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 "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 "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 WebRtc_UWord16 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 WebRtc_Word16 kAvgDecayTime = 250; // frames; < 3000 60 61 WebRtc_Word32 WebRtcAgc_CalculateGainTable(WebRtc_Word32 *gainTable, // Q16 62 WebRtc_Word16 digCompGaindB, // Q0 63 WebRtc_Word16 targetLevelDbfs,// Q0 64 WebRtc_UWord8 limiterEnable, 65 WebRtc_Word16 analogTarget) // Q0 66 { 67 // This function generates the compressor gain table used in the fixed digital part. 68 WebRtc_UWord32 tmpU32no1, tmpU32no2, absInLevel, logApprox; 69 WebRtc_Word32 inLevel, limiterLvl; 70 WebRtc_Word32 tmp32, tmp32no1, tmp32no2, numFIX, den, y32; 71 const WebRtc_UWord16 kLog10 = 54426; // log2(10) in Q14 72 const WebRtc_UWord16 kLog10_2 = 49321; // 10*log10(2) in Q14 73 const WebRtc_UWord16 kLogE_1 = 23637; // log2(e) in Q14 74 WebRtc_UWord16 constMaxGain; 75 WebRtc_UWord16 tmpU16, intPart, fracPart; 76 const WebRtc_Word16 kCompRatio = 3; 77 const WebRtc_Word16 kSoftLimiterLeft = 1; 78 WebRtc_Word16 limiterOffset = 0; // Limiter offset 79 WebRtc_Word16 limiterIdx, limiterLvlX; 80 WebRtc_Word16 constLinApprox, zeroGainLvl, maxGain, diffGain; 81 WebRtc_Word16 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((WebRtc_Word32)limiterLvlX, 13), 121 WEBRTC_SPL_RSHIFT_U16(kLog10_2, 1)); 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 = (WebRtc_Word16)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((WebRtc_Word32)diffGain, 14) - inLevel; // Q14 148 149 // Make calculations on abs(inLevel) and compensate for the sign afterwards. 150 absInLevel = (WebRtc_UWord32)WEBRTC_SPL_ABS_W32(inLevel); // Q14 151 152 // LUT with interpolation 153 intPart = (WebRtc_UWord16)WEBRTC_SPL_RSHIFT_U32(absInLevel, 14); 154 fracPart = (WebRtc_UWord16)(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((WebRtc_UWord32)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((WebRtc_Word32)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 = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(tmp32, 14); 235 fracPart = (WebRtc_UWord16)(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 = (WebRtc_UWord16)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 WebRtc_Word32 WebRtcAgc_InitDigital(DigitalAgc_t *stt, WebRtc_Word16 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; // (WebRtc_Word32)(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 WebRtc_Word32 WebRtcAgc_AddFarendToDigital(DigitalAgc_t *stt, const WebRtc_Word16 *in_far, 289 WebRtc_Word16 nrSamples) 290 { 291 // Check for valid pointer 292 if (&stt->vadFarend == NULL) 293 { 294 return -1; 295 } 296 297 // VAD for far end 298 WebRtcAgc_ProcessVad(&stt->vadFarend, in_far, nrSamples); 299 300 return 0; 301 } 302 303 WebRtc_Word32 WebRtcAgc_ProcessDigital(DigitalAgc_t *stt, const WebRtc_Word16 *in_near, 304 const WebRtc_Word16 *in_near_H, WebRtc_Word16 *out, 305 WebRtc_Word16 *out_H, WebRtc_UWord32 FS, 306 WebRtc_Word16 lowlevelSignal) 307 { 308 // array for gains (one value per ms, incl start & end) 309 WebRtc_Word32 gains[11]; 310 311 WebRtc_Word32 out_tmp, tmp32; 312 WebRtc_Word32 env[10]; 313 WebRtc_Word32 nrg, max_nrg; 314 WebRtc_Word32 cur_level; 315 WebRtc_Word32 gain32, delta; 316 WebRtc_Word16 logratio; 317 WebRtc_Word16 lower_thr, upper_thr; 318 WebRtc_Word16 zeros, zeros_fast, frac; 319 WebRtc_Word16 decay; 320 WebRtc_Word16 gate, gain_adj; 321 WebRtc_Word16 k, n; 322 WebRtc_Word16 L, L2; // samples/subframe 323 324 // determine number of samples per ms 325 if (FS == 8000) 326 { 327 L = 8; 328 L2 = 3; 329 } else if (FS == 16000) 330 { 331 L = 16; 332 L2 = 4; 333 } else if (FS == 32000) 334 { 335 L = 16; 336 L2 = 4; 337 } else 338 { 339 return -1; 340 } 341 342 // TODO(andrew): again, we don't need input and output pointers... 343 if (in_near != out) 344 { 345 // Only needed if they don't already point to the same place. 346 memcpy(out, in_near, 10 * L * sizeof(WebRtc_Word16)); 347 } 348 if (FS == 32000) 349 { 350 if (in_near_H != out_H) 351 { 352 memcpy(out_H, in_near_H, 10 * L * sizeof(WebRtc_Word16)); 353 } 354 } 355 // VAD for near end 356 logratio = WebRtcAgc_ProcessVad(&stt->vadNearend, out, L * 10); 357 358 // Account for far end VAD 359 if (stt->vadFarend.counter > 10) 360 { 361 tmp32 = WEBRTC_SPL_MUL_16_16(3, logratio); 362 logratio = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(tmp32 - stt->vadFarend.logRatio, 2); 363 } 364 365 // Determine decay factor depending on VAD 366 // upper_thr = 1.0f; 367 // lower_thr = 0.25f; 368 upper_thr = 1024; // Q10 369 lower_thr = 0; // Q10 370 if (logratio > upper_thr) 371 { 372 // decay = -2^17 / DecayTime; -> -65 373 decay = -65; 374 } else if (logratio < lower_thr) 375 { 376 decay = 0; 377 } else 378 { 379 // decay = (WebRtc_Word16)(((lower_thr - logratio) 380 // * (2^27/(DecayTime*(upper_thr-lower_thr)))) >> 10); 381 // SUBSTITUTED: 2^27/(DecayTime*(upper_thr-lower_thr)) -> 65 382 tmp32 = WEBRTC_SPL_MUL_16_16((lower_thr - logratio), 65); 383 decay = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(tmp32, 10); 384 } 385 386 // adjust decay factor for long silence (detected as low standard deviation) 387 // This is only done in the adaptive modes 388 if (stt->agcMode != kAgcModeFixedDigital) 389 { 390 if (stt->vadNearend.stdLongTerm < 4000) 391 { 392 decay = 0; 393 } else if (stt->vadNearend.stdLongTerm < 8096) 394 { 395 // decay = (WebRtc_Word16)(((stt->vadNearend.stdLongTerm - 4000) * decay) >> 12); 396 tmp32 = WEBRTC_SPL_MUL_16_16((stt->vadNearend.stdLongTerm - 4000), decay); 397 decay = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(tmp32, 12); 398 } 399 400 if (lowlevelSignal != 0) 401 { 402 decay = 0; 403 } 404 } 405 #ifdef AGC_DEBUG 406 stt->frameCounter++; 407 fprintf(stt->logFile, "%5.2f\t%d\t%d\t%d\t", (float)(stt->frameCounter) / 100, logratio, decay, stt->vadNearend.stdLongTerm); 408 #endif 409 // Find max amplitude per