generic-library/webrtc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Remove the different block lengths in ns_core

Browse Source 
Relanding the CL: https://webrtc-codereview.appspot.com/30539004/
It had to be reverted because some development code was uploaded by mistake.

TBR=bjornv@webrtc.org

BUG=webrtc:3811

Review URL: https://webrtc-codereview.appspot.com/28589005

git-svn-id: http://webrtc.googlecode.com/svn/trunk@7307 4adac7df-926f-26a2-2b94-8c16560cd09d
This commit is contained in:
aluebs@webrtc.org 2014-09-26 14:41:19 +00:00
parent 5088377d70
commit 384d05f362
2 changed files with 389 additions and 436 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

822
webrtc/modules/audio_processing/ns/ns_core.c
View File

@ -90,24 +90,18 @@ int WebRtcNs_InitCore(NSinst_t* inst, uint32_t fs) {
if (fs == 8000) { if (fs == 8000) {
// We only support 10ms frames // We only support 10ms frames
inst->blockLen = 80; inst->blockLen = 80;
inst->blockLen10ms = 80;
inst->anaLen = 128; inst->anaLen = 128;
inst->window = kBlocks80w128; inst->window = kBlocks80w128;
inst->outLen = 0;
} else if (fs == 16000) { } else if (fs == 16000) {
// We only support 10ms frames // We only support 10ms frames
inst->blockLen = 160; inst->blockLen = 160;
inst->blockLen10ms = 160;
inst->anaLen = 256; inst->anaLen = 256;
inst->window = kBlocks160w256; inst->window = kBlocks160w256;
inst->outLen = 0;
} else if (fs == 32000) { } else if (fs == 32000) {
// We only support 10ms frames // We only support 10ms frames
inst->blockLen = 160; inst->blockLen = 160;
inst->blockLen10ms = 160;
inst->anaLen = 256; inst->anaLen = 256;
inst->window = kBlocks160w256; inst->window = kBlocks160w256;
inst->outLen = 0;
} }
inst->magnLen = inst->anaLen / 2 + 1; // Number of frequency bins inst->magnLen = inst->anaLen / 2 + 1; // Number of frequency bins
@ -148,13 +142,13 @@ int WebRtcNs_InitCore(NSinst_t* inst, uint32_t fs) {
// initialize variables for new method // initialize variables for new method
inst->priorSpeechProb = (float)0.5; // prior prob for speech/noise inst->priorSpeechProb = (float)0.5; // prior prob for speech/noise
for (i = 0; i < HALF_ANAL_BLOCKL; i++) { for (i = 0; i < HALF_ANAL_BLOCKL; i++) {
inst->magnPrev[i] = (float)0.0; // previous mag spectrum inst->magnPrev[i] = (float)0.0; // previous mag spectrum
inst->noisePrev[i] = (float)0.0; // previous noise-spectrum inst->noisePrev[i] = (float)0.0; // previous noise-spectrum
inst->logLrtTimeAvg[i] = inst->logLrtTimeAvg[i] =
LRT_FEATURE_THR; // smooth LR ratio (same as threshold) LRT_FEATURE_THR; // smooth LR ratio (same as threshold)
inst->magnAvgPause[i] = (float)0.0; // conservative noise spectrum estimate inst->magnAvgPause[i] = (float)0.0; // conservative noise spectrum estimate
inst->speechProb[i] = (float)0.0; // for estimation of HB in second pass inst->speechProb[i] = (float)0.0; // for estimation of HB in second pass
inst->initMagnEst[i] = (float)0.0; // initial average mag spectrum inst->initMagnEst[i] = (float)0.0; // initial average mag spectrum
} }
// feature quantities // feature quantities
@ -215,8 +209,6 @@ int WebRtcNs_InitCore(NSinst_t* inst, uint32_t fs) {
// default mode // default mode
WebRtcNs_set_policy_core(inst, 0); WebRtcNs_set_policy_core(inst, 0);
memset(inst->outBuf, 0, sizeof(float) * 3 * BLOCKL_MAX);
inst->initFlag = 1; inst->initFlag = 1;
return 0; return 0;
} }
@ -789,250 +781,245 @@ int WebRtcNs_AnalyzeCore(NSinst_t* inst, float* speechFrame) {
// update analysis buffer for L band // update analysis buffer for L band
memcpy(inst->analyzeBuf, memcpy(inst->analyzeBuf,
inst->analyzeBuf + inst->blockLen10ms, inst->analyzeBuf + inst->blockLen,
sizeof(float) * (inst->anaLen - inst->blockLen10ms)); sizeof(float) * (inst->anaLen - inst->blockLen));
memcpy(inst->analyzeBuf + inst->anaLen - inst->blockLen10ms, memcpy(inst->analyzeBuf + inst->anaLen - inst->blockLen,
speechFrame, speechFrame,
sizeof(float) * inst->blockLen10ms); sizeof(float) * inst->blockLen);
// check if processing needed // windowing
if (inst->outLen == 0) { energy = 0.0;
// windowing for (i = 0; i < inst->anaLen; i++) {
energy = 0.0; winData[i] = inst->window[i] * inst->analyzeBuf[i];
for (i = 0; i < inst->anaLen; i++) { energy += winData[i] * winData[i];
winData[i] = inst->window[i] * inst->analyzeBuf[i]; }
energy += winData[i] * winData[i]; if (energy == 0.0) {
} // we want to avoid updating statistics in this case:
if (energy == 0.0) { // Updating feature statistics when we have zeros only will cause
// we want to avoid updating statistics in this case: // thresholds to move towards zero signal situations. This in turn has the
// Updating feature statistics when we have zeros only will cause // effect that once the signal is "turned on" (non-zero values) everything
// thresholds to move towards zero signal situations. This in turn has the // will be treated as speech and there is no noise suppression effect.
