/* * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. * * Use of this source code is governed by a BSD-style license * that can be found in the LICENSE file in the root of the source * tree. An additional intellectual property rights grant can be found * in the file PATENTS. All contributing project authors may * be found in the AUTHORS file in the root of the source tree. */ /* * The core AEC algorithm, which is presented with time-aligned signals. */ #include "webrtc/modules/audio_processing/aec/aec_core.h" #ifdef WEBRTC_AEC_DEBUG_DUMP #include #endif #include #include #include // size_t #include #include #include "webrtc/common_audio/ring_buffer.h" #include "webrtc/common_audio/signal_processing/include/signal_processing_library.h" #include "webrtc/modules/audio_processing/aec/aec_common.h" #include "webrtc/modules/audio_processing/aec/aec_core_internal.h" #include "webrtc/modules/audio_processing/aec/aec_rdft.h" #include "webrtc/modules/audio_processing/logging/aec_logging.h" #include "webrtc/modules/audio_processing/utility/delay_estimator_wrapper.h" #include "webrtc/system_wrappers/include/cpu_features_wrapper.h" #include "webrtc/typedefs.h" // Buffer size (samples) static const size_t kBufSizePartitions = 250; // 1 second of audio in 16 kHz. // Metrics static const int subCountLen = 4; static const int countLen = 50; static const int kDelayMetricsAggregationWindow = 1250; // 5 seconds at 16 kHz. // Quantities to control H band scaling for SWB input static const float cnScaleHband = (float)0.4; // scale for comfort noise in H band // Initial bin for averaging nlp gain in low band static const int freqAvgIc = PART_LEN / 2; // Matlab code to produce table: // win = sqrt(hanning(63)); win = [0 ; win(1:32)]; // fprintf(1, '\t%.14f, %.14f, %.14f,\n', win); ALIGN16_BEG const float ALIGN16_END WebRtcAec_sqrtHanning[65] = { 0.00000000000000f, 0.02454122852291f, 0.04906767432742f, 0.07356456359967f, 0.09801714032956f, 0.12241067519922f, 0.14673047445536f, 0.17096188876030f, 0.19509032201613f, 0.21910124015687f, 0.24298017990326f, 0.26671275747490f, 0.29028467725446f, 0.31368174039889f, 0.33688985339222f, 0.35989503653499f, 0.38268343236509f, 0.40524131400499f, 0.42755509343028f, 0.44961132965461f, 0.47139673682600f, 0.49289819222978f, 0.51410274419322f, 0.53499761988710f, 0.55557023301960f, 0.57580819141785f, 0.59569930449243f, 0.61523159058063f, 0.63439328416365f, 0.65317284295378f, 0.67155895484702f, 0.68954054473707f, 0.70710678118655f, 0.72424708295147f, 0.74095112535496f, 0.75720884650648f, 0.77301045336274f, 0.78834642762661f, 0.80320753148064f, 0.81758481315158f, 0.83146961230255f, 0.84485356524971f, 0.85772861000027f, 0.87008699110871f, 0.88192126434835f, 0.89322430119552f, 0.90398929312344f, 0.91420975570353f, 0.92387953251129f, 0.93299279883474f, 0.94154406518302f, 0.94952818059304f, 0.95694033573221f, 0.96377606579544f, 0.97003125319454f, 0.97570213003853f, 0.98078528040323f, 0.98527764238894f, 0.98917650996478f, 0.99247953459871f, 0.99518472667220f, 0.99729045667869f, 0.99879545620517f, 0.99969881869620f, 1.00000000000000f}; // Matlab code to produce table: // weightCurve = [0 ; 0.3 * sqrt(linspace(0,1,64))' + 0.1]; // fprintf(1, '\t%.4f, %.4f, %.4f, %.4f, %.4f, %.4f,\n', weightCurve); ALIGN16_BEG const float ALIGN16_END WebRtcAec_weightCurve[65] = { 0.0000f, 0.1000f, 0.1378f, 0.1535f, 0.1655f, 0.1756f, 0.1845f, 0.1926f, 0.2000f, 0.2069f, 0.2134f, 0.2195f, 0.2254f, 0.2309f, 0.2363f, 0.2414f, 0.2464f, 0.2512f, 0.2558f, 0.2604f, 0.2648f, 0.2690f, 0.2732f, 0.2773f, 0.2813f, 0.2852f, 0.2890f, 0.2927f, 0.2964f, 0.3000f, 0.3035f, 0.3070f, 0.3104f, 0.3138f, 0.3171f, 0.3204f, 0.3236f, 0.3268f, 0.3299f, 0.3330f, 0.3360f, 0.3390f, 0.3420f, 0.3449f, 0.3478f, 0.3507f, 0.3535f, 0.3563f, 0.3591f, 0.3619f, 0.3646f, 0.3673f, 0.3699f, 0.3726f, 0.3752f, 0.3777f, 0.3803f, 0.3828f, 0.3854f, 0.3878f, 0.3903f, 0.3928f, 0.3952f, 0.3976f, 0.4000f}; // Matlab code to produce table: // overDriveCurve = [sqrt(linspace(0,1,65))' + 1]; // fprintf(1, '\t%.4f, %.4f, %.4f, %.4f, %.4f, %.4f,\n', overDriveCurve); ALIGN16_BEG const float ALIGN16_END WebRtcAec_overDriveCurve[65] = { 1.0000f, 1.1250f, 1.1768f, 1.2165f, 1.2500f, 1.2795f, 1.3062f, 1.3307f, 1.3536f, 1.3750f, 1.3953f, 1.4146f, 1.4330f, 1.4507f, 1.4677f, 1.4841f, 1.5000f, 1.5154f, 1.5303f, 1.5449f, 1.5590f, 1.5728f, 1.5863f, 1.5995f, 1.6124f, 1.6250f, 1.6374f, 1.6495f, 1.6614f, 1.6731f, 1.6847f, 1.6960f, 1.7071f, 1.7181f, 1.7289f, 1.7395f, 1.7500f, 1.7603f, 1.7706f, 1.7806f, 1.7906f, 1.8004f, 1.8101f, 1.8197f, 1.8292f, 1.8385f, 1.8478f, 1.8570f, 1.8660f, 1.8750f, 1.8839f, 1.8927f, 1.9014f, 1.9100f, 1.9186f, 1.9270f, 1.9354f, 1.9437f, 1.9520f, 1.9601f, 1.9682f, 1.9763f, 1.9843f, 1.9922f, 2.0000f}; // Delay Agnostic AEC parameters, still under development and may change. static const float kDelayQualityThresholdMax = 0.07f; static const float kDelayQualityThresholdMin = 0.01f; static const int kInitialShiftOffset = 5; #if !defined(WEBRTC_ANDROID) static const int kDelayCorrectionStart = 1500; // 10 ms chunks #endif // Target suppression levels for nlp modes. // log{0.001, 0.00001, 0.00000001} static const float kTargetSupp[3] = {-6.9f, -11.5f, -18.4f}; // Two sets of parameters, one for the extended filter mode. static const float kExtendedMinOverDrive[3] = {3.0f, 6.0f, 15.0f}; static const float kNormalMinOverDrive[3] = {1.0f, 2.0f, 5.0f}; const float WebRtcAec_kExtendedSmoothingCoefficients[2][2] = {{0.9f, 0.1f}, {0.92f, 0.08f}}; const float WebRtcAec_kNormalSmoothingCoefficients[2][2] = {{0.9f, 0.1f}, {0.93f, 0.07f}}; // Number of partitions forming the NLP's "preferred" bands. enum { kPrefBandSize = 24 }; #ifdef WEBRTC_AEC_DEBUG_DUMP extern int webrtc_aec_instance_count; #endif WebRtcAecFilterFar WebRtcAec_FilterFar; WebRtcAecScaleErrorSignal WebRtcAec_ScaleErrorSignal; WebRtcAecFilterAdaptation WebRtcAec_FilterAdaptation; WebRtcAecOverdriveAndSuppress WebRtcAec_OverdriveAndSuppress; WebRtcAecComfortNoise WebRtcAec_ComfortNoise; WebRtcAecSubBandCoherence WebRtcAec_SubbandCoherence; WebRtcAecStoreAsComplex WebRtcAec_StoreAsComplex; WebRtcAecPartitionDelay WebRtcAec_PartitionDelay; WebRtcAecWindowData WebRtcAec_WindowData; __inline static float MulRe(float aRe, float aIm, float bRe, float bIm) { return aRe * bRe - aIm * bIm; } __inline static float MulIm(float aRe, float aIm, float bRe, float bIm) { return aRe * bIm + aIm * bRe; } static int CmpFloat(const void* a, const void* b) { const float* da = (const float*)a; const float* db = (const float*)b; return (*da > *db) - (*da < *db); } static void FilterFar( int num_partitions, int x_fft_buf_block_pos, float x_fft_buf[2][kExtendedNumPartitions * PART_LEN1], float h_fft_buf[2][kExtendedNumPartitions * PART_LEN1], float y_fft[2][PART_LEN1]) { int i; for (i = 0; i < num_partitions; i++) { int j; int xPos = (i + x_fft_buf_block_pos) * PART_LEN1; int pos = i * PART_LEN1; // Check for wrap if (i + x_fft_buf_block_pos >= num_partitions) { xPos -= num_partitions * (PART_LEN1); } for (j = 0; j < PART_LEN1; j++) { y_fft[0][j] += MulRe(x_fft_buf[0][xPos + j], x_fft_buf[1][xPos + j], h_fft_buf[0][pos + j], h_fft_buf[1][pos + j]); y_fft[1][j] += MulIm(x_fft_buf[0][xPos + j], x_fft_buf[1][xPos + j], h_fft_buf[0][pos + j], h_fft_buf[1][pos + j]); } } } static void ScaleErrorSignal(int extended_filter_enabled, float normal_mu, float normal_error_threshold, float x_pow[PART_LEN1], float ef[2][PART_LEN1]) { const float mu = extended_filter_enabled ? kExtendedMu : normal_mu; const float error_threshold = extended_filter_enabled ? kExtendedErrorThreshold : normal_error_threshold; int i; float abs_ef; for (i = 0; i < (PART_LEN1); i++) { ef[0][i] /= (x_pow[i] + 1e-10f); ef[1][i] /= (x_pow[i] + 1e-10f); abs_ef = sqrtf(ef[0][i] * ef[0][i] + ef[1][i] * ef[1][i]); if (abs_ef > error_threshold) { abs_ef = error_threshold / (abs_ef + 1e-10f); ef[0][i] *= abs_ef; ef[1][i] *= abs_ef; } // Stepsize factor ef[0][i] *= mu; ef[1][i] *= mu; } } static void FilterAdaptation( int num_partitions, int x_fft_buf_block_pos, float x_fft_buf[2][kExtendedNumPartitions * PART_LEN1], float e_fft[2][PART_LEN1], float h_fft_buf[2][kExtendedNumPartitions * PART_LEN1]) { int i, j; float fft[PART_LEN2]; for (i = 0; i < num_partitions; i++) { int xPos = (i + x_fft_buf_block_pos) * (PART_LEN1); int pos; // Check for wrap if (i + x_fft_buf_block_pos >= num_partitions) { xPos -= num_partitions * PART_LEN1; } pos = i * PART_LEN1; for (j = 0; j < PART_LEN; j++) { fft[2 * j] = MulRe(x_fft_buf[0][xPos + j], -x_fft_buf[1][xPos + j], e_fft[0][j], e_fft[1][j]); fft[2 * j + 1] = MulIm(x_fft_buf[0][xPos + j], -x_fft_buf[1][xPos + j], e_fft[0][j], e_fft[1][j]); } fft[1] = MulRe(x_fft_buf[0][xPos + PART_LEN], -x_fft_buf[1][xPos + PART_LEN], e_fft[0][PART_LEN], e_fft[1][PART_LEN]); aec_rdft_inverse_128(fft); memset(fft + PART_LEN, 0, sizeof(float) * PART_LEN); // fft scaling { float scale = 2.0f / PART_LEN2; for (j = 0; j < PART_LEN; j++) { fft[j] *= scale; } } aec_rdft_forward_128(fft); h_fft_buf[0][pos] += fft[0]; h_fft_buf[0][pos + PART_LEN] += fft[1]; for (j = 1; j < PART_LEN; j++) { h_fft_buf[0][pos + j] += fft[2 * j]; h_fft_buf[1][pos + j] += fft[2 * j + 1]; } } } static void OverdriveAndSuppress(AecCore* aec, float hNl[PART_LEN1], const float hNlFb, float efw[2][PART_LEN1]) { int i; for (i = 0; i < PART_LEN1; i++) { // Weight subbands if (hNl[i] > hNlFb) { hNl[i] = WebRtcAec_weightCurve[i] * hNlFb + (1 - WebRtcAec_weightCurve[i]) * hNl[i]; } hNl[i] = powf(hNl[i], aec->overDriveSm * WebRtcAec_overDriveCurve[i]); // Suppress error signal efw[0][i] *= hNl[i]; efw[1][i] *= hNl[i]; // Ooura fft returns incorrect sign on imaginary component. It matters here // because we are making an additive change with comfort noise. efw[1][i] *= -1; } } static int PartitionDelay(const AecCore* aec) { // Measures the energy in each filter partition and returns the partition with // highest energy. // TODO(bjornv): Spread computational cost by computing one partition per // block? float wfEnMax = 0; int i; int delay = 0; for (i = 0; i < aec->num_partitions; i++) { int j; int pos = i * PART_LEN1; float wfEn = 0; for (j = 0; j < PART_LEN1; j++) { wfEn += aec->wfBuf[0][pos + j] * aec->wfBuf[0][pos + j] + aec->wfBuf[1][pos + j] * aec->wfBuf[1][pos + j]; } if (wfEn > wfEnMax) { wfEnMax = wfEn; delay = i; } } return delay; } // Threshold to protect against the ill-effects of a zero far-end. const float WebRtcAec_kMinFarendPSD = 15; // Updates the following smoothed Power Spectral Densities (PSD): // - sd : near-end // - se : residual echo // - sx : far-end // - sde : cross-PSD of near-end and residual echo // - sxd : cross-PSD of near-end and far-end // // In addition to updating the PSDs, also the filter diverge state is // determined. static void SmoothedPSD(AecCore* aec, float efw[2][PART_LEN1], float dfw[2][PART_LEN1], float xfw[2][PART_LEN1], int* extreme_filter_divergence) { // Power estimate smoothing coefficients. const float* ptrGCoh = aec->extended_filter_enabled ? WebRtcAec_kExtendedSmoothingCoefficients[aec->mult - 1] : WebRtcAec_kNormalSmoothingCoefficients[aec->mult - 1]; int i; float sdSum = 0, seSum = 0; for (i = 0; i < PART_LEN1; i++) { aec->sd[i] = ptrGCoh[0] * aec->sd[i] + ptrGCoh[1] * (dfw[0][i] * dfw[0][i] + dfw[1][i] * dfw[1][i]); aec->se[i] = ptrGCoh[0] * aec->se[i] + ptrGCoh[1] * (efw[0][i] * efw[0][i] + efw[1][i] * efw[1][i]); // We threshold here to protect against the ill-effects of a zero farend. // The threshold is not arbitrarily chosen, but balances protection and // adverse interaction with the algorithm's tuning. // TODO(bjornv): investigate further why this is so sensitive. aec->sx[i] = ptrGCoh[0] * aec->sx[i] + ptrGCoh[1] * WEBRTC_SPL_MAX( xfw[0][i] * xfw[0][i] + xfw[1][i] * xfw[1][i], WebRtcAec_kMinFarendPSD); aec->sde[i][0] = ptrGCoh[0] * aec->sde[i][0] + ptrGCoh[1] * (dfw[0][i] * efw[0][i] + dfw[1][i] * efw[1][i]); aec->sde[i][1] = ptrGCoh[0] * aec->sde[i][1] + ptrGCoh[1] * (dfw[0][i] * efw[1][i] - dfw[1][i] * efw[0][i]); aec->sxd[i][0] = ptrGCoh[0] * aec->sxd[i][0] + ptrGCoh[1] * (dfw[0][i] * xfw[0][i] + dfw[1][i] * xfw[1][i]); aec->sxd[i][1] = ptrGCoh[0] * aec->sxd[i][1] + ptrGCoh[1] * (dfw[0][i] * xfw[1][i] - dfw[1][i] * xfw[0][i]); sdSum += aec->sd[i]; seSum += aec->se[i]; } // Divergent filter safeguard update. aec->divergeState = (aec->divergeState ? 1.05f : 1.0f) * seSum > sdSum; // Signal extreme filter divergence if the error is significantly larger // than the nearend (13 dB). *extreme_filter_divergence = (seSum > (19.95f * sdSum)); } // Window time domain data to be used by the fft. __inline static void WindowData(float* x_windowed, const float* x) { int i; for (i = 0; i < PART_LEN; i++) { x_windowed[i] = x[i] * WebRtcAec_sqrtHanning[i]; x_windowed[PART_LEN + i] = x[PART_LEN + i] * WebRtcAec_sqrtHanning[PART_LEN - i]; } } // Puts fft output data into a complex valued array. __inline static void StoreAsComplex(const float* data, float data_complex[2][PART_LEN1]) { int i; data_complex[0][0] = data[0]; data_complex[1][0] = 0; for (i = 1; i < PART_LEN; i++) { data_complex[0][i] = data[2 * i]; data_complex[1][i] = data[2 * i + 1]; } data_complex[0][PART_LEN] = data[1]; data_complex[1][PART_LEN] = 0; } static void SubbandCoherence(AecCore* aec, float efw[2][PART_LEN1], float dfw[2][PART_LEN1], float xfw[2][PART_LEN1], float* fft, float* cohde, float* cohxd, int* extreme_filter_divergence) { int i; SmoothedPSD(aec, efw, dfw, xfw, extreme_filter_divergence); // Subband coherence for (i = 0; i < PART_LEN1; i++) { cohde[i] = (aec->sde[i][0] * aec->sde[i][0] + aec->sde[i][1] * aec->sde[i][1]) / (aec->sd[i] * aec->se[i] + 1e-10f); cohxd[i] = (aec->sxd[i][0] * aec->sxd[i][0] + aec->sxd[i][1] * aec->sxd[i][1]) / (aec->sx[i] * aec->sd[i] + 1e-10f); } } static void GetHighbandGain(const float* lambda, float* nlpGainHband) { int i; *nlpGainHband = (float)0.0; for (i = freqAvgIc; i < PART_LEN1 - 1; i++) { *nlpGainHband += lambda[i]; } *nlpGainHband /= (float)(PART_LEN1 - 1 - freqAvgIc); } static void ComfortNoise(AecCore* aec, float efw[2][PART_LEN1], float comfortNoiseHband[2][PART_LEN1], const float* noisePow, const float* lambda) { int i, num; float rand[PART_LEN]; float noise, noiseAvg, tmp, tmpAvg; int16_t randW16[PART_LEN]; float u[2][PART_LEN1]; const float pi2 = 6.28318530717959f; // Generate a uniform random array on [0 1] WebRtcSpl_RandUArray(randW16, PART_LEN, &aec->seed); for (i = 0; i < PART_LEN; i++) { rand[i] = ((float)randW16[i]) / 32768; } // Reject LF noise u[0][0] = 0; u[1][0] = 0; for (i = 1; i < PART_LEN1; i++) { tmp = pi2 * rand[i - 1]; noise = sqrtf(noisePow[i]); u[0][i] = noise * cosf(tmp); u[1][i] = -noise * sinf(tmp); } u[1][PART_LEN] = 0; for (i = 0; i < PART_LEN1; i++) { // This is the proper weighting to match the background noise power tmp = sqrtf(WEBRTC_SPL_MAX(1 - lambda[i] * lambda[i], 0)); // tmp = 1 - lambda[i]; efw[0][i] += tmp * u[0][i]; efw[1][i] += tmp * u[1][i]; } // For H band comfort noise // TODO: don't compute noise and "tmp" twice. Use the previous results. noiseAvg = 0.0; tmpAvg = 0.0; num = 0; if (aec->num_bands > 1) { // average noise scale // average over second half of freq spectrum (i.e., 4->8khz) // TODO: we shouldn't need num. We know how many elements we're summing. for (i = PART_LEN1 >> 1; i < PART_LEN1; i++) { num++; noiseAvg += sqrtf(noisePow[i]); } noiseAvg /= (float)num; // average nlp scale // average over second half of freq spectrum (i.e., 4->8khz) // TODO: we shouldn't need num. We know how many elements we're summing. num = 0; for (i = PART_LEN1 >> 1; i < PART_LEN1; i++) { num++; tmpAvg += sqrtf(WEBRTC_SPL_MAX(1 - lambda[i] * lambda[i], 0)); } tmpAvg /= (float)num; // Use average noise for H band // TODO: we should probably have a new random vector here. // Reject LF noise u[0][0] = 0; u[1][0] = 0; for (i = 1; i < PART_LEN1; i++) { tmp = pi2 * rand[i - 1]; // Use average noise for H band u[0][i] = noiseAvg * (float)cos(tmp); u[1][i] = -noiseAvg * (float)sin(tmp); } u[1][PART_LEN] = 0; for (i = 0; i < PART_LEN1; i++) { // Use average NLP weight for H band comfortNoiseHband[0][i] = tmpAvg * u[0][i]; comfortNoiseHband[1][i] = tmpAvg * u[1][i]; } } else { memset(comfortNoiseHband, 0, 2 * PART_LEN1 * sizeof(comfortNoiseHband[0][0])); } } static void InitLevel(PowerLevel* level) { const float kBigFloat = 1E17f; level->averagelevel = 0; level->framelevel = 0; level->minlevel = kBigFloat; level->frsum = 0; level->sfrsum = 0; level->frcounter = 0; level->sfrcounter = 0; } static void InitStats(Stats* stats) { stats->instant = kOffsetLevel; stats->average = kOffsetLevel; stats->max = kOffsetLevel; stats->min = kOffsetLevel * (-1); stats->sum = 0; stats->hisum = 0; stats->himean = kOffsetLevel; stats->counter = 0; stats->hicounter = 0; } static void InitMetrics(AecCore* self) { self->stateCounter = 0; InitLevel(&self->farlevel); InitLevel(&self->nearlevel); InitLevel(&self->linoutlevel); InitLevel(&self->nlpoutlevel); InitStats(&self->erl); InitStats(&self->erle); InitStats(&self->aNlp); InitStats(&self->rerl); } static void UpdateLevel(PowerLevel* level, float in[2][PART_LEN1]) { // Do the energy calculation in the frequency domain. The FFT is performed on // a segment of PART_LEN2 samples due to overlap, but we only want the energy // of half that data (the last PART_LEN samples). Parseval's relation states // that the energy is preserved according to // // \sum_{n=0}^{N-1} |x(n)|^2 = 1/N * \sum_{n=0}^{N-1} |X(n)|^2 // = ENERGY, // // where N = PART_LEN2. Since we are only interested in calculating the energy // for the last PART_LEN samples we approximate by calculating ENERGY and // divide by 2, // // \sum_{n=N/2}^{N-1} |x(n)|^2 ~= ENERGY / 2 // // Since we deal with real valued time domain signals we only store frequency // bins [0, PART_LEN], which is what |in| consists of. To calculate ENERGY we // need to add the contribution from the missing part in // [PART_LEN+1, PART_LEN2-1]. These values are, up to a phase shift, identical // with the values in [1, PART_LEN-1], hence multiply those values by 2. This // is the values in the for loop below, but multiplication by 2 and division // by 2 cancel. // TODO(bjornv): Investigate reusing energy calculations performed at other // places in the code. int k = 1; // Imaginary parts are zero at end points and left out of the calculation. float energy = (in[0][0] * in[0][0]) / 2; energy += (in[0][PART_LEN] * in[0][PART_LEN]) / 2; for (k = 1; k < PART_LEN; k++) { energy += (in[0][k] * in[0][k] + in[1][k] * in[1][k]); } energy /= PART_LEN2; level->sfrsum += energy; level->sfrcounter++; if (level->sfrcounter > subCountLen) { level->framelevel = level->sfrsum / (subCountLen * PART_LEN); level->sfrsum = 0; level->sfrcounter = 0; if (level->framelevel > 0) { if (level->framelevel < level->minlevel) { level->minlevel = level->framelevel; // New minimum. } else { level->minlevel *= (1 + 0.001f); // Small increase. } } level->frcounter++; level->frsum += level->framelevel; if (level->frcounter > countLen) { level->averagelevel = level->frsum / countLen; level->frsum = 0; level->frcounter = 0; } } } static void UpdateMetrics(AecCore* aec) { float dtmp, dtmp2; const float actThresholdNoisy = 8.0f; const float actThresholdClean = 40.0f; const float safety = 0.99995f; const float noisyPower = 300000.0f; float actThreshold; float echo, suppressedEcho; if (aec->echoState) { // Check if echo is likely present aec->stateCounter++; } if (aec->farlevel.frcounter == 0) { if (aec->farlevel.minlevel < noisyPower) { actThreshold = actThresholdClean; } else { actThreshold = actThresholdNoisy; } if ((aec->stateCounter > (0.5f * countLen * subCountLen)) && (aec->farlevel.sfrcounter == 0) // Estimate in active far-end segments only && (aec->farlevel.averagelevel > (actThreshold * aec->farlevel.minlevel))) { // Subtract noise power echo = aec->nearlevel.averagelevel - safety * aec->nearlevel.minlevel; // ERL dtmp = 10 * (float)log10(aec->farlevel.averagelevel / aec->nearlevel.averagelevel + 1e-10f); dtmp2 = 10 * (float)log10(aec->farlevel.averagelevel / echo + 1e-10f); aec->erl.instant = dtmp; if (dtmp > aec->erl.max) { aec->erl.max = dtmp; } if (dtmp < aec->erl.min) { aec->erl.min = dtmp; } aec->erl.counter++; aec->erl.sum += dtmp; aec->erl.average = aec->erl.sum / aec->erl.counter; // Upper mean if (dtmp > aec->erl.average) { aec->erl.hicounter++; aec->erl.hisum += dtmp; aec->erl.himean = aec->erl.hisum / aec->erl.hicounter; } // A_NLP dtmp = 10 * (float)log10(aec->nearlevel.averagelevel / (2 * aec->linoutlevel.averagelevel) + 1e-10f); // subtract noise power suppressedEcho = 2 * (aec->linoutlevel.averagelevel - safety * aec->linoutlevel.minlevel); dtmp2 = 10 * (float)log10(echo / suppressedEcho + 1e-10f); aec->aNlp.instant = dtmp2; if (dtmp > aec->aNlp.max) { aec->aNlp.max = dtmp; } if (dtmp < aec->aNlp.min) { aec->aNlp.min = dtmp; } aec->aNlp.counter++; aec->aNlp.sum += dtmp; aec->aNlp.average = aec->aNlp.sum / aec->aNlp.counter; // Upper mean if (dtmp > aec->aNlp.average) { aec->aNlp.hicounter++; aec->aNlp.hisum += dtmp; aec->aNlp.himean = aec->aNlp.hisum / aec->aNlp.hicounter; } // ERLE // subtract noise power suppressedEcho = 2 * (aec->nlpoutlevel.averagelevel - safety * aec->nlpoutlevel.minlevel); dtmp = 10 * (float)log10(aec->nearlevel.averagelevel / (2 * aec->nlpoutlevel.averagelevel) + 1e-10f); dtmp2 = 10 * (float)log10(echo / suppressedEcho + 1e-10f); dtmp = dtmp2; aec->erle.instant = dtmp; if (dtmp > aec->erle.max) { aec->erle.max = dtmp; } if (dtmp < aec->erle.min) { aec->erle.min = dtmp; } aec->erle.counter++; aec->erle.sum += dtmp; aec->erle.average = aec->erle.sum / aec->erle.counter; // Upper mean if (dtmp > aec->erle.average) { aec->erle.hicounter++; aec->erle.hisum += dtmp; aec->erle.himean = aec->erle.hisum / aec->erle.hicounter; } } aec->stateCounter = 0; } } static void UpdateDelayMetrics(AecCore* self) { int i = 0; int delay_values = 0; int median = 0; int lookahead = WebRtc_lookahead(self->delay_estimator); const int kMsPerBlock = PART_LEN / (self->mult * 8); int64_t l1_norm = 0; if (self->num_delay_values == 0) { // We have no new delay value data. Even though -1 is a valid |median| in // the sense that we allow negative values, it will practically never be // used since multiples of |kMsPerBlock| will always be returned. // We therefore use -1 to indicate in the logs that the delay estimator was // not able to estimate the delay. self->delay_median = -1; self->delay_std = -1; self->fraction_poor_delays = -1; return; } // Start value for median count down. delay_values = self->num_delay_values >> 1; // Get median of delay values since last update. for (i = 0; i < kHistorySizeBlocks; i++) { delay_values -= self->delay_histogram[i]; if (delay_values < 0) { median = i; break; } } // Account for lookahead. self->delay_median = (median - lookahead) * kMsPerBlock; // Calculate the L1 norm, with median value as central moment. for (i = 0; i < kHistorySizeBlocks; i++) { l1_norm += abs(i - median) * self->delay_histogram[i]; } self->delay_std = (int)((l1_norm + self->num_delay_values / 2) / self->num_delay_values) * kMsPerBlock; // Determine fraction of delays that are out of bounds, that is, either // negative (anti-causal system) or larger than the AEC filter length. { int num_delays_out_of_bounds = self->num_delay_values; const int histogram_length = sizeof(self->delay_histogram) / sizeof(self->delay_histogram[0]); for (i = lookahead; i < lookahead + self->num_partitions; ++i) { if (i < histogram_length) num_delays_out_of_bounds -= self->delay_histogram[i]; } self->fraction_poor_delays = (float)num_delays_out_of_bounds / self->num_delay_values; } // Reset histogram. memset(self->delay_histogram, 0, sizeof(self->delay_histogram)); self->num_delay_values = 0; return; } static void ScaledInverseFft(float freq_data[2][PART_LEN1], float time_data[PART_LEN2], float scale, int conjugate) { int i; const float normalization = scale / ((float)PART_LEN2); const float sign = (conjugate ? -1 : 1); time_data[0] = freq_data[0][0] * normalization; time_data[1] = freq_data[0][PART_LEN] * normalization; for (i = 1; i < PART_LEN; i++) { time_data[2 * i] = freq_data[0][i] * normalization; time_data[2 * i + 1] = sign * freq_data[1][i] * normalization; } aec_rdft_inverse_128(time_data); } static void Fft(float time_data[PART_LEN2], float freq_data[2][PART_LEN1]) { int i; aec_rdft_forward_128(time_data); // Reorder fft output data. freq_data[1][0] = 0; freq_data[1][PART_LEN] = 0; freq_data[0][0] = time_data[0]; freq_data[0][PART_LEN] = time_data[1]; for (i = 1; i < PART_LEN; i++) { freq_data[0][i] = time_data[2 * i]; freq_data[1][i] = time_data[2 * i + 1]; } } static int SignalBasedDelayCorrection(AecCore* self) { int delay_correction = 0; int last_delay = -2; assert(self != NULL); #if !defined(WEBRTC_ANDROID) // On desktops, turn on correction after |kDelayCorrectionStart| frames. This // is to let the delay estimation get a chance to converge. Also, if the // playout audio volume is low (or even muted) the delay estimation can return // a very large delay, which will break the AEC if it is applied. if (self->frame_count < kDelayCorrectionStart) { return 0; } #endif // 1. Check for non-negative delay estimate. Note that the estimates we get // from the delay estimation are not compensated for lookahead. Hence, a // negative |last_delay| is an invalid one. // 2. Verify that there is a delay change. In addition, only allow a change // if the delay is outside a certain region taking the AEC filter length // into account. // TODO(bjornv): Investigate if we can remove the non-zero delay change check. // 3. Only allow delay correction if the delay estimation quality exceeds // |delay_quality_threshold|. // 4. Finally, verify that the proposed |delay_correction| is feasible by // comparing with the size of the far-end buffer. last_delay = WebRtc_last_delay(self->delay_estimator); if ((last_delay >= 0) && (last_delay != self->previous_delay) && (WebRtc_last_delay_quality(self->delay_estimator) > self->delay_quality_threshold)) { int delay = last_delay - WebRtc_lookahead(self->delay_estimator); // Allow for a slack in the actual delay, defined by a |lower_bound| and an // |upper_bound|. The adaptive echo cancellation filter is currently // |num_partitions| (of 64 samples) long. If the delay estimate is negative // or at least 3/4 of the filter length we open up for correction. const int lower_bound = 0; const int upper_bound = self->num_partitions * 3 / 4; const int do_correction = delay <= lower_bound || delay > upper_bound; if (do_correction == 1) { int available_read = (int)WebRtc_available_read(self->far_time_buf); // With |shift_offset| we gradually rely on the delay estimates. For // positive delays we reduce the correction by |shift_offset| to lower the // risk of pushing the AEC into a non causal state. For negative delays // we rely on the values up to a rounding error, hence compensate by 1 // element to make sure to push the delay into the causal region. delay_correction = -delay; delay_correction += delay > self->shift_offset ? self->shift_offset : 1; self->shift_offset--; self->shift_offset = (self->shift_offset <= 1 ? 1 : self->shift_offset); if (delay_correction > available_read - self->mult - 1) { // There is not enough data in the buffer to perform this shift. Hence, // we do not rely on the delay estimate and do nothing. delay_correction = 0; } else { self->previous_delay = last_delay; ++self->delay_correction_count; } } } // Update the |delay_quality_threshold| once we have our first delay // correction. if (self->delay_correction_count > 0) { float delay_quality = WebRtc_last_delay_quality(self->delay_estimator); delay_quality = (delay_quality > kDelayQualityThresholdMax ? kDelayQualityThresholdMax : delay_quality); self->delay_quality_threshold = (delay_quality > self->delay_quality_threshold ? delay_quality : self->delay_quality_threshold); } return delay_correction; } static void EchoSubtraction( AecCore* aec, int num_partitions, int x_fft_buf_block_pos, int metrics_mode, int extended_filter_enabled, float normal_mu, float normal_error_threshold, float x_fft_buf[2][kExtendedNumPartitions * PART_LEN1], float* const y, float x_pow[PART_LEN1], float h_fft_buf[2][kExtendedNumPartitions * PART_LEN1], PowerLevel* linout_level, float echo_subtractor_output[PART_LEN]) { float s_fft[2][PART_LEN1]; float e_extended[PART_LEN2]; float s_extended[PART_LEN2]; float *s; float e[PART_LEN]; float e_fft[2][PART_LEN1]; int i; memset(s_fft, 0, sizeof(s_fft)); // Conditionally reset the echo subtraction filter if the filter has diverged // significantly. if (!aec->extended_filter_enabled && aec->extreme_filter_divergence) { memset(aec->wfBuf, 0, sizeof(aec->wfBuf)); aec->extreme_filter_divergence = 