sub frame 410 // iterate over sub frames 411 for (k = 0; k < 10; k++) 412 { 413 // iterate over samples 414 max_nrg = 0; 415 for (n = 0; n < L; n++) 416 { 417 nrg = WEBRTC_SPL_MUL_16_16(out[k * L + n], out[k * L + n]); 418 if (nrg > max_nrg) 419 { 420 max_nrg = nrg; 421 } 422 } 423 env[k] = max_nrg; 424 } 425 426 // Calculate gain per sub frame 427 gains[0] = stt->gain; 428 for (k = 0; k < 10; k++) 429 { 430 // Fast envelope follower 431 // decay time = -131000 / -1000 = 131 (ms) 432 stt->capacitorFast = AGC_SCALEDIFF32(-1000, stt->capacitorFast, stt->capacitorFast); 433 if (env[k] > stt->capacitorFast) 434 { 435 stt->capacitorFast = env[k]; 436 } 437 // Slow envelope follower 438 if (env[k] > stt->capacitorSlow) 439 { 440 // increase capacitorSlow 441 stt->capacitorSlow 442 = AGC_SCALEDIFF32(500, (env[k] - stt->capacitorSlow), stt->capacitorSlow); 443 } else 444 { 445 // decrease capacitorSlow 446 stt->capacitorSlow 447 = AGC_SCALEDIFF32(decay, stt->capacitorSlow, stt->capacitorSlow); 448 } 449 450 // use maximum of both capacitors as current level 451 if (stt->capacitorFast > stt->capacitorSlow) 452 { 453 cur_level = stt->capacitorFast; 454 } else 455 { 456 cur_level = stt->capacitorSlow; 457 } 458 // Translate signal level into gain, using a piecewise linear approximation 459 // find number of leading zeros 460 zeros = WebRtcSpl_NormU32((WebRtc_UWord32)cur_level); 461 if (cur_level == 0) 462 { 463 zeros = 31; 464 } 465 tmp32 = (WEBRTC_SPL_LSHIFT_W32(cur_level, zeros) & 0x7FFFFFFF); 466 frac = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(tmp32, 19); // Q12 467 tmp32 = WEBRTC_SPL_MUL((stt->gainTable[zeros-1] - stt->gainTable[zeros]), frac); 468 gains[k + 1] = stt->gainTable[zeros] + WEBRTC_SPL_RSHIFT_W32(tmp32, 12); 469 #ifdef AGC_DEBUG 470 if (k == 0) 471 { 472 fprintf(stt->logFile, "%d\t%d\t%d\t%d\t%d\n", env[0], cur_level, stt->capacitorFast, stt->capacitorSlow, zeros); 473 } 474 #endif 475 } 476 477 // Gate processing (lower gain during absence of speech) 478 zeros = WEBRTC_SPL_LSHIFT_W16(zeros, 9) - WEBRTC_SPL_RSHIFT_W16(frac, 3); 479 // find number of leading zeros 480 zeros_fast = WebRtcSpl_NormU32((WebRtc_UWord32)stt->capacitorFast); 481 if (stt->capacitorFast == 0) 482 { 483 zeros_fast = 31; 484 } 485 tmp32 = (WEBRTC_SPL_LSHIFT_W32(stt->capacitorFast, zeros_fast) & 0x7FFFFFFF); 486 zeros_fast = WEBRTC_SPL_LSHIFT_W16(zeros_fast, 9); 487 zeros_fast -= (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(tmp32, 22); 488 489 gate = 1000 + zeros_fast - zeros - stt->vadNearend.stdShortTerm; 490 491 if (gate < 0) 492 { 493 stt->gatePrevious = 0; 494 } else 495 { 496 tmp32 = WEBRTC_SPL_MUL_16_16(stt->gatePrevious, 7); 497 gate = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32((WebRtc_Word32)gate + tmp32, 3); 498 stt->gatePrevious = gate; 499 } 500 // gate < 0 -> no gate 501 // gate > 2500 -> max gate 502 if (gate > 0) 503 { 504 if (gate < 2500) 505 { 506 gain_adj = WEBRTC_SPL_RSHIFT_W16(2500 - gate, 5); 507 } else 508 { 509 gain_adj = 0; 510 } 511 for (k = 0; k < 10; k++) 512 { 513 if ((gains[k + 1] - stt->gainTable[0]) > 8388608) 514 { 515 // To prevent wraparound 516 tmp32 = WEBRTC_SPL_RSHIFT_W32((gains[k+1] - stt->gainTable[0]), 8); 517 tmp32 = WEBRTC_SPL_MUL(tmp32, (178 + gain_adj)); 518 } else 519 { 520 tmp32 = WEBRTC_SPL_MUL((gains[k+1] - stt->gainTable[0]), (178 + gain_adj)); 521 tmp32 = WEBRTC_SPL_RSHIFT_W32(tmp32, 8); 522 } 523 gains[k + 1] = stt->gainTable[0] + tmp32; 524 } 525 } 526 527 // Limit gain to avoid overload distortion 528 for (k = 0; k < 10; k++) 529 { 530 // To prevent wrap around 531 zeros = 10; 532 if (gains[k + 1] > 47453132) 533 { 534 zeros = 16 - WebRtcSpl_NormW32(gains[k + 1]); 535 } 536 gain32 = WEBRTC_SPL_RSHIFT_W32(gains[k+1], zeros) + 1; 537 gain32 = WEBRTC_SPL_MUL(gain32, gain32); 538 // check for overflow 539 while (AGC_MUL32(WEBRTC_SPL_RSHIFT_W32(env[k], 12) + 1, gain32) 540 > WEBRTC_SPL_SHIFT_W32((WebRtc_Word32)32767, 2 * (1 - zeros + 10))) 541 { 542 // multiply by 253/256 ==> -0.1 dB 543 if (gains[k + 1] > 8388607) 544 { 545 // Prevent wrap around 546 gains[k + 1] = WEBRTC_SPL_MUL(WEBRTC_SPL_RSHIFT_W32(gains[k+1], 8), 253); 547 } else 548 { 549 gains[k + 1] = WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL(gains[k+1], 253), 8); 550 } 551 gain32 = WEBRTC_SPL_RSHIFT_W32(gains[k+1], zeros) + 1; 552 gain32 = WEBRTC_SPL_MUL(gain32, gain32); 553 } 554 } 555 // gain reductions should be done 1 ms earlier than gain increases 556 for (k = 1; k < 10; k++) 557 { 558 if (gains[k] > gains[k + 1]) 559 { 560 gains[k] = gains[k + 1]; 561 } 562 } 563 // save start gain for next frame 564 stt->gain = gains[10]; 565 566 // Apply gain 567 // handle first sub frame separately 568 delta = WEBRTC_SPL_LSHIFT_W32(gains[1] - gains[0], (4 - L2)); 569 gain32 = WEBRTC_SPL_LSHIFT_W32(gains[0], 4); 570 // iterate over samples 571 for (n = 0; n < L; n++) 572 { 573 // For lower band 574 tmp32 = WEBRTC_SPL_MUL((WebRtc_Word32)out[n], WEBRTC_SPL_RSHIFT_W32(gain32 + 127, 7)); 575 out_tmp = WEBRTC_SPL_RSHIFT_W32(tmp32 , 16); 576 if (out_tmp > 4095) 577 { 578 out[n] = (WebRtc_Word16)32767; 579 } else if (out_tmp < -4096) 580 { 581 out[n] = (WebRtc_Word16)-32768; 582 } else 583 { 584 tmp32 = WEBRTC_SPL_MUL((WebRtc_Word32)out[n], WEBRTC_SPL_RSHIFT_W32(gain32, 4)); 585 out[n] = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(tmp32 , 16); 586 } 587 // For higher band 588 if (FS == 32000) 589 { 590 tmp32 = WEBRTC_SPL_MUL((WebRtc_Word32)out_H[n], 591 WEBRTC_SPL_RSHIFT_W32(gain32 + 127, 7)); 592 out_tmp = WEBRTC_SPL_RSHIFT_W32(tmp32 , 16); 593 if (out_tmp > 4095) 594 { 595 out_H[n] = (WebRtc_Word16)32767; 596 } else if (out_tmp < -4096) 597 { 598 out_H[n] = (WebRtc_Word16)-32768; 599 } else 600 { 601 tmp32 = WEBRTC_SPL_MUL((WebRtc_Word32)out_H[n], 602 WEBRTC_SPL_RSHIFT_W32(gain32, 4)); 603 out_H[n] = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(tmp32 , 16); 604 } 605 } 606 // 607 608 gain32 += delta; 609 } 610 // iterate over subframes 611 