// effect that once the signal is "turned on" (non-zero values) everything // Depending on the duration of the inactive signal it takes a
// will be treated as speech and there is no noise suppression effect. // considerable amount of time for the system to learn what is noise and
// Depending on the duration of the inactive signal it takes a // what is speech.
// considerable amount of time for the system to learn what is noise and return 0;
// what is speech. }
return 0;
}
// //
inst->blockInd++; // Update the block index only when we process a block. inst->blockInd++; // Update the block index only when we process a block.
// FFT // FFT
WebRtc_rdft(inst->anaLen, 1, winData, inst->ip, inst->wfft); WebRtc_rdft(inst->anaLen, 1, winData, inst->ip, inst->wfft);
imag[0] = 0; imag[0] = 0;
real[0] = winData[0]; real[0] = winData[0];
magn[0] = (float)(fabs(real[0]) + 1.0f); magn[0] = (float)(fabs(real[0]) + 1.0f);
imag[inst->magnLen - 1] = 0; imag[inst->magnLen - 1] = 0;
real[inst->magnLen - 1] = winData[1]; real[inst->magnLen - 1] = winData[1];
magn[inst->magnLen - 1] = (float)(fabs(real[inst->magnLen - 1]) + 1.0f); magn[inst->magnLen - 1] = (float)(fabs(real[inst->magnLen - 1]) + 1.0f);
signalEnergy = (float)(real[0] * real[0]) + signalEnergy = (float)(real[0] * real[0]) +
(float)(real[inst->magnLen - 1] * real[inst->magnLen - 1]); (float)(real[inst->magnLen - 1] * real[inst->magnLen - 1]);
sumMagn = magn[0] + magn[inst->magnLen - 1]; sumMagn = magn[0] + magn[inst->magnLen - 1];
if (inst->blockInd < END_STARTUP_SHORT) {
tmpFloat2 = log((float)(inst->magnLen - 1));
sum_log_i = tmpFloat2;
sum_log_i_square = tmpFloat2 * tmpFloat2;
tmpFloat1 = log(magn[inst->magnLen - 1]);
sum_log_magn = tmpFloat1;
sum_log_i_log_magn = tmpFloat2 * tmpFloat1;
}
for (i = 1; i < inst->magnLen - 1; i++) {
real[i] = winData[2 * i];
imag[i] = winData[2 * i + 1];
// magnitude spectrum
fTmp = real[i] * real[i];
fTmp += imag[i] * imag[i];
signalEnergy += fTmp;
magn[i] = ((float)sqrt(fTmp)) + 1.0f;
sumMagn += magn[i];
if (inst->blockInd < END_STARTUP_SHORT) { if (inst->blockInd < END_STARTUP_SHORT) {
tmpFloat2 = log((float)(inst->magnLen - 1)); if (i >= kStartBand) {
sum_log_i = tmpFloat2; tmpFloat2 = log((float)i);
sum_log_i_square = tmpFloat2 * tmpFloat2; sum_log_i += tmpFloat2;
tmpFloat1 = log(magn[inst->magnLen - 1]); sum_log_i_square += tmpFloat2 * tmpFloat2;
sum_log_magn = tmpFloat1; tmpFloat1 = log(magn[i]);
sum_log_i_log_magn = tmpFloat2 * tmpFloat1; sum_log_magn += tmpFloat1;
} sum_log_i_log_magn += tmpFloat2 * tmpFloat1;
for (i = 1; i < inst->magnLen - 1; i++) {
real[i] = winData[2 * i];
imag[i] = winData[2 * i + 1];
// magnitude spectrum
fTmp = real[i] * real[i];
fTmp += imag[i] * imag[i];
signalEnergy += fTmp;
magn[i] = ((float)sqrt(fTmp)) + 1.0f;
sumMagn += magn[i];
if (inst->blockInd < END_STARTUP_SHORT) {
if (i >= kStartBand) {
tmpFloat2 = log((float)i);
sum_log_i += tmpFloat2;
sum_log_i_square += tmpFloat2 * tmpFloat2;
tmpFloat1 = log(magn[i]);
sum_log_magn += tmpFloat1;
sum_log_i_log_magn += tmpFloat2 * tmpFloat1;
}
} }
} }
signalEnergy = signalEnergy / ((float)inst->magnLen); }
inst->signalEnergy = signalEnergy; signalEnergy = signalEnergy / ((float)inst->magnLen);
inst->sumMagn = sumMagn; inst->signalEnergy = signalEnergy;
inst->sumMagn = sumMagn;
// compute spectral flatness on input spectrum // compute spectral flatness on input spectrum
WebRtcNs_ComputeSpectralFlatness(inst, magn); WebRtcNs_ComputeSpectralFlatness(inst, magn);
// quantile noise estimate // quantile noise estimate
WebRtcNs_NoiseEstimation(inst, magn, noise); WebRtcNs_NoiseEstimation(inst, magn, noise);
// compute simplified noise model during startup // compute simplified noise model during startup
if (inst->blockInd < END_STARTUP_SHORT) { if (inst->blockInd < END_STARTUP_SHORT) {
// Estimate White noise // Estimate White noise
inst->whiteNoiseLevel += inst->whiteNoiseLevel += sumMagn / ((float)inst->magnLen) * inst->overdrive;
sumMagn / ((float)inst->magnLen) * inst->overdrive; // Estimate Pink noise parameters
