if (FS == 32000) 344 { 345 if (in_near_H != out_H) 346 { 347 memcpy(out_H, in_near_H, 10 * L * sizeof(int16_t)); 348 } 349 } 350 // VAD for near end 351 logratio = WebRtcAgc_ProcessVad(&stt->vadNearend, out, L * 10); 352 353 // Account for far end VAD 354 if (stt->vadFarend.counter > 10) 355 { 356 tmp32 = WEBRTC_SPL_MUL_16_16(3, logratio); 357 logratio = (int16_t)WEBRTC_SPL_RSHIFT_W32(tmp32 - stt->vadFarend.logRatio, 2); 358 } 359 360 // Determine decay factor depending on VAD 361 // upper_thr = 1.0f; 362 // lower_thr = 0.25f; 363 upper_thr = 1024; // Q10 364 lower_thr = 0; // Q10 365 if (logratio > upper_thr) 366 { 367 // decay = -2^17 / DecayTime; -> -65 368 decay = -65; 369 } else if (logratio < lower_thr) 370 { 371 decay = 0; 372 } else 373 { 374 // decay = (int16_t)(((lower_thr - logratio) 375 // * (2^27/(DecayTime*(upper_thr-lower_thr)))) >> 10); 376 // SUBSTITUTED: 2^27/(DecayTime*(upper_thr-lower_thr)) -> 65 377 tmp32 = WEBRTC_SPL_MUL_16_16((lower_thr - logratio), 65); 378 decay = (int16_t)WEBRTC_SPL_RSHIFT_W32(tmp32, 10); 379 } 380 381 // adjust decay factor for long silence (detected as low standard deviation) 382 // This is only done in the adaptive modes 383 if (stt->agcMode != kAgcModeFixedDigital) 384 { 385 if (stt->vadNearend.stdLongTerm < 4000) 386 { 387 decay = 0; 388 } else if (stt->vadNearend.stdLongTerm < 8096) 389 { 390 // decay = (int16_t)(((stt->vadNearend.stdLongTerm - 4000) * decay) >> 12); 391 tmp32 = WEBRTC_SPL_MUL_16_16((stt->vadNearend.stdLongTerm - 4000), decay); 392 decay = (int16_t)WEBRTC_SPL_RSHIFT_W32(tmp32, 12); 393 } 394 395 if (lowlevelSignal != 0) 396 { 397 decay = 0; 398 } 399 } 400 #ifdef AGC_DEBUG 401 stt->frameCounter++; 402 fprintf(stt->logFile, "%5.2f\t%d\t%d\t%d\t", (float)(stt->frameCounter) / 100, logratio, decay, stt->vadNearend.stdLongTerm); 403 #endif 404 // Find max amplitude per sub frame 405 // iterate over sub frames 406 for (k = 0; k < 10; k++) 407 { 408 // iterate over samples 409 max_nrg = 0; 410 for (n = 0; n < L; n++) 411 { 412 nrg = WEBRTC_SPL_MUL_16_16(out[k * L + n], out[k * L + n]); 413 if (nrg > max_nrg) 414 { 415 max_nrg = nrg; 416 } 417 } 418 env[k] = max_nrg; 419 } 420 421 // Calculate gain per sub frame 422 gains[0] = stt->gain; 423 for (k = 0; k < 10; k++) 424 { 425 // Fast envelope follower 426 // decay time = -131000 / -1000 = 131 (ms) 427 stt->capacitorFast = AGC_SCALEDIFF32(-1000, stt->capacitorFast, stt->capacitorFast); 428 if (env[k] > stt->capacitorFast) 429 { 430 stt->capacitorFast = env[k]; 431 } 432 // Slow envelope follower 433 if (env[k] > stt->capacitorSlow) 434 { 435 // increase capacitorSlow 436 stt->capacitorSlow 437 = AGC_SCALEDIFF32(500, (env[k] - stt->capacitorSlow), stt->capacitorSlow); 438 } else 439 { 440 // decrease capacitorSlow 441 stt->capacitorSlow 442 = AGC_SCALEDIFF32(decay, stt->capacitorSlow, stt->capacitorSlow); 443 } 444 445 // use maximum of both capacitors as current level 446 if (stt->capacitorFast > stt->capacitorSlow) 447 { 448 cur_level = stt->capacitorFast; 449 } else 450 { 451 cur_level = stt->capacitorSlow; 452 } 453 // Translate signal level into gain, using a piecewise linear approximation 454 // find number of leading zeros 455 zeros = WebRtcSpl_NormU32((uint32_t)cur_level); 456 if (cur_level == 0) 457 { 458 zeros = 31; 459 } 460 tmp32 = (WEBRTC_SPL_LSHIFT_W32(cur_level, zeros) & 0x7FFFFFFF); 461 frac = (int16_t)WEBRTC_SPL_RSHIFT_W32(tmp32, 19); // Q12 462 tmp32 = WEBRTC_SPL_MUL((stt->gainTable[zeros-1] - stt->gainTable[zeros]), frac); 463 gains[k + 1] = stt->gainTable[zeros] + WEBRTC_SPL_RSHIFT_W32(tmp32, 12); 464 #ifdef AGC_DEBUG 465 if (k == 0) 466 { 467 fprintf(stt->logFile, "%d\t%d\t%d\t%d\t%d\n", env[0], cur_level, stt->capacitorFast, stt->capacitorSlow, zeros); 468 } 469 #endif 470 } 471 472 // Gate processing (lower gain during absence of speech) 473 zeros = WEBRTC_SPL_LSHIFT_W16(zeros, 9) - WEBRTC_SPL_RSHIFT_W16(frac, 3); 474 // find number of leading zeros 475 zeros_fast = WebRtcSpl_NormU32((uint32_t)stt->capacitorFast); 476 if (stt->capacitorFast == 0) 477 { 478 zeros_fast = 31; 479 } 480 tmp32 = (WEBRTC_SPL_LSHIFT_W32(stt->capacitorFast, zeros_fast) & 0x7FFFFFFF); 481 zeros_fast = WEBRTC_SPL_LSHIFT_W16(zeros_fast, 9); 482 zeros_fast -= (int16_t)WEBRTC_SPL_RSHIFT_W32(tmp32, 22); 483 484 gate = 1000 + zeros_fast - zeros - stt->vadNearend.stdShortTerm; 485 486 if (gate < 0) 487 { 488 stt->gatePrevious = 0; 489 } else 490 { 491 tmp32 = WEBRTC_SPL_MUL_16_16(stt->gatePrevious, 7); 492 gate = (int16_t)WEBRTC_SPL_RSHIFT_W32((int32_t)gate + tmp32, 3); 493 stt->gatePrevious = gate; 494 } 495 // gate < 0 -> no gate 496 // gate > 2500 -> max gate 497 if (gate > 0) 498 { 499 if (gate < 2500) 500 { 501 gain_adj = WEBRTC_SPL_RSHIFT_W16(2500 - gate, 5); 502 } else 503 { 504 gain_adj = 0; 505 } 506 for (k = 0; k < 10; k++) 507 { 508 if ((gains[k + 1] - stt->gainTable[0]) > 8388608) 509 { 510 // To prevent wraparound 511 tmp32 = WEBRTC_SPL_RSHIFT_W32((gains[k+1] - stt->gainTable[0]), 8); 512 tmp32 = WEBRTC_SPL_MUL(tmp32, (178 + gain_adj)); 513 } else 514 { 515 tmp32 = WEBRTC_SPL_MUL((gains[k+1] - stt->gainTable[0]), (178 + gain_adj)); 516 tmp32 = WEBRTC_SPL_RSHIFT_W32(tmp32, 8); 517 } 518 gains[k + 1] = stt->gainTable[0] + tmp32; 519 } 520 } 521 522 // Limit gain to avoid overload distortion 523 for (k = 0; k < 10; k++) 524 { 525 // To prevent wrap