for (k = 1; k < 10; k++) 612 { 613 delta = WEBRTC_SPL_LSHIFT_W32(gains[k+1] - gains[k], (4 - L2)); 614 gain32 = WEBRTC_SPL_LSHIFT_W32(gains[k], 4); 615 // iterate over samples 616 for (n = 0; n < L; n++) 617 { 618 // For lower band 619 tmp32 = WEBRTC_SPL_MUL((WebRtc_Word32)out[k * L + n], 620 WEBRTC_SPL_RSHIFT_W32(gain32, 4)); 621 out[k * L + n] = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(tmp32 , 16); 622 // For higher band 623 if (FS == 32000) 624 { 625 tmp32 = WEBRTC_SPL_MUL((WebRtc_Word32)out_H[k * L + n], 626 WEBRTC_SPL_RSHIFT_W32(gain32, 4)); 627 out_H[k * L + n] = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(tmp32 , 16); 628 } 629 gain32 += delta; 630 } 631 } 632 633 return 0; 634 } 635 636 void WebRtcAgc_InitVad(AgcVad_t *state) 637 { 638 WebRtc_Word16 k; 639 640 state->HPstate = 0; // state of high pass filter 641 state->logRatio = 0; // log( P(active) / P(inactive) ) 642 // average input level (Q10) 643 state->meanLongTerm = WEBRTC_SPL_LSHIFT_W16(15, 10); 644 645 // variance of input level (Q8) 646 state->varianceLongTerm = WEBRTC_SPL_LSHIFT_W32(500, 8); 647 648 state->stdLongTerm = 0; // standard deviation of input level in dB 649 // short-term average input level (Q10) 650 state->meanShortTerm = WEBRTC_SPL_LSHIFT_W16(15, 10); 651 652 // short-term variance of input level (Q8) 653 state->varianceShortTerm = WEBRTC_SPL_LSHIFT_W32(500, 8); 654 655 state->stdShortTerm = 0; // short-term standard deviation of input level in dB 656 state->counter = 3; // counts updates 657 for (k = 0; k < 8; k++) 658 { 659 // downsampling filter 660 state->downState[k] = 0; 661 } 662 } 663 664 WebRtc_Word16 WebRtcAgc_ProcessVad(AgcVad_t *state, // (i) VAD state 665 const WebRtc_Word16 *in, // (i) Speech signal 666 WebRtc_Word16 nrSamples) // (i) number of samples 667 { 668 WebRtc_Word32 out, nrg, tmp32, tmp32b; 669 WebRtc_UWord16 tmpU16; 670 WebRtc_Word16 k, subfr, tmp16; 671 WebRtc_Word16 buf1[8]; 672 WebRtc_Word16 buf2[4]; 673 WebRtc_Word16 HPstate; 674 WebRtc_Word16 zeros, dB; 675 676 // process in 10 sub frames of 1 ms (to save on memory) 677 nrg = 0; 678 HPstate = state->HPstate; 679 for (subfr = 0; subfr < 10; subfr++) 680 { 681 // downsample to 4 kHz 682 if (nrSamples == 160) 683 { 684 for (k = 0; k < 8; k++) 685 { 686 tmp32 = (WebRtc_Word32)in[2 * k] + (WebRtc_Word32)in[2 * k + 1]; 687 tmp32 = WEBRTC_SPL_RSHIFT_W32(tmp32, 1); 688 buf1[k] = (WebRtc_Word16)tmp32; 689 } 690 in += 16; 691 692 WebRtcSpl_DownsampleBy2(buf1, 8, buf2, state->downState); 693 } else 694 { 695 WebRtcSpl_DownsampleBy2(in, 8, buf2, state->downState); 696 in += 8; 697 } 698 699 // high pass filter and compute energy 700 for (k = 0; k < 4; k++) 701 { 702 out = buf2[k] + HPstate; 703 tmp32 = WEBRTC_SPL_MUL(600, out); 704 HPstate = (WebRtc_Word16)(WEBRTC_SPL_RSHIFT_W32(tmp32, 10) - buf2[k]); 705 tmp32 = WEBRTC_SPL_MUL(out, out); 706 nrg += WEBRTC_SPL_RSHIFT_W32(tmp32, 6); 707 } 708 } 709 state->HPstate = HPstate; 710 711 // find number of leading zeros 712 if (!