// Estimate Pink noise parameters tmpFloat1 = sum_log_i_square * ((float)(inst->magnLen - kStartBand));
tmpFloat1 = sum_log_i_square * ((float)(inst->magnLen - kStartBand)); tmpFloat1 -= (sum_log_i * sum_log_i);
tmpFloat1 -= (sum_log_i * sum_log_i); tmpFloat2 =
tmpFloat2 = (sum_log_i_square * sum_log_magn - sum_log_i * sum_log_i_log_magn);
(sum_log_i_square * sum_log_magn - sum_log_i * sum_log_i_log_magn); tmpFloat3 = tmpFloat2 / tmpFloat1;
tmpFloat3 = tmpFloat2 / tmpFloat1; // Constrain the estimated spectrum to be positive
// Constrain the estimated spectrum to be positive if (tmpFloat3 < 0.0f) {
if (tmpFloat3 < 0.0f) { tmpFloat3 = 0.0f;
tmpFloat3 = 0.0f; }
} inst->pinkNoiseNumerator += tmpFloat3;
inst->pinkNoiseNumerator += tmpFloat3; tmpFloat2 = (sum_log_i * sum_log_magn);
tmpFloat2 = (sum_log_i * sum_log_magn); tmpFloat2 -= ((float)(inst->magnLen - kStartBand)) * sum_log_i_log_magn;
tmpFloat2 -= ((float)(inst->magnLen - kStartBand)) * sum_log_i_log_magn; tmpFloat3 = tmpFloat2 / tmpFloat1;
tmpFloat3 = tmpFloat2 / tmpFloat1; // Constrain the pink noise power to be in the interval [0, 1];
// Constrain the pink noise power to be in the interval [0, 1]; if (tmpFloat3 < 0.0f) {
if (tmpFloat3 < 0.0f) { tmpFloat3 = 0.0f;
tmpFloat3 = 0.0f; }
} if (tmpFloat3 > 1.0f) {
if (tmpFloat3 > 1.0f) { tmpFloat3 = 1.0f;
tmpFloat3 = 1.0f; }
} inst->pinkNoiseExp += tmpFloat3;
inst->pinkNoiseExp += tmpFloat3;
// Calculate frequency independent parts of parametric noise estimate. // Calculate frequency independent parts of parametric noise estimate.
if (inst->pinkNoiseExp > 0.0f) { if (inst->pinkNoiseExp > 0.0f) {
// Use pink noise estimate
parametric_num =
exp(inst->pinkNoiseNumerator / (float)(inst->blockInd + 1));
parametric_num *= (float)(inst->blockInd + 1);
parametric_exp = inst->pinkNoiseExp / (float)(inst->blockInd + 1);
}
for (i = 0; i < inst->magnLen; i++) {
// Estimate the background noise using the white and pink noise
// parameters
if (inst->pinkNoiseExp == 0.0f) {
// Use white noise estimate
inst->parametricNoise[i] = inst->whiteNoiseLevel;
} else {
// Use pink noise estimate // Use pink noise estimate
parametric_num = float use_band = (float)(i < kStartBand ? kStartBand : i);
exp(inst->pinkNoiseNumerator / (float)(inst->blockInd + 1)); inst->parametricNoise[i] =
parametric_num *= (float)(inst->blockInd + 1); parametric_num / pow(use_band, parametric_exp);
parametric_exp = inst->pinkNoiseExp / (float)(inst->blockInd + 1);
}
for (i = 0; i < inst->magnLen; i++) {
// Estimate the background noise using the white and pink noise
// parameters
if (inst->pinkNoiseExp == 0.0f) {
// Use white noise estimate
inst->parametricNoise[i] = inst->whiteNoiseLevel;
} else {
// Use pink noise estimate
float use_band = (float)(i < kStartBand ? kStartBand : i);
inst->parametricNoise[i] =
parametric_num / pow(use_band, parametric_exp);
}
// Weight quantile noise with modeled noise
noise[i] *= (inst->blockInd);
tmpFloat2 =
inst->parametricNoise[i] * (END_STARTUP_SHORT - inst->blockInd);
noise[i] += (tmpFloat2 / (float)(inst->blockInd + 1));
noise[i] /= END_STARTUP_SHORT;
} }
// Weight quantile noise with modeled noise
noise[i] *= (inst->blockInd);
tmpFloat2 =
inst->parametricNoise[i] * (END_STARTUP_SHORT - inst->blockInd);
noise[i] += (tmpFloat2 / (float)(inst->blockInd + 1));
noise[i] /= END_STARTUP_SHORT;
} }
// compute average signal during END_STARTUP_LONG time: }
// used to normalize spectral difference measure // compute average signal during END_STARTUP_LONG time:
if (inst->blockInd < END_STARTUP_LONG) { // used to normalize spectral difference measure
inst->featureData[5] *= inst->blockInd; if (inst->blockInd < END_STARTUP_LONG) {
inst->featureData[5] += signalEnergy; inst->featureData[5] *= inst->blockInd;
inst->featureData[5] /= (inst->blockInd + 1); inst->featureData[5] += signalEnergy;
} inst->featureData[5] /= (inst->blockInd + 1);
}
// start processing at frames == converged+1 // start processing at frames == converged+1