0; } // Produce echo estimate s_fft. WebRtcAec_FilterFar(num_partitions, x_fft_buf_block_pos, x_fft_buf, h_fft_buf, s_fft); // Compute the time-domain echo estimate s. ScaledInverseFft(s_fft, s_extended, 2.0f, 0); s = &s_extended[PART_LEN]; // Compute the time-domain echo prediction error. for (i = 0; i < PART_LEN; ++i) { e[i] = y[i] - s[i]; } // Compute the frequency domain echo prediction error. memset(e_extended, 0, sizeof(float) * PART_LEN); memcpy(e_extended + PART_LEN, e, sizeof(float) * PART_LEN); Fft(e_extended, e_fft); RTC_AEC_DEBUG_RAW_WRITE(aec->e_fft_file, &e_fft[0][0], sizeof(e_fft[0][0]) * PART_LEN1 * 2); if (metrics_mode == 1) { // Note that the first PART_LEN samples in fft (before transformation) are // zero. Hence, the scaling by two in UpdateLevel() should not be // performed. That scaling is taken care of in UpdateMetrics() instead. UpdateLevel(linout_level, e_fft); } // Scale error signal inversely with far power. WebRtcAec_ScaleErrorSignal(extended_filter_enabled, normal_mu, normal_error_threshold, x_pow, e_fft); WebRtcAec_FilterAdaptation(num_partitions, x_fft_buf_block_pos, x_fft_buf, e_fft, h_fft_buf); memcpy(echo_subtractor_output, e, sizeof(float) * PART_LEN); } static void EchoSuppression(AecCore* aec, float farend[PART_LEN2], float* echo_subtractor_output, float* output, float* const* outputH) { float efw[2][PART_LEN1]; float xfw[2][PART_LEN1]; float dfw[2][PART_LEN1]; float comfortNoiseHband[2][PART_LEN1]; float fft[PART_LEN2]; float nlpGainHband; int i; size_t j; // Coherence and non-linear filter float cohde[PART_LEN1], cohxd[PART_LEN1]; float hNlDeAvg, hNlXdAvg; float hNl[PART_LEN1]; float hNlPref[kPrefBandSize]; float hNlFb = 0, hNlFbLow = 0; const float prefBandQuant = 0.75f, prefBandQuantLow = 0.5f; const int prefBandSize = kPrefBandSize / aec->mult; const int minPrefBand = 4 / aec->mult; // Power estimate smoothing coefficients. const float* min_overdrive = aec->extended_filter_enabled ? kExtendedMinOverDrive : kNormalMinOverDrive; // Filter energy const int delayEstInterval = 10 * aec->mult; float* xfw_ptr = NULL; // Update eBuf with echo subtractor output. memcpy(aec->eBuf + PART_LEN, echo_subtractor_output, sizeof(float) * PART_LEN); // Analysis filter banks for the echo suppressor. // Windowed near-end ffts. WindowData(fft, aec->dBuf); aec_rdft_forward_128(fft); StoreAsComplex(fft, dfw); // Windowed echo suppressor output ffts. WindowData(fft, aec->eBuf); aec_rdft_forward_128(fft); StoreAsComplex(fft, efw); // NLP // Convert far-end partition to the frequency domain with windowing. WindowData(fft, farend); Fft(fft, xfw); xfw_ptr = &xfw[0][0]; // Buffer far. memcpy(aec->xfwBuf, xfw_ptr, sizeof(float) * 2 * PART_LEN1); aec->delayEstCtr++; if (aec->delayEstCtr == delayEstInterval) { aec->delayEstCtr = 0; aec->delayIdx = WebRtcAec_PartitionDelay(aec); } // Use delayed far. memcpy(xfw, aec->xfwBuf + aec->delayIdx * PART_LEN1, sizeof(xfw[0][0]) * 2 * PART_LEN1); WebRtcAec_SubbandCoherence(aec, efw, dfw, xfw, fft, cohde, cohxd, &aec->extreme_filter_divergence); // Select the microphone signal as output if the filter is deemed to have // diverged. if (aec->divergeState) { memcpy(efw, dfw, sizeof(efw[0][0]) * 2 * PART_LEN1); } hNlXdAvg = 0; for (i = minPrefBand; i < prefBandSize + minPrefBand; i++) { hNlXdAvg += cohxd[i]; } hNlXdAvg /= prefBandSize; hNlXdAvg = 1 - hNlXdAvg; hNlDeAvg = 0; for (i = minPrefBand; i < prefBandSize + minPrefBand; i++) { hNlDeAvg += cohde[i]; } hNlDeAvg /= prefBandSize; if (hNlXdAvg < 0.75f && hNlXdAvg < aec->hNlXdAvgMin) { aec->hNlXdAvgMin = hNlXdAvg; } if (hNlDeAvg > 0.98f && hNlXdAvg > 0.9f) { aec->stNearState = 1; } else if (hNlDeAvg < 0.95f || hNlXdAvg < 0.8f) { aec->stNearState = 0; } if (aec->hNlXdAvgMin == 1) { aec->echoState = 0; aec->overDrive = min_overdrive[aec->nlp_mode]; if (aec->stNearState == 1) { memcpy(hNl, cohde, sizeof(hNl)); hNlFb = hNlDeAvg; hNlFbLow = hNlDeAvg; } else { for (i = 0; i < PART_LEN1; i++) { hNl[i] = 1 - cohxd[i]; } hNlFb = hNlXdAvg; hNlFbLow = hNlXdAvg; } } else { if (aec->stNearState == 1) { aec->echoState = 0; memcpy(hNl, cohde, sizeof(hNl)); hNlFb = hNlDeAvg; hNlFbLow = hNlDeAvg; } else { aec->echoState = 1; for (i = 0; i < PART_LEN1; i++) { hNl[i] = WEBRTC_SPL_MIN(cohde[i], 1 - cohxd[i]); } // Select an order statistic from the preferred bands. // TODO: Using quicksort now, but a selection algorithm may be preferred. memcpy(hNlPref, &hNl[minPrefBand], sizeof(float) * prefBandSize); qsort(hNlPref, prefBandSize, sizeof(float), CmpFloat); hNlFb = hNlPref[(int)floor(prefBandQuant * (prefBandSize - 1))]; hNlFbLow = hNlPref[(int)floor(prefBandQuantLow * (prefBandSize - 1))]; } } // Track the local filter minimum to determine suppression overdrive. if (hNlFbLow < 0.6f && hNlFbLow < aec->hNlFbLocalMin) { aec->hNlFbLocalMin = hNlFbLow; aec->hNlFbMin = hNlFbLow; aec->hNlNewMin = 1; aec->hNlMinCtr = 0; } aec->hNlFbLocalMin = WEBRTC_SPL_MIN(aec->hNlFbLocalMin + 0.0008f / aec->mult, 1); aec->hNlXdAvgMin = WEBRTC_SPL_MIN(aec->hNlXdAvgMin + 0.0006f / aec->mult, 1); if (aec->hNlNewMin == 1) { aec->hNlMinCtr++; } if (aec->hNlMinCtr == 2) { aec->hNlNewMin = 0; aec->hNlMinCtr = 0; aec->overDrive = WEBRTC_SPL_MAX(kTargetSupp[aec->nlp_mode] / ((float)log(aec->hNlFbMin + 1e-10f) + 1e-10f), min_overdrive[aec->nlp_mode]); } // Smooth the overdrive. if (aec->overDrive < aec->overDriveSm) { aec->overDriveSm = 0.99f * aec->overDriveSm + 0.01f * aec->overDrive; } else { aec->overDriveSm = 0.9f * aec->overDriveSm + 0.1f * aec->overDrive; } WebRtcAec_OverdriveAndSuppress(aec, hNl, hNlFb, efw); // Add comfort noise. WebRtcAec_ComfortNoise(aec, efw, comfortNoiseHband, aec->noisePow, hNl); // TODO(bjornv): Investigate how to take the windowing below into account if // needed. if (aec->metricsMode == 1) { // Note that we have a scaling by two in the time domain |eBuf|. // In addition the time domain signal is windowed before transformation, // losing half the energy on the average. We take care of the first // scaling only in UpdateMetrics(). UpdateLevel(&aec->nlpoutlevel, efw); } // Inverse error fft. ScaledInverseFft(efw, fft, 2.0f, 1); // Overlap and add to obtain output. for (i = 0; i < PART_LEN; i++) { output[i] = (fft[i] * WebRtcAec_sqrtHanning[i] + aec->outBuf[i] * WebRtcAec_sqrtHanning[PART_LEN - i]); // Saturate output to keep it in the allowed range. output[i] = WEBRTC_SPL_SAT( WEBRTC_SPL_WORD16_MAX, output[i], WEBRTC_SPL_WORD16_MIN); } memcpy(aec->outBuf, &fft[PART_LEN], PART_LEN * sizeof(aec->outBuf[0])); // For H band if (aec->num_bands > 1) { // H band gain // average nlp over low band: average over second half of freq spectrum // (4->8khz) GetHighbandGain(hNl, &nlpGainHband); // Inverse comfort_noise ScaledInverseFft(comfortNoiseHband, fft, 2.0f, 0); // compute gain factor for (j = 0; j < aec->num_bands - 1; ++j) { for (i = 0; i < PART_LEN; i++) { outputH[j][i] = aec->dBufH[j][i] * nlpGainHband; } } // Add some comfort noise where Hband is attenuated. for (i = 0; i < PART_LEN; i++) { outputH[0][i] += cnScaleHband * fft[i]; } // Saturate output to keep it in the allowed range. for (j = 0; j < aec->num_bands - 1; ++j) { for (i = 0; i < PART_LEN; i++) { outputH[j][i] = WEBRTC_SPL_SAT( WEBRTC_SPL_WORD16_MAX, outputH[j][i], WEBRTC_SPL_WORD16_MIN); } } } // Copy the current block to the old position. memcpy(aec->dBuf, aec->dBuf + PART_LEN, sizeof(float) * PART_LEN); memcpy(aec->eBuf, aec->eBuf + PART_LEN, sizeof(float) * PART_LEN); // Copy the current block to the old position for H band for (j = 0; j < aec->num_bands - 1; ++j) { memcpy(aec->dBufH[j], aec->dBufH[j] + PART_LEN, sizeof(float) * PART_LEN); } memmove(aec->xfwBuf + PART_LEN1, aec->xfwBuf, sizeof(aec->xfwBuf) - sizeof(complex_t) * PART_LEN1); } static void ProcessBlock(AecCore* aec) { size_t i; float fft[PART_LEN2]; float xf[2][PART_LEN1]; float df[2][PART_LEN1]; float far_spectrum = 0.0f; float near_spectrum = 0.0f; float abs_far_spectrum[PART_LEN1]; float abs_near_spectrum[PART_LEN1]; const float gPow[2] = {0.9f, 0.1f}; // Noise estimate constants. const int noiseInitBlocks = 500 * aec->mult; const float step = 0.1f; const float ramp = 1.0002f; const float gInitNoise[2] = {0.999f, 0.001f}; float nearend[PART_LEN]; float* nearend_ptr = NULL; float farend[PART_LEN2]; float* farend_ptr = NULL; float echo_subtractor_output[PART_LEN]; float output[PART_LEN]; float outputH[NUM_HIGH_BANDS_MAX][PART_LEN]; float* outputH_ptr[NUM_HIGH_BANDS_MAX]; float* xf_ptr = NULL; for (i = 0; i < NUM_HIGH_BANDS_MAX; ++i) { outputH_ptr[i] = outputH[i]; } // Concatenate old and new nearend blocks. for (i = 0; i < aec->num_bands - 1; ++i) { WebRtc_ReadBuffer(aec->nearFrBufH[i], (void**)&nearend_ptr, nearend, PART_LEN); memcpy(aec->dBufH[i] + PART_LEN, nearend_ptr, sizeof(nearend)); } WebRtc_ReadBuffer(aec->nearFrBuf, (void**)&nearend_ptr, nearend, PART_LEN); memcpy(aec->dBuf + PART_LEN, nearend_ptr, sizeof(nearend)); // We should always have at least one element stored in |far_buf|. assert(WebRtc_available_read(aec->far_time_buf) > 0); WebRtc_ReadBuffer(aec->far_time_buf, (void**)&farend_ptr, farend, 1); #ifdef WEBRTC_AEC_DEBUG_DUMP { // TODO(minyue): |farend_ptr| starts from buffered samples. This will be // modified when |aec->far_time_buf| is revised. RTC_AEC_DEBUG_WAV_WRITE(aec->farFile, &farend_ptr[PART_LEN], PART_LEN); RTC_AEC_DEBUG_WAV_WRITE(aec->nearFile, nearend_ptr, PART_LEN); } #endif // Convert far-end signal to the frequency domain. memcpy(fft, farend_ptr, sizeof(float) * PART_LEN2); Fft(fft, xf); xf_ptr = &xf[0][0]; // Near fft memcpy(fft, aec->dBuf, sizeof(float) * PART_LEN2); Fft(fft, df); // Power smoothing for (i = 0; i < PART_LEN1; i++) { far_spectrum = (xf_ptr[i] * xf_ptr[i]) + (xf_ptr[PART_LEN1 + i] * xf_ptr[PART_LEN1 + i]); aec->xPow[i] = gPow[0] * aec->xPow[i] + gPow[1] * aec->num_partitions * far_spectrum; // Calculate absolute spectra abs_far_spectrum[i] = sqrtf(far_spectrum); near_spectrum = df[0][i] * df[0][i] + df[1][i] * df[1][i]; aec->dPow[i] = gPow[0] * aec->dPow[i] + gPow[1] * near_spectrum; // Calculate absolute spectra abs_near_spectrum[i] = sqrtf(near_spectrum); } // Estimate noise power. Wait until dPow is more stable. if (aec->noiseEstCtr > 50) { for (i = 0; i < PART_LEN1; i++) { if (aec->dPow[i] < aec->dMinPow[i]) { aec->dMinPow[i] = (aec->dPow[i] + step * (aec->dMinPow[i] - aec->dPow[i])) * ramp; } else { aec->dMinPow[i] *= ramp; } } } // Smooth increasing noise power from zero at the start, // to avoid a sudden burst of comfort noise. if (aec->noiseEstCtr < noiseInitBlocks) { aec->noiseEstCtr++; for (i = 0; i < PART_LEN1; i++) { if (aec->dMinPow[i] > aec->dInitMinPow[i]) { aec->dInitMinPow[i] = gInitNoise[0] * aec->dInitMinPow[i] + gInitNoise[1] * aec->dMinPow[i]; } else { aec->dInitMinPow[i] = aec->dMinPow[i]; } } aec->noisePow = aec->dInitMinPow; } else { aec->noisePow = aec->dMinPow; } // Block wise delay estimation used for logging if (aec->delay_logging_enabled) { if (WebRtc_AddFarSpectrumFloat( aec->delay_estimator_farend, abs_far_spectrum, PART_LEN1) == 0) { int delay_estimate = WebRtc_DelayEstimatorProcessFloat( aec->delay_estimator, abs_near_spectrum, PART_LEN1); if (delay_estimate >= 0) { // Update delay estimate buffer. aec->delay_histogram[delay_estimate]++; aec->num_delay_values++; } if (aec->delay_metrics_delivered == 1 && aec->num_delay_values >= kDelayMetricsAggregationWindow) { UpdateDelayMetrics(aec); } } } // Update the xfBuf block position. aec->xfBufBlockPos--; if (aec->xfBufBlockPos == -1) { aec->xfBufBlockPos = aec->num_partitions - 1; } // Buffer xf memcpy(aec->xfBuf[0] + aec->xfBufBlockPos * PART_LEN1, xf_ptr, sizeof(float) * PART_LEN1); memcpy(aec->xfBuf[1] + aec->xfBufBlockPos * PART_LEN1, &xf_ptr[PART_LEN1], sizeof(float) * PART_LEN1); // Perform echo subtraction. EchoSubtraction(aec, aec->num_partitions, aec->xfBufBlockPos, aec->metricsMode, aec->extended_filter_enabled, aec->normal_mu, aec->normal_error_threshold, aec->xfBuf, nearend_ptr, aec->xPow, aec->wfBuf, &aec->linoutlevel, echo_subtractor_output); RTC_AEC_DEBUG_WAV_WRITE(aec->outLinearFile, echo_subtractor_output, PART_LEN); // Perform echo suppression. EchoSuppression(aec, farend_ptr, echo_subtractor_output, output, outputH_ptr); if (aec->metricsMode == 1) { // Update power levels and echo metrics UpdateLevel(&aec->farlevel, (float(*)[PART_LEN1])xf_ptr); UpdateLevel(&aec->nearlevel, df); UpdateMetrics(aec); } // Store the output block. WebRtc_WriteBuffer(aec->outFrBuf, output, PART_LEN); // For high bands for (i = 0; i < aec->num_bands - 1; ++i) { WebRtc_WriteBuffer(aec->outFrBufH[i], outputH[i], PART_LEN); } RTC_AEC_DEBUG_WAV_WRITE(aec->outFile, output, PART_LEN); } AecCore* WebRtcAec_CreateAec() { int i; AecCore* aec = malloc(sizeof(AecCore)); if (!aec) { return NULL; } aec->nearFrBuf = WebRtc_CreateBuffer(FRAME_LEN + PART_LEN, sizeof(float)); if (!aec->nearFrBuf) { WebRtcAec_FreeAec(aec); return NULL; } aec->outFrBuf = WebRtc_CreateBuffer(FRAME_LEN + PART_LEN, sizeof(float)); if (!aec->outFrBuf) { WebRtcAec_FreeAec(aec); return NULL; } for (i = 0; i < NUM_HIGH_BANDS_MAX; ++i) { aec->nearFrBufH[i] = WebRtc_CreateBuffer(FRAME_LEN + PART_LEN, sizeof(float)); if (!aec->nearFrBufH[i]) { WebRtcAec_FreeAec(aec); return NULL; } aec->outFrBufH[i] = WebRtc_CreateBuffer(FRAME_LEN + PART_LEN, sizeof(float)); if (!aec->outFrBufH[i]) { WebRtcAec_FreeAec(aec); return NULL; } } // Create far-end buffers. // For bit exactness with legacy code, each element in |far_time_buf| is // supposed to contain |PART_LEN2| samples with an overlap of |PART_LEN| // samples from the last frame. // TODO(minyue): reduce |far_time_buf| to non-overlapped |PART_LEN| samples. aec->far_time_buf = WebRtc_CreateBuffer(kBufSizePartitions, sizeof(float) * PART_LEN2); if (!aec->far_time_buf) { WebRtcAec_FreeAec(aec); return NULL; } #ifdef WEBRTC_AEC_DEBUG_DUMP aec->instance_index = webrtc_aec_instance_count; aec->farFile = aec->nearFile = aec->outFile = aec->outLinearFile = NULL; aec->debug_dump_count = 0; #endif aec->delay_estimator_farend = WebRtc_CreateDelayEstimatorFarend(PART_LEN1, kHistorySizeBlocks); if (aec->delay_estimator_farend == NULL) { WebRtcAec_FreeAec(aec); return NULL; } // We create the delay_estimator with the same amount of maximum lookahead as // the delay history size (kHistorySizeBlocks) for symmetry reasons. aec->delay_estimator = WebRtc_CreateDelayEstimator( aec->delay_estimator_farend, kHistorySizeBlocks); if (aec->delay_estimator == NULL) { WebRtcAec_FreeAec(aec); return NULL; } #ifdef WEBRTC_ANDROID aec->delay_agnostic_enabled = 1; // DA-AEC enabled by default. // DA-AEC assumes the system is causal from the beginning and will self adjust // the lookahead when shifting is required. WebRtc_set_lookahead(aec->delay_estimator, 0); #else aec->delay_agnostic_enabled = 0; WebRtc_set_lookahead(aec->delay_estimator, kLookaheadBlocks); #endif aec->extended_filter_enabled = 0; // Assembly optimization WebRtcAec_FilterFar = FilterFar; WebRtcAec_ScaleErrorSignal = ScaleErrorSignal; WebRtcAec_FilterAdaptation = FilterAdaptation; WebRtcAec_OverdriveAndSuppress = OverdriveAndSuppress; WebRtcAec_ComfortNoise = ComfortNoise; WebRtcAec_SubbandCoherence = SubbandCoherence; WebRtcAec_StoreAsComplex = StoreAsComplex; WebRtcAec_PartitionDelay = PartitionDelay; WebRtcAec_WindowData = WindowData; #if defined(WEBRTC_ARCH_X86_FAMILY) if (WebRtc_GetCPUInfo(kSSE2)) { WebRtcAec_InitAec_SSE2(); } #endif #if defined(MIPS_FPU_LE) WebRtcAec_InitAec_mips(); #endif #if defined(WEBRTC_HAS_NEON) WebRtcAec_InitAec_neon(); #elif defined(WEBRTC_DETECT_NEON) if ((WebRtc_GetCPUFeaturesARM() & kCPUFeatureNEON) != 0) { WebRtcAec_InitAec_neon(); } #endif aec_rdft_init(); return aec; } void WebRtcAec_FreeAec(AecCore* aec) { int i; if (aec == NULL) { return; } WebRtc_FreeBuffer(aec->nearFrBuf); WebRtc_FreeBuffer(aec->outFrBuf); for (i = 0; i < NUM_HIGH_BANDS_MAX; ++i) { WebRtc_FreeBuffer(aec->nearFrBufH[i]); WebRtc_FreeBuffer(aec->outFrBufH[i]); } WebRtc_FreeBuffer(aec->far_time_buf); RTC_AEC_DEBUG_WAV_CLOSE(aec->farFile); RTC_AEC_DEBUG_WAV_CLOSE(aec->nearFile); RTC_AEC_DEBUG_WAV_CLOSE(aec->outFile); RTC_AEC_DEBUG_WAV_CLOSE(aec->outLinearFile); RTC_AEC_DEBUG_RAW_CLOSE(aec->e_fft_file); WebRtc_FreeDelayEstimator(aec->delay_estimator); WebRtc_FreeDelayEstimatorFarend(aec->delay_estimator_farend); free(aec); } int WebRtcAec_InitAec(AecCore* aec, int sampFreq) { int i; aec->sampFreq = sampFreq; if (sampFreq == 8000) { aec->normal_mu = 0.6f; aec->normal_error_threshold = 2e-6f; aec->num_bands = 1; } else { aec->normal_mu = 0.5f; aec->normal_error_threshold = 1.5e-6f; aec->num_bands = (size_t)(sampFreq / 16000); } WebRtc_InitBuffer(aec->nearFrBuf); WebRtc_InitBuffer(aec->outFrBuf); for (i = 0; i < NUM_HIGH_BANDS_MAX; ++i) { WebRtc_InitBuffer(aec->nearFrBufH[i]); WebRtc_InitBuffer(aec->outFrBufH[i]); } // Initialize far-end buffers. WebRtc_InitBuffer(aec->far_time_buf); #ifdef WEBRTC_AEC_DEBUG_DUMP { int process_rate = sampFreq > 16000 ? 16000 : sampFreq; RTC_AEC_DEBUG_WAV_REOPEN("aec_far", aec->instance_index, aec->debug_dump_count, process_rate, &aec->farFile ); RTC_AEC_DEBUG_WAV_REOPEN("aec_near", aec->instance_index, aec->debug_dump_count, process_rate, &aec->nearFile); RTC_AEC_DEBUG_WAV_REOPEN("aec_out", aec->instance_index, aec->debug_dump_count, process_rate, &aec->outFile ); RTC_AEC_DEBUG_WAV_REOPEN("aec_out_linear", aec->instance_index, aec->debug_dump_count, process_rate, &aec->outLinearFile); } RTC_AEC_DEBUG_RAW_OPEN("aec_e_fft", aec->debug_dump_count, &aec->e_fft_file); ++aec->debug_dump_count; #endif aec->system_delay = 0; if (WebRtc_InitDelayEstimatorFarend(aec->delay_estimator_farend) != 0) { return -1; } if (WebRtc_InitDelayEstimator(aec->delay_estimator) != 0) { return -1; } aec->delay_logging_enabled = 0; aec->delay_metrics_delivered = 0; memset(aec->delay_histogram, 0, sizeof(aec->delay_histogram)); aec->num_delay_values = 0; aec->delay_median = -1; aec->delay_std = -1; aec->fraction_poor_delays = -1.0f; aec->signal_delay_correction = 0; aec->previous_delay = -2; // (-2): Uninitialized. aec->delay_correction_count = 0; aec->shift_offset = kInitialShiftOffset; aec->delay_quality_threshold = kDelayQualityThresholdMin; aec->num_partitions = kNormalNumPartitions; // Update the delay estimator with filter length. We use half the // |num_partitions| to take the echo path into account. In practice we say // that the echo has a duration of maximum half |num_partitions|, which is not // true, but serves as a crude measure. WebRtc_set_allowed_offset(aec->delay_estimator, aec->num_partitions / 2); // TODO(bjornv): I currently hard coded the enable. Once we've established // that AECM has no performance regression, robust_validation will be enabled // all the time and the APIs to turn it on/off will be removed. Hence, remove // this line then. WebRtc_enable_robust_validation(aec->delay_estimator, 1); aec->frame_count = 0; // Default target suppression mode. aec->nlp_mode = 1; // Sampling frequency multiplier w.r.t. 8 kHz. // In case of multiple bands we process the lower band in 16 kHz, hence the // multiplier is always 2. if (aec->num_bands > 1) { aec->mult = 2; } else { aec->mult = (short)aec->sampFreq / 8000; } aec->farBufWritePos = 0; aec->farBufReadPos = 0; aec->inSamples = 0; aec->outSamples = 0; aec->knownDelay = 0; // Initialize buffers memset(aec->dBuf, 0, sizeof(aec->dBuf)); memset(aec->eBuf, 0, sizeof(aec->eBuf)); // For H bands for (i = 0; i < NUM_HIGH_BANDS_MAX; ++i) { memset(aec->dBufH[i], 0, sizeof(aec->dBufH[i])); } memset(aec->xPow, 0, sizeof(aec->xPow)); memset(aec->dPow, 0, sizeof(aec->dPow)); memset(aec->dInitMinPow, 0, sizeof(aec->dInitMinPow)); aec->noisePow = aec->dInitMinPow; aec->noiseEstCtr = 0; // Initial comfort noise power for (i = 0; i < PART_LEN1; i++) { aec->dMinPow[i] = 1.0e6f; } // Holds the last block written to aec->xfBufBlockPos = 0; // TODO: Investigate need for these initializations. Deleting them doesn't // change the output at all and yields 0.4% overall speedup. memset(aec->xfBuf, 0, sizeof(complex_t) * kExtendedNumPartitions * PART_LEN1); memset(aec->wfBuf, 0, sizeof(complex_t) * kExtendedNumPartitions * PART_LEN1); memset(aec->sde, 0, sizeof(complex_t) * PART_LEN1); memset(aec->sxd, 0, sizeof(complex_t) * PART_LEN1); memset( aec->xfwBuf, 0, sizeof(complex_t) * kExtendedNumPartitions * PART_LEN1); memset(aec->se, 0, sizeof(float) * PART_LEN1); // To prevent numerical instability in the first block. for (i = 0; i < PART_LEN1; i++) { aec->sd[i] = 1; } for (i = 0; i < PART_LEN1; i++) { aec->sx[i] = 1; } memset(aec->hNs, 0, sizeof(aec->hNs)); memset(aec->outBuf, 0, sizeof(float) * PART_LEN); aec->hNlFbMin = 1; aec->hNlFbLocalMin = 1; aec->hNlXdAvgMin = 1; aec->hNlNewMin = 0; aec->hNlMinCtr = 0; aec->overDrive = 2; aec->overDriveSm = 2; aec->delayIdx = 0; aec->stNearState = 0; aec->echoState = 0; aec->divergeState = 0; aec->seed = 777; aec->delayEstCtr = 0; aec->extreme_filter_divergence = 0; // Metrics disabled by default aec->metricsMode = 0; InitMetrics(aec); return 0; } // For bit exactness with a legacy code, |farend| is supposed to contain // |PART_LEN2| samples with an overlap of |PART_LEN| samples from the last // frame. // TODO(minyue): reduce |farend| to non-overlapped |PART_LEN| samples. void WebRtcAec_BufferFarendPartition(AecCore* aec, const float* farend) { // Check if the buffer is full, and in that case flush the oldest data. if (WebRtc_available_write(aec->far_time_buf) < 1) { WebRtcAec_MoveFarReadPtr(aec, 1); } WebRtc_WriteBuffer(aec->far_time_buf, farend, 1); } int WebRtcAec_MoveFarReadPtr(AecCore* aec, int elements) { int