(0xFFFF0000 & nrg)) 713 { 714 zeros = 16; 715 } else 716 { 717 zeros = 0; 718 } 719 if (!(0xFF000000 & (nrg << zeros))) 720 { 721 zeros += 8; 722 } 723 if (!(0xF0000000 & (nrg << zeros))) 724 { 725 zeros += 4; 726 } 727 if (!(0xC0000000 & (nrg << zeros))) 728 { 729 zeros += 2; 730 } 731 if (!(0x80000000 & (nrg << zeros))) 732 { 733 zeros += 1; 734 } 735 736 // energy level (range {-32..30}) (Q10) 737 dB = WEBRTC_SPL_LSHIFT_W16(15 - zeros, 11); 738 739 // Update statistics 740 741 if (state->counter < kAvgDecayTime) 742 { 743 // decay time = AvgDecTime * 10 ms 744 state->counter++; 745 } 746 747 // update short-term estimate of mean energy level (Q10) 748 tmp32 = (WEBRTC_SPL_MUL_16_16(state->meanShortTerm, 15) + (WebRtc_Word32)dB); 749 state->meanShortTerm = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(tmp32, 4); 750 751 // update short-term estimate of variance in energy level (Q8) 752 tmp32 = WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_16_16(dB, dB), 12); 753 tmp32 += WEBRTC_SPL_MUL(state->varianceShortTerm, 15); 754 state->varianceShortTerm = WEBRTC_SPL_RSHIFT_W32(tmp32, 4); 755 756 // update short-term estimate of standard deviation in energy level (Q10) 757 tmp32 = WEBRTC_SPL_MUL_16_16(state->meanShortTerm, state->meanShortTerm); 758 tmp32 = WEBRTC_SPL_LSHIFT_W32(state->varianceShortTerm, 12) - tmp32; 759 state->stdShortTerm = (WebRtc_Word16)WebRtcSpl_Sqrt(tmp32); 760 761 // update long-term estimate of mean energy level (Q10) 762 tmp32 = WEBRTC_SPL_MUL_16_16(state->meanLongTerm, state->counter) + (WebRtc_Word32)dB; 763 state->meanLongTerm = WebRtcSpl_DivW32W16ResW16(tmp32, 764 WEBRTC_SPL_ADD_SAT_W16(state->counter, 1)); 765 766 // update long-term estimate of variance in energy level (Q8) 767 tmp32 = WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_16_16(dB, dB), 12); 768 tmp32 += WEBRTC_SPL_MUL(state->varianceLongTerm, state->counter); 769 state->varianceLongTerm = WebRtcSpl_DivW32W16(tmp32, 770 WEBRTC_SPL_ADD_SAT_W16(state->counter, 1)); 771 772 // update long-term estimate of standard deviation in energy level (Q10) 773 tmp32 = WEBRTC_SPL_MUL_16_16(state->meanLongTerm, state->meanLongTerm); 774 tmp32 = WEBRTC_SPL_LSHIFT_W32(state->varianceLongTerm, 12) - tmp32; 775 state->stdLongTerm = (WebRtc_Word16)WebRtcSpl_Sqrt(tmp32); 776 777 // update voice activity measure (Q10) 778 tmp16 = WEBRTC_SPL_LSHIFT_W16(3, 12); 779 tmp32 = WEBRTC_SPL_MUL_16_16(tmp16, (dB - state->meanLongTerm)); 780 tmp32 = WebRtcSpl_DivW32W16(tmp32, state->stdLongTerm); 781 tmpU16 = WEBRTC_SPL_LSHIFT_U16((WebRtc_UWord16)13, 12); 782 tmp32b = WEBRTC_SPL_MUL_16_U16(state->logRatio, tmpU16); 783 tmp32 += WEBRTC_SPL_RSHIFT_W32(tmp32b, 10); 784 785 state->logRatio = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(tmp32, 6); 786 787 // limit 788 if (state->logRatio > 2048) 789 { 790 state->logRatio = 2048; 791 } 792 if (state->logRatio < -2048) 793 { 794 state->logRatio = -2048; 795 } 796 797 return state->logRatio; // Q10 798 } 799