// STEP 1: compute prior and post snr based on quantile noise est // STEP 1: compute prior and post snr based on quantile noise est
// compute DD estimate of prior SNR: needed for new method // compute DD estimate of prior SNR: needed for new method
for (i = 0; i < inst->magnLen; i++) { for (i = 0; i < inst->magnLen; i++) {
// post snr // post snr
snrLocPost[i] = (float)0.0; snrLocPost[i] = (float)0.0;
if (magn[i] > noise[i]) { if (magn[i] > noise[i]) {
snrLocPost[i] = magn[i] / (noise[i] + (float)0.0001) - (float)1.0; snrLocPost[i] = magn[i] / (noise[i] + (float)0.0001) - (float)1.0;
} }
// previous post snr // previous post snr
// previous estimate: based on previous frame with gain filter // previous estimate: based on previous frame with gain filter
inst->previousEstimateStsa[i] = inst->magnPrev[i] / inst->previousEstimateStsa[i] = inst->magnPrev[i] /
(inst->noisePrev[i] + (float)0.0001) * (inst->noisePrev[i] + (float)0.0001) *
(inst->smooth[i]); (inst->smooth[i]);
// DD estimate is sum of two terms: current estimate and previous estimate // DD estimate is sum of two terms: current estimate and previous estimate
// directed decision update of snrPrior // directed decision update of snrPrior
snrLocPrior[i] = DD_PR_SNR * inst->previousEstimateStsa[i] + snrLocPrior[i] = DD_PR_SNR * inst->previousEstimateStsa[i] +
((float)1.0 - DD_PR_SNR) * snrLocPost[i]; ((float)1.0 - DD_PR_SNR) * snrLocPost[i];
// post and prior snr needed for step 2 // post and prior snr needed for step 2
} // end of loop over freqs } // end of loop over freqs
// done with step 1: dd computation of prior and post snr // done with step 1: dd computation of prior and post snr
// STEP 2: compute speech/noise likelihood // STEP 2: compute speech/noise likelihood
// compute difference of input spectrum with learned/estimated noise // compute difference of input spectrum with learned/estimated noise
// spectrum // spectrum
WebRtcNs_ComputeSpectralDifference(inst, magn); WebRtcNs_ComputeSpectralDifference(inst, magn);
// compute histograms for parameter decisions (thresholds and weights for // compute histograms for parameter decisions (thresholds and weights for
// features) // features)
// parameters are extracted once every window time // parameters are extracted once every window time
// (=inst->modelUpdatePars[1]) // (=inst->modelUpdatePars[1])
if (updateParsFlag >= 1) { if (updateParsFlag >= 1) {
// counter update // counter update
inst->modelUpdatePars[3]--; inst->modelUpdatePars[3]--;
// update histogram // update histogram
if (inst->modelUpdatePars[3] > 0) { if (inst->modelUpdatePars[3] > 0) {
WebRtcNs_FeatureParameterExtraction(inst, 0); WebRtcNs_FeatureParameterExtraction(inst, 0);
} }
// compute model parameters // compute model parameters
if (inst->modelUpdatePars[3] == 0) { if (inst->modelUpdatePars[3] == 0) {
WebRtcNs_FeatureParameterExtraction(inst, 1); WebRtcNs_FeatureParameterExtraction(inst, 1);
inst->modelUpdatePars[3] = inst->modelUpdatePars[1]; inst->modelUpdatePars[3] = inst->modelUpdatePars[1];
// if wish to update only once, set flag to zero // if wish to update only once, set flag to zero
if (updateParsFlag == 1) { if (updateParsFlag == 1) {
inst->modelUpdatePars[0] = 0; inst->modelUpdatePars[0] = 0;
} else { } else {
// update every window: // update every window:
// get normalization for spectral difference for next window estimate // get normalization for spectral difference for next window estimate
inst->featureData[6] = inst->featureData[6] =
inst->featureData[6] / ((float)inst->modelUpdatePars[1]); inst->featureData[6] / ((float)inst->modelUpdatePars[1]);
inst->featureData[5] = inst->featureData[5] =
(float)0.5 * (inst->featureData[6] + inst->featureData[5]); (float)0.5 * (inst->featureData[6] + inst->featureData[5]);
inst->featureData[6] = (float)0.0; inst->featureData[6] = (float)0.0;
}
} }
} }
// compute speech/noise probability }
WebRtcNs_SpeechNoiseProb(inst, inst->speechProb, snrLocPrior, snrLocPost); // compute speech/noise probability
// time-avg parameter for noise update WebRtcNs_SpeechNoiseProb(inst, inst->speechProb, snrLocPrior, snrLocPost);