around 526 zeros = 10; 527 if (gains[k + 1] > 47453132) 528 { 529 zeros = 16 - WebRtcSpl_NormW32(gains[k + 1]); 530 } 531 gain32 = WEBRTC_SPL_RSHIFT_W32(gains[k+1], zeros) + 1; 532 gain32 = WEBRTC_SPL_MUL(gain32, gain32); 533 // check for overflow 534 while (AGC_MUL32(WEBRTC_SPL_RSHIFT_W32(env[k], 12) + 1, gain32) 535 > WEBRTC_SPL_SHIFT_W32((int32_t)32767, 2 * (1 - zeros + 10))) 536 { 537 // multiply by 253/256 ==> -0.1 dB 538 if (gains[k + 1] > 8388607) 539 { 540 // Prevent wrap around 541 gains[k + 1] = WEBRTC_SPL_MUL(WEBRTC_SPL_RSHIFT_W32(gains[k+1], 8), 253); 542 } else 543 { 544 gains[k + 1] = WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL(gains[k+1], 253), 8); 545 } 546 gain32 = WEBRTC_SPL_RSHIFT_W32(gains[k+1], zeros) + 1; 547 gain32 = WEBRTC_SPL_MUL(gain32, gain32); 548 } 549 } 550 // gain reductions should be done 1 ms earlier than gain increases 551 for (k = 1; k < 10; k++) 552 { 553 if (gains[k] > gains[k + 1]) 554 { 555 gains[k] = gains[k + 1]; 556 } 557 } 558 // save start gain for next frame 559 stt->gain = gains[10]; 560 561 // Apply gain 562 // handle first sub frame separately 563 delta = WEBRTC_SPL_LSHIFT_W32(gains[1] - gains[0], (4 - L2)); 564 gain32 = WEBRTC_SPL_LSHIFT_W32(gains[0], 4); 565 // iterate over samples 566 for (n = 0; n < L; n++) 567 { 568 // For lower band 569 tmp32 = WEBRTC_SPL_MUL((int32_t)out[n], WEBRTC_SPL_RSHIFT_W32(gain32 + 127, 7)); 570 out_tmp = WEBRTC_SPL_RSHIFT_W32(tmp32 , 16); 571 if (out_tmp > 4095) 572 { 573 out[n] = (int16_t)32767; 574 } else if (out_tmp < -4096) 575 { 576 out[n] = (int16_t)-32768; 577 } else 578 { 579 tmp32 = WEBRTC_SPL_MUL((int32_t)out[n], WEBRTC_SPL_RSHIFT_W32(gain32, 4)); 580 out[n] = (int16_t)WEBRTC_SPL_RSHIFT_W32(tmp32 , 16); 581 } 582 // For higher band 583 if (FS == 32000) 584 { 585 tmp32 = WEBRTC_SPL_MUL((int32_t)out_H[n], 586 WEBRTC_SPL_RSHIFT_W32(gain32 + 127, 7)); 587 out_tmp = WEBRTC_SPL_RSHIFT_W32(tmp32 , 16); 588 if (out_tmp > 4095) 589 { 590 out_H[n] = (int16_t)32767; 591 } else if (out_tmp < -4096) 592 { 593 out_H[n] = (int16_t)-32768; 594 } else 595 { 596 tmp32 = WEBRTC_SPL_MUL((int32_t)out_H[n], 597 WEBRTC_SPL_RSHIFT_W32(gain32, 4)); 598 out_H[n] = (int16_t)WEBRTC_SPL_RSHIFT_W32(tmp32 , 16); 599 } 600 } 601 // 602 603 gain32 += delta; 604 } 605 // iterate over subframes 606 for (k = 1; k < 10; k++) 607 { 608 delta = WEBRTC_SPL_LSHIFT_W32(gains[k+1] - gains[k], (4 - L2)); 609 gain32 = WEBRTC_SPL_LSHIFT_W32(gains[k], 4); 610 // iterate over samples 611 for (n = 0; n < L; n++) 612 { 613 // For lower band 614 tmp32 = WEBRTC_SPL_MUL((int32_t)out[k * L + n], 615 WEBRTC_SPL_RSHIFT_W32(gain32, 4)); 616 out[k * L + n] = (int16_t)WEBRTC_SPL_RSHIFT_W32(tmp32 , 16); 617 // For higher band 618 if (FS == 32000) 619 { 620 tmp32 = WEBRTC_SPL_MUL((int32_t)out_H[k * L + n], 621 WEBRTC_SPL_RSHIFT_W32(gain32, 4)); 622 out_H[k * L + n] = (int16_t)WEBRTC_SPL_RSHIFT_W32(tmp32 , 16); 623 } 624 gain32 += delta; 625 } 626 } 627 628 return 0; 629 } 630 631 void WebRtcAgc_InitVad(AgcVad_t *state) 632 { 633 int16_t k; 634 635 state->HPstate = 0; // state of high pass filter 636 state->logRatio = 0; // log( P(active) / P(inactive) ) 637 // average input level (Q10) 638 state->meanLongTerm = WEBRTC_SPL_LSHIFT_W16(15, 10); 639 640 // variance of input level (Q8) 641 state->varianceLongTerm = WEBRTC_SPL_LSHIFT_W32(500, 8); 642 643 state->stdLongTerm = 0; // standard deviation of input level in dB 644 // short-term average input level (Q10) 645 state->meanShortTerm = WEBRTC_SPL_LSHIFT_W16(15, 10); 646 647 // short-term variance of input level (Q8) 648 state->varianceShortTerm = WEBRTC_SPL_LSHIFT_W32(500, 8); 649 650 state->stdShortTerm = 0; // short-term standard deviation of input level in dB 651 state->counter = 3; // counts updates 652 for (k = 0; k < 8; k++) 653 { 654 // downsampling filter 655 state->downState[k] = 0; 656 } 657 } 658 659 int16_t WebRtcAgc_ProcessVad(AgcVad_t *state, // (i) VAD state 660 const int16_t *in, // (i) Speech signal 661 int16_t nrSamples) // (i) number of samples 662 { 663 int32_t out, nrg, tmp32, tmp32b; 664 uint16_t tmpU16; 665 int16_t k, subfr, tmp16; 666 int16_t buf1[8]; 667 int16_t buf2[4]; 668 int16_t HPstate; 669 int16_t zeros, dB; 670 671 // process in 10 sub frames of 1 ms (to save on memory) 672 nrg = 0; 673 HPstate = state->HPstate; 674 for (subfr = 0; subfr < 10; subfr++) 675 { 676 // downsample to 4 kHz 677 if (nrSamples == 160) 678 { 679 for (k = 0; k < 8; k++) 680 { 681 tmp32 = (int32_t)in[2 * k] + (int32_t)in[2 * k + 1]; 682 tmp32 = WEBRTC_SPL_RSHIFT_W32(tmp32, 1); 683 buf1[k] = (int16_t)tmp32; 684 } 685 in += 16; 686 687 WebRtcSpl_DownsampleBy2(buf1, 8, buf2, state->downState); 688 } else 689 { 690 WebRtcSpl_DownsampleBy2(in, 8, buf2, state->downState); 691 in += 8; 692 } 693 694 // high pass filter and compute energy 695 for (k = 0; k < 4; k++) 696 { 697 out = buf2[k] + HPstate; 698 tmp32 = WEBRTC_SPL_MUL(600, out); 699 HPstate = (int16_t)(WEBRTC_SPL_RSHIFT_W32(tmp32, 10) - buf2[k]); 700 tmp32 = WEBRTC_SPL_MUL(out, out); 701 nrg += WEBRTC_SPL_RSHIFT_W32(tmp32, 6); 702 } 703 } 704 state->HPstate = HPstate; 705 706 // find number of leading zeros 707 if (!(0xFFFF0000 & nrg)) 708 { 709 zeros = 16; 710 } else 711 { 712 zeros = 0; 713 } 714 if (!(0xFF000000 & (nrg << zeros))) 715 { 716 zeros += 8; 717 } 718 if (!(0xF0000000 & (nrg << zeros))) 719 { 720 zeros += 4; 721 } 722 if (!