elements_moved = WebRtc_MoveReadPtr(aec->far_time_buf, elements); aec->system_delay -= elements_moved * PART_LEN; return elements_moved; } void WebRtcAec_ProcessFrames(AecCore* aec, const float* const* nearend, size_t num_bands, size_t num_samples, int knownDelay, float* const* out) { size_t i, j; int out_elements = 0; aec->frame_count++; // For each frame the process is as follows: // 1) If the system_delay indicates on being too small for processing a // frame we stuff the buffer with enough data for 10 ms. // 2 a) Adjust the buffer to the system delay, by moving the read pointer. // b) Apply signal based delay correction, if we have detected poor AEC // performance. // 3) TODO(bjornv): Investigate if we need to add this: // If we can't move read pointer due to buffer size limitations we // flush/stuff the buffer. // 4) Process as many partitions as possible. // 5) Update the |system_delay| with respect to a full frame of FRAME_LEN // samples. Even though we will have data left to process (we work with // partitions) we consider updating a whole frame, since that's the // amount of data we input and output in audio_processing. // 6) Update the outputs. // The AEC has two different delay estimation algorithms built in. The // first relies on delay input values from the user and the amount of // shifted buffer elements is controlled by |knownDelay|. This delay will // give a guess on how much we need to shift far-end buffers to align with // the near-end signal. The other delay estimation algorithm uses the // far- and near-end signals to find the offset between them. This one // (called "signal delay") is then used to fine tune the alignment, or // simply compensate for errors in the system based one. // Note that the two algorithms operate independently. Currently, we only // allow one algorithm to be turned on. assert(aec->num_bands == num_bands); for (j = 0; j < num_samples; j+= FRAME_LEN) { // TODO(bjornv): Change the near-end buffer handling to be the same as for // far-end, that is, with a near_pre_buf. // Buffer the near-end frame. WebRtc_WriteBuffer(aec->nearFrBuf, &nearend[0][j], FRAME_LEN); // For H band for (i = 1; i < num_bands; ++i) { WebRtc_WriteBuffer(aec->nearFrBufH[i - 1], &nearend[i][j], FRAME_LEN); } // 1) At most we process |aec->mult|+1 partitions in 10 ms. Make sure we // have enough far-end data for that by stuffing the buffer if the // |system_delay| indicates others. if (aec->system_delay < FRAME_LEN) { // We don't have enough data so we rewind 10 ms. WebRtcAec_MoveFarReadPtr(aec, -(aec->mult + 1)); } if (!aec->delay_agnostic_enabled) { // 2 a) Compensate for a possible change in the system delay. // TODO(bjornv): Investigate how we should round the delay difference; // right now we know that incoming |knownDelay| is underestimated when // it's less than |aec->knownDelay|. We therefore, round (-32) in that // direction. In the other direction, we don't have this situation, but // might flush one partition too little. This can cause non-causality, // which should be investigated. Maybe, allow for a non-symmetric // rounding, like -16. int move_elements = (aec->knownDelay - knownDelay - 32) / PART_LEN; int moved_elements = WebRtc_MoveReadPtr(aec->far_time_buf, move_elements); aec->knownDelay -= moved_elements * PART_LEN; } else { // 2 b) Apply signal based delay correction. int move_elements = SignalBasedDelayCorrection(aec); int moved_elements = WebRtc_MoveReadPtr(aec->far_time_buf, move_elements); int far_near_buffer_diff = WebRtc_available_read(aec->far_time_buf) - WebRtc_available_read(aec->nearFrBuf) / PART_LEN; WebRtc_SoftResetDelayEstimator(aec->delay_estimator, moved_elements); WebRtc_SoftResetDelayEstimatorFarend(aec->delay_estimator_farend, moved_elements); aec->signal_delay_correction += moved_elements; // If we rely on reported system delay values only, a buffer underrun here // can never occur since we've taken care of that in 1) above. Here, we // apply signal based delay correction and can therefore end up with // buffer underruns since the delay estimation can be wrong. We therefore // stuff the buffer with enough elements if needed. if (far_near_buffer_diff < 0) { WebRtcAec_MoveFarReadPtr(aec, far_near_buffer_diff); } } // 4) Process as many blocks as possible. while (WebRtc_available_read(aec->nearFrBuf) >= PART_LEN) { ProcessBlock(aec); } // 5) Update system delay with respect to the entire frame. aec->system_delay -= FRAME_LEN; // 6) Update output frame. // Stuff the out buffer if we have less than a frame to output. // This should only happen for the first frame. out_elements = (int)WebRtc_available_read(aec->outFrBuf); if (out_elements < FRAME_LEN) { WebRtc_MoveReadPtr(aec->outFrBuf, out_elements - FRAME_LEN); for (i = 0; i < num_bands - 1; ++i) { WebRtc_MoveReadPtr(aec->outFrBufH[i], out_elements - FRAME_LEN); } } // Obtain an output frame. WebRtc_ReadBuffer(aec->outFrBuf, NULL, &out[0][j], FRAME_LEN); // For H bands. for (i = 1; i < num_bands; ++i) { WebRtc_ReadBuffer(aec->outFrBufH[i - 1], NULL, &out[i][j], FRAME_LEN); } } } int WebRtcAec_GetDelayMetricsCore(AecCore* self, int* median, int* std, float* fraction_poor_delays) { assert(self != NULL); assert(median != NULL); assert(std != NULL); if (self->delay_logging_enabled == 0) { // Logging disabled. return -1; } if (self->delay_metrics_delivered == 0) { UpdateDelayMetrics(self); self->delay_metrics_delivered = 1; } *median = self->delay_median; *std = self->delay_std; *fraction_poor_delays = self->fraction_poor_delays; return 0; } int WebRtcAec_echo_state(AecCore* self) { return self->echoState; } void WebRtcAec_GetEchoStats(AecCore* self, Stats* erl, Stats* erle, Stats* a_nlp) { assert(erl != NULL); assert(erle != NULL); assert(a_nlp != NULL); *erl = self->erl; *erle = self->erle; *a_nlp = self->aNlp; } void WebRtcAec_SetConfigCore(AecCore* self, int nlp_mode, int metrics_mode, int delay_logging) { assert(nlp_mode >= 0 && nlp_mode < 3); self->nlp_mode = nlp_mode; self->metricsMode = metrics_mode; if (self->metricsMode) { InitMetrics(self); } // Turn on delay logging if it is either set explicitly or if delay agnostic // AEC is enabled (which requires delay estimates). self->delay_logging_enabled = delay_logging || self->delay_agnostic_enabled; if (self->delay_logging_enabled) { memset(self->delay_histogram, 0, sizeof(self->delay_histogram)); } } void WebRtcAec_enable_delay_agnostic(AecCore* self, int enable) { self->delay_agnostic_enabled = enable; } int WebRtcAec_delay_agnostic_enabled(AecCore* self) { return self->delay_agnostic_enabled; } void WebRtcAec_enable_extended_filter(AecCore* self, int enable) { self->extended_filter_enabled = enable; self->num_partitions = enable ? kExtendedNumPartitions : kNormalNumPartitions; // Update the delay estimator with filter length. See InitAEC() for details. WebRtc_set_allowed_offset(self->delay_estimator, self->num_partitions / 2); } int WebRtcAec_extended_filter_enabled(AecCore* self) { return self->extended_filter_enabled; } int WebRtcAec_system_delay(AecCore* self) { return self->system_delay; } void WebRtcAec_SetSystemDelay(AecCore* self, int delay) { assert(delay >= 0); self->system_delay = delay; }