// time-avg parameter for noise update
gammaNoiseTmp = NOISE_UPDATE;
for (i = 0; i < inst->magnLen; i++) {
probSpeech = inst->speechProb[i];
probNonSpeech = (float)1.0 - probSpeech;
// temporary noise update:
// use it for speech frames if update value is less than previous
noiseUpdateTmp =
gammaNoiseTmp * inst->noisePrev[i] +
((float)1.0 - gammaNoiseTmp) *
(probNonSpeech * magn[i] + probSpeech * inst->noisePrev[i]);
//
// time-constant based on speech/noise state
gammaNoiseOld = gammaNoiseTmp;
gammaNoiseTmp = NOISE_UPDATE; gammaNoiseTmp = NOISE_UPDATE;
for (i = 0; i < inst->magnLen; i++) { // increase gamma (i.e., less noise update) for frame likely to be speech
probSpeech = inst->speechProb[i]; if (probSpeech > PROB_RANGE) {
probNonSpeech = (float)1.0 - probSpeech; gammaNoiseTmp = SPEECH_UPDATE;
// temporary noise update: }
// use it for speech frames if update value is less than previous // conservative noise update
noiseUpdateTmp = if (probSpeech < PROB_RANGE) {
inst->magnAvgPause[i] += GAMMA_PAUSE * (magn[i] - inst->magnAvgPause[i]);
}
// noise update
if (gammaNoiseTmp == gammaNoiseOld) {
noise[i] = noiseUpdateTmp;
} else {
noise[i] =
gammaNoiseTmp * inst->noisePrev[i] + gammaNoiseTmp * inst->noisePrev[i] +
((float)1.0 - gammaNoiseTmp) * ((float)1.0 - gammaNoiseTmp) *
(probNonSpeech * magn[i] + probSpeech * inst->noisePrev[i]); (probNonSpeech * magn[i] + probSpeech * inst->noisePrev[i]);
// // allow for noise update downwards:
// time-constant based on speech/noise state // if noise update decreases the noise, it is safe, so allow it to
gammaNoiseOld = gammaNoiseTmp; // happen
gammaNoiseTmp = NOISE_UPDATE; if (noiseUpdateTmp < noise[i]) {
// increase gamma (i.e., less noise update) for frame likely to be speech
if (probSpeech > PROB_RANGE) {
gammaNoiseTmp = SPEECH_UPDATE;
}
// conservative noise update
if (probSpeech < PROB_RANGE) {
inst->magnAvgPause[i] +=
GAMMA_PAUSE * (magn[i] - inst->magnAvgPause[i]);
}
// noise update
if (gammaNoiseTmp == gammaNoiseOld) {
noise[i] = noiseUpdateTmp; noise[i] = noiseUpdateTmp;
} else {
noise[i] =
gammaNoiseTmp * inst->noisePrev[i] +
((float)1.0 - gammaNoiseTmp) *
(probNonSpeech * magn[i] + probSpeech * inst->noisePrev[i]);
// allow for noise update downwards:
// if noise update decreases the noise, it is safe, so allow it to
// happen
if (noiseUpdateTmp < noise[i]) {
noise[i] = noiseUpdateTmp;
}
} }
} // end of freq loop
// done with step 2: noise update
// keep track of noise spectrum for next frame
for (i = 0; i < inst->magnLen; i++) {
inst->noisePrev[i] = noise[i];
} }
} // end of if inst->outLen == 0 } // end of freq loop
// done with step 2: noise update
// keep track of noise spectrum for next frame
for (i = 0; i < inst->magnLen; i++) {
inst->noisePrev[i] = noise[i];
}
return 0; return 0;
} }
@ -1081,194 +1068,30 @@ int WebRtcNs_ProcessCore(NSinst_t* inst,
// update analysis buffer for L band // update analysis buffer for L band
memcpy(inst->dataBuf, memcpy(inst->dataBuf,
inst->dataBuf + inst->blockLen10ms, inst->dataBuf + inst->blockLen,
sizeof(float) * (inst->anaLen - inst->blockLen10ms)); sizeof(float) * (inst->anaLen - inst->blockLen));
memcpy(inst->dataBuf + inst->anaLen - inst->blockLen10ms, memcpy(inst->dataBuf + inst->anaLen - inst->blockLen,
speechFrame, speechFrame,
sizeof(float) * inst->blockLen10ms); sizeof(float) * inst->blockLen);
if (flagHB == 1) { if (flagHB == 1) {
// update analysis buffer for H band // update analysis buffer for H band
memcpy(inst->dataBufHB, memcpy(inst->dataBufHB,
inst->dataBufHB + inst->blockLen10ms, inst->dataBufHB + inst->blockLen,
sizeof(float) * (inst->anaLen - inst->blockLen10ms)); sizeof(float) * (inst->anaLen - inst->blockLen));
memcpy(inst->dataBufHB + inst->anaLen - inst->blockLen10ms, memcpy(inst->dataBufHB + inst->anaLen - inst->blockLen,
speechFrameHB, speechFrameHB,
sizeof(float) * inst->blockLen10ms); sizeof(float) * inst->blockLen);
} }
// check if processing needed // windowing
if (inst->outLen == 0) { energy1 = 0.0;