(0xC0000000 & (nrg << zeros))) 723 { 724 zeros += 2; 725 } 726 if (!(0x80000000 & (nrg << zeros))) 727 { 728 zeros += 1; 729 } 730 731 // energy level (range {-32..30}) (Q10) 732 dB = WEBRTC_SPL_LSHIFT_W16(15 - zeros, 11); 733 734 // Update statistics 735 736 if (state->counter < kAvgDecayTime) 737 { 738 // decay time = AvgDecTime * 10 ms 739 state->counter++; 740 } 741 742 // update short-term estimate of mean energy level (Q10) 743 tmp32 = (WEBRTC_SPL_MUL_16_16(state->meanShortTerm, 15) + (int32_t)dB); 744 state->meanShortTerm = (int16_t)WEBRTC_SPL_RSHIFT_W32(tmp32, 4); 745 746 // update short-term estimate of variance in energy level (Q8) 747 tmp32 = WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_16_16(dB, dB), 12); 748 tmp32 += WEBRTC_SPL_MUL(state->varianceShortTerm, 15); 749 state->varianceShortTerm = WEBRTC_SPL_RSHIFT_W32(tmp32, 4); 750 751 // update short-term estimate of standard deviation in energy level (Q10) 752 tmp32 = WEBRTC_SPL_MUL_16_16(state->meanShortTerm, state->meanShortTerm); 753 tmp32 = WEBRTC_SPL_LSHIFT_W32(state->varianceShortTerm, 12) - tmp32; 754 state->stdShortTerm = (int16_t)WebRtcSpl_Sqrt(tmp32); 755 756 // update long-term estimate of mean energy level (Q10) 757 tmp32 = WEBRTC_SPL_MUL_16_16(state->meanLongTerm, state->counter) + (int32_t)dB; 758 state->meanLongTerm = WebRtcSpl_DivW32W16ResW16(tmp32, 759 WEBRTC_SPL_ADD_SAT_W16(state->counter, 1)); 760 761 // update long-term estimate of variance in energy level (Q8) 762 tmp32 = WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_16_16(dB, dB), 12); 763 tmp32 += WEBRTC_SPL_MUL(state->varianceLongTerm, state->counter); 764 state->varianceLongTerm = WebRtcSpl_DivW32W16(tmp32, 765 WEBRTC_SPL_ADD_SAT_W16(state->counter, 1)); 766 767 // update long-term estimate of standard deviation in energy level (Q10) 768 tmp32 = WEBRTC_SPL_MUL_16_16(state->meanLongTerm, state->meanLongTerm); 769 tmp32 = WEBRTC_SPL_LSHIFT_W32(state->varianceLongTerm, 12) - tmp32; 770 state->stdLongTerm = (int16_t)WebRtcSpl_Sqrt(tmp32); 771 772 // update voice activity measure (Q10) 773 tmp16 = WEBRTC_SPL_LSHIFT_W16(3, 12); 774 tmp32 = WEBRTC_SPL_MUL_16_16(tmp16, (dB - state->meanLongTerm)); 775 tmp32 = WebRtcSpl_DivW32W16(tmp32, state->stdLongTerm); 776 tmpU16 = (13 << 12); 777 tmp32b = WEBRTC_SPL_MUL_16_U16(state->logRatio, tmpU16); 778 tmp32 += WEBRTC_SPL_RSHIFT_W32(tmp32b, 10); 779 780 state->logRatio = (int16_t)WEBRTC_SPL_RSHIFT_W32(tmp32, 6); 781 782 // limit 783 if (state->logRatio > 2048) 784 { 785 state->logRatio = 2048; 786 } 787 if (state->logRatio < -2048) 788 { 789 state->logRatio = -2048; 790 } 791 792 return state->logRatio; // Q10 793 } 794