// windowing for (i = 0; i < inst->anaLen; i++) {
energy1 = 0.0; winData[i] = inst->window[i] * inst->dataBuf[i];
for (i = 0; i < inst->anaLen; i++) { energy1 += winData[i] * winData[i];
winData[i] = inst->window[i] * inst->dataBuf[i]; }
energy1 += winData[i] * winData[i]; if (energy1 == 0.0) {
} // synthesize the special case of zero input
if (energy1 == 0.0) {
// synthesize the special case of zero input
// read out fully processed segment
for (i = inst->windShift; i < inst->blockLen + inst->windShift; i++) {
fout[i - inst->windShift] = inst->syntBuf[i];
}
// update synthesis buffer
memcpy(inst->syntBuf,
inst->syntBuf + inst->blockLen,
sizeof(float) * (inst->anaLen - inst->blockLen));
memset(inst->syntBuf + inst->anaLen - inst->blockLen,
0,
sizeof(float) * inst->blockLen);
// out buffer
inst->outLen = inst->blockLen - inst->blockLen10ms;
if (inst->blockLen > inst->blockLen10ms) {
for (i = 0; i < inst->outLen; i++) {
inst->outBuf[i] = fout[i + inst->blockLen10ms];
}
}
for (i = 0; i < inst->blockLen10ms; ++i)
outFrame[i] = WEBRTC_SPL_SAT(
WEBRTC_SPL_WORD16_MAX, fout[i], WEBRTC_SPL_WORD16_MIN);
// for time-domain gain of HB
if (flagHB == 1)
for (i = 0; i < inst->blockLen10ms; ++i)
outFrameHB[i] = WEBRTC_SPL_SAT(
WEBRTC_SPL_WORD16_MAX, inst->dataBufHB[i], WEBRTC_SPL_WORD16_MIN);
return 0;
}
// FFT
WebRtc_rdft(inst->anaLen, 1, winData, inst->ip, inst->wfft);
imag[0] = 0;
real[0] = winData[0];
magn[0] = (float)(fabs(real[0]) + 1.0f);
imag[inst->magnLen - 1] = 0;
real[inst->magnLen - 1] = winData[1];
magn[inst->magnLen - 1] = (float)(fabs(real[inst->magnLen - 1]) + 1.0f);
if (inst->blockInd < END_STARTUP_SHORT) {
inst->initMagnEst[0] += magn[0];
inst->initMagnEst[inst->magnLen - 1] += magn[inst->magnLen - 1];
}
for (i = 1; i < inst->magnLen - 1; i++) {
real[i] = winData[2 * i];
imag[i] = winData[2 * i + 1];
// magnitude spectrum
fTmp = real[i] * real[i];
fTmp += imag[i] * imag[i];
magn[i] = ((float)sqrt(fTmp)) + 1.0f;
if (inst->blockInd < END_STARTUP_SHORT) {
inst->initMagnEst[i] += magn[i];
}
}
// Compute dd update of prior snr and post snr based on new noise estimate
for (i = 0; i < inst->magnLen; i++) {
// post and prior snr
currentEstimateStsa = (float)0.0;
if (magn[i] > inst->noisePrev[i]) {
currentEstimateStsa =
magn[i] / (inst->noisePrev[i] + (float)0.0001) - (float)1.0;
}
// DD estimate is sume of two terms: current estimate and previous
// estimate
// directed decision update of snrPrior
snrPrior = DD_PR_SNR * inst->previousEstimateStsa[i] +
((float)1.0 - DD_PR_SNR) * currentEstimateStsa;
// gain filter
tmpFloat1 = inst->overdrive + snrPrior;
tmpFloat2 = (float)snrPrior / tmpFloat1;
theFilter[i] = (float)tmpFloat2;
} // end of loop over freqs
for (i = 0; i < inst->magnLen; i++) {
// flooring bottom
if (theFilter[i] < inst->denoiseBound) {
theFilter[i] = inst->denoiseBound;
}
// flooring top
if (theFilter[i] > (float)1.0) {
theFilter[i] = 1.0;
}
if (inst->blockInd < END_STARTUP_SHORT) {
theFilterTmp[i] =
(inst->initMagnEst[i] - inst->overdrive * inst->parametricNoise[i]);
theFilterTmp[i] /= (inst->initMagnEst[i] + (float)0.0001);
// flooring bottom
if (theFilterTmp[i] < inst->denoiseBound) {
theFilterTmp[i] = inst->denoiseBound;
}
// flooring top
if (theFilterTmp[i] > (float)1.0) {
theFilterTmp[i] = 1.0;
}
// Weight the two suppression filters
theFilter[i] *= (inst->blockInd);
theFilterTmp[i] *= (END_STARTUP_SHORT - inst->blockInd);
theFilter[i] += theFilterTmp[i];
theFilter[i] /= (END_STARTUP_SHORT);
}
// smoothing
inst->smooth[i] = theFilter[i];
real[i] *= inst->smooth[i];
imag[i] *= inst->smooth[i];
}
// keep track of magn spectrum for next frame
for (i = 0; i < inst->magnLen; i++) {
inst->magnPrev[i] = magn[i];
}
// back to time domain
winData[0] = real[0];
winData[1] = real[inst->magnLen - 1];
for (i = 1; i < inst->magnLen - 1; i++) {
winData[2 * i] = real[i];
winData[2 * i + 1] = imag[i];
}
WebRtc_rdft(inst->anaLen, -1, winData, inst->ip, inst->wfft);
for (i = 0; i < inst->anaLen; i++) {
real[i] = 2.0f * winData[i] / inst->anaLen; // fft scaling
}
// scale factor: only do it after END_STARTUP_LONG time
factor = (float)1.0;
if (inst->gainmap == 1 && inst->blockInd > END_STARTUP_LONG) {
factor1 = (float)1.0;
factor2 = (float)1.0;
energy2 = 0.0;
for (i = 0; i < inst->anaLen; i++) {
energy2 += (float)real[i] * (float)real[i];
}
gain = (float)sqrt(energy2 / (energy1 + (float)1.0));
// scaling for new version
if (gain > B_LIM) {
factor1 = (float)1.0 + (float)1.3 * (gain - B_LIM);
if (gain * factor1 > (float)1.0) {
factor1 = (float)1.0 / gain;
}
}
if (gain < B_LIM) {
// don't reduce scale too much for pause regions:
// attenuation here should be controlled by flooring
if (gain <= inst->denoiseBound) {
gain = inst->denoiseBound;
}
factor2 = (float)1.0 - (float)0.3 * (B_LIM - gain);
}
// combine both scales with speech/noise prob:
// note prior (priorSpeechProb) is not frequency dependent
factor = inst->priorSpeechProb * factor1 +
((float)1.0 - inst->priorSpeechProb) * factor2;
} // out of inst->gainmap==1
// synthesis
for (i = 0; i < inst->anaLen; i++) {
inst->syntBuf[i] += factor * inst->window[i] * (float)real[i];
}
// read out fully processed segment // read out fully processed segment
for (i = inst->windShift; i < inst->blockLen + inst->windShift; i++) { for (i = inst->windShift; i < inst->blockLen + inst->windShift; i++) {
fout[i - inst->windShift] = inst->syntBuf[i]; fout[i - inst->windShift] = inst->syntBuf[i];
@ -1281,28 +1104,161 @@ int WebRtcNs_ProcessCore(NSinst_t* inst,
0, 0,
sizeof(float) * inst->blockLen); sizeof(float) * inst->blockLen);
// out buffer for (i = 0; i < inst->blockLen; ++i)
inst->outLen = inst->blockLen - inst->blockLen10ms; outFrame[i] =
if (inst->blockLen > inst->blockLen10ms) { WEBRTC_SPL_SAT(WEBRTC_SPL_WORD16_MAX, fout[i], WEBRTC_SPL_WORD16_MIN);
for (i = 0; i < inst->outLen; i++) {
inst->outBuf[i] = fout[i + inst->blockLen10ms]; // for time-domain gain of HB
} if (flagHB == 1)
for (i = 0; i < inst->blockLen; ++i)
outFrameHB[i] = WEBRTC_SPL_SAT(
WEBRTC_SPL_WORD16_MAX, inst->dataBufHB[i], WEBRTC_SPL_WORD16_MIN);
return 0;
}
// FFT
WebRtc_rdft(inst->anaLen, 1, winData, inst->ip, inst->wfft);
imag[0] = 0;
real[0] = winData[0];
magn[0] = (float)(fabs(real[0]) + 1.0f);
imag[inst->magnLen - 1] = 0;
real[inst->magnLen - 1] = winData[1];
magn[inst->magnLen - 1] = (float)(fabs(real[inst->magnLen - 1]) + 1.0f);
if (inst->blockInd < END_STARTUP_SHORT) {
inst->initMagnEst[0] += magn[0];
inst->initMagnEst[inst->magnLen - 1] += magn[inst->magnLen - 1];
}
for (i = 1; i < inst->magnLen - 1; i++) {
real[i] = winData[2 * i];
imag[i] = winData[2 * i + 1];
// magnitude spectrum
fTmp = real[i] * real[i];
fTmp += imag[i] * imag[i];
magn[i] = ((float)sqrt(fTmp)) + 1.0f;
if (inst->blockInd < END_STARTUP_SHORT) {
inst->initMagnEst[i] += magn[i];
} }
} // end of if out.len==0
else {
for (i = 0; i < inst->blockLen10ms; i++) {
fout[i] = inst->outBuf[i];
}
memcpy(inst->outBuf,
inst->outBuf + inst->blockLen10ms,
sizeof(float) * (inst->outLen - inst->blockLen10ms));
memset(inst->outBuf + inst->outLen - inst->blockLen10ms,
0,
sizeof(float) * inst->blockLen10ms);
inst->outLen -= inst->blockLen10ms;
} }
for (i = 0; i < inst->blockLen10ms; ++i) // Compute dd update of prior snr and post snr based on new noise estimate
for (i = 0; i < inst->magnLen; i++) {
// post and prior snr
currentEstimateStsa = (float)0.0;
if (magn[i] > inst->noisePrev[i]) {
currentEstimateStsa =
magn[i] / (inst->noisePrev[i] + (float)0.0001) - (float)1.0;
}
// DD estimate is sume of two terms: current estimate and previous
// estimate
// directed decision update of snrPrior
snrPrior = DD_PR_SNR * inst->previousEstimateStsa[i] +
((float)1.0 - DD_PR_SNR) * currentEstimateStsa;
// gain filter
tmpFloat1 = inst->overdrive + snrPrior;
tmpFloat2 = (float)snrPrior / tmpFloat1;
theFilter[i] = (float)tmpFloat2;
} // end of loop over freqs
for (i = 0; i < inst->magnLen; i++) {
// flooring bottom
if (theFilter[i] < inst->denoiseBound) {
theFilter[i] = inst->denoiseBound;
}
// flooring top
if (theFilter[i] > (float)1.0) {
theFilter[i] = 1.0;
}
if (inst->blockInd < END_STARTUP_SHORT) {
theFilterTmp[i] =
(inst->initMagnEst[i] - inst->overdrive * inst->parametricNoise[i]);
theFilterTmp[i] /= (inst->initMagnEst[i] + (float)0.0001);
// flooring bottom
if (theFilterTmp[i] < inst->denoiseBound) {
theFilterTmp[i] = inst->denoiseBound;
}
// flooring top
if (theFilterTmp[i] > (float)1.0) {
theFilterTmp[i] = 1.0;
}
// Weight the two suppression filters
theFilter[i] *= (inst->blockInd);
theFilterTmp[i] *= (END_STARTUP_SHORT - inst->blockInd);
theFilter[i] += theFilterTmp[i];
theFilter[i] /= (END_STARTUP_SHORT);
}
// smoothing
inst->smooth[i] = theFilter[i];
real[i] *= inst->smooth[i];
imag[i] *= inst->smooth[i];
}
// keep track of magn spectrum for next frame
for (i = 0; i < inst->magnLen; i++) {
inst->magnPrev[i] = magn[i];
}
// back to time domain
winData[0] = real[0];
winData[1] = real[inst->magnLen - 1];
for (i = 1; i < inst->magnLen - 1; i++) {
winData[2 * i] = real[i];
winData[2 * i + 1] = imag[i];
}
WebRtc_rdft(inst->anaLen, -1, winData, inst->ip, inst->wfft);
for (i = 0; i < inst->anaLen; i++) {
real[i] = 2.0f * winData[i] / inst->anaLen; // fft scaling
}
// scale factor: only do it after END_STARTUP_LONG time
factor = (float)1.0;
if (inst->gainmap == 1 && inst->blockInd > END_STARTUP_LONG) {
factor1 = (float)1.0;
factor2 = (float)1.0;
energy2 = 0.0;
for (i = 0; i < inst->anaLen; i++) {
energy2 += (float)real[i] * (float)real[i];
}
gain = (float)sqrt(energy2 / (energy1 + (float)1.0));
// scaling for new version
if (gain > B_LIM) {
factor1 = (float)1.0 + (float)1.3 * (gain - B_LIM);
if (gain * factor1 > (float)1.0) {
factor1 = (float)1.0 / gain;
}
}
if (gain < B_LIM) {
// don't reduce scale too much for pause regions:
// attenuation here should be controlled by flooring
if (gain <= inst->denoiseBound) {
gain = inst->denoiseBound;
}
factor2 = (float)1.0 - (float)0.3 * (B_LIM - gain);
}
// combine both scales with speech/noise prob:
// note prior (priorSpeechProb) is not frequency dependent
factor = inst->priorSpeechProb * factor1 +
((float)1.0 - inst->priorSpeechProb) * factor2;
} // out of inst->gainmap==1
// synthesis
for (i = 0; i < inst->anaLen; i++) {
inst->syntBuf[i] += factor * inst->window[i] * (float)real[i];
}
// read out fully processed segment
for (i = inst->windShift; i < inst->blockLen + inst->windShift; i++) {
fout[i - inst->windShift] = inst->syntBuf[i];
}
// update synthesis buffer
memcpy(inst->syntBuf,
inst->syntBuf + inst->blockLen,
sizeof(float) * (inst->anaLen - inst->blockLen));
memset(inst->syntBuf + inst->anaLen - inst->blockLen,
0,
sizeof(float) * inst->blockLen);
for (i = 0; i < inst->blockLen; ++i)
outFrame[i] = outFrame[i] =
WEBRTC_SPL_SAT(WEBRTC_SPL_WORD16_MAX, fout[i], WEBRTC_SPL_WORD16_MIN); WEBRTC_SPL_SAT(WEBRTC_SPL_WORD16_MAX, fout[i], WEBRTC_SPL_WORD16_MIN);
@ -1343,7 +1299,7 @@ int WebRtcNs_ProcessCore(NSinst_t* inst,
gainTimeDomainHB = 1.0; gainTimeDomainHB = 1.0;
} }
// apply gain // apply gain
for (i = 0; i < inst->blockLen10ms; i++) { for (i = 0; i < inst->blockLen; i++) {
float o = gainTimeDomainHB * inst->dataBufHB[i]; float o = gainTimeDomainHB * inst->dataBufHB[i];
outFrameHB[i] = outFrameHB[i] =
WEBRTC_SPL_SAT(WEBRTC_SPL_WORD16_MAX, o, WEBRTC_SPL_WORD16_MIN); WEBRTC_SPL_SAT(WEBRTC_SPL_WORD16_MAX, o, WEBRTC_SPL_WORD16_MIN);

3
webrtc/modules/audio_processing/ns/ns_core.h
View File

@ -52,9 +52,7 @@ typedef struct NSParaExtract_t_ {
typedef struct NSinst_t_ { typedef struct NSinst_t_ {
uint32_t fs; uint32_t fs;
int blockLen; int blockLen;
int blockLen10ms;
int windShift; int windShift;
int outLen;
int anaLen; int anaLen;
int magnLen; int magnLen;
int aggrMode; int aggrMode;
@ -62,7 +60,6 @@ typedef struct NSinst_t_ {
float analyzeBuf[ANAL_BLOCKL_MAX]; float analyzeBuf[ANAL_BLOCKL_MAX];
float dataBuf[ANAL_BLOCKL_MAX]; float dataBuf[ANAL_BLOCKL_MAX];
float syntBuf[ANAL_BLOCKL_MAX]; float syntBuf[ANAL_BLOCKL_MAX];
float outBuf[3 * BLOCKL_MAX];
int initFlag; int initFlag;
// parameters for quantile noise estimation // parameters for quantile noise estimation