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

Revert 4837 "Add an extended filter mode to AEC."

Browse Source 
> Add an extended filter mode to AEC.
> 
> This mode extends the filter length from the current 48 ms to 128 ms.
> It is runtime selectable which allows it to be enabled through
> experiment. We reuse the DelayCorrection infrastructure to avoid having
> to replumb everything up to libjingle.
> 
> Increases AEC complexity by ~50% on modern x86 CPUs.
> Measurements (in percent of usage on one core):
> 
> Machine/CPU                                     Normal Extended
> MacBook Retina (Early 2013),
> Core i7 Ivy Bridge (2.7 GHz, hyperthreaded)     0.6%   0.9%
> 
> MacBook Air (Late 2010), Core 2 Duo (2.13 GHz)  1.4%   2.7%
> 
> Chromebook Pixel, Core i5 Ivy Bridge (1.8 GHz)  0.6%   1.0%
> 
> Samsung ARM Chromebook,
> Samsung Exynos 5 Dual (1.7 GHz)                 3.2%   5.6%
> 
> The relative value is large of course but the absolute should be
> acceptable in order to have a working AEC on some platforms.
> 
> Detailed changes to the algorithm:
> - The filter length is changed from 48 to 128 ms. This comes with tuning
> of several parameters: i) filter adaptation stepsize and error
> threshold; ii) non-linear processing smoothing and overdrive.
> - Option to ignore the reported delays on platforms which we deem
> sufficiently unreliable. Currently this will be enabled in Chromium for
> Mac.
> - Faster startup times by removing the excessive "startup phase"
> processing of reported delays.
> - Much more conservative adjustments to the far-end read pointer. We
> smooth the delay difference more heavily, and back off from the
> difference more. Adjustments force a readaptation of the filter, so they
> should be avoided except when really necessary.
> 
> Corresponds to these changes:
> https://chromereviews.googleplex.com/9412014
> https://chromereviews.googleplex.com/9514013
> https://chromereviews.googleplex.com/9960013
> 
> BUG=454,827,1261
> R=bjornv@webrtc.org
> 
> Review URL: https://webrtc-codereview.appspot.com/2151007

TBR=andrew@webrtc.org

Review URL: https://webrtc-codereview.appspot.com/2296005

git-svn-id: http://webrtc.googlecode.com/svn/trunk@4839 4adac7df-926f-26a2-2b94-8c16560cd09d
This commit is contained in:
asapersson@webrtc.org 2013-09-25 12:27:27 +00:00
parent 2f240b43f5
commit ce014d97cd
11 changed files with 270 additions and 595 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

124
webrtc/modules/audio_processing/aec/aec_core.c
View File

@ -109,17 +109,7 @@ const float WebRtcAec_overDriveCurve[65] = {
// Target suppression levels for nlp modes. // Target suppression levels for nlp modes.
// log{0.001, 0.00001, 0.00000001} // log{0.001, 0.00001, 0.00000001}
static const float kTargetSupp[3] = { -6.9f, -11.5f, -18.4f }; static const float kTargetSupp[3] = { -6.9f, -11.5f, -18.4f };
static const float kMinOverDrive[3] = { 1.0f, 2.0f, 5.0f };
// 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 };
static const float kExtendedSmoothingCoefficients[2][2] =
{ { 0.9f, 0.1f }, { 0.92f, 0.08f } };
static const float 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 #ifdef WEBRTC_AEC_DEBUG_DUMP
extern int webrtc_aec_instance_count; extern int webrtc_aec_instance_count;
@ -291,13 +281,13 @@ int WebRtcAec_FreeAec(AecCore* aec)
static void FilterFar(AecCore* aec, float yf[2][PART_LEN1]) static void FilterFar(AecCore* aec, float yf[2][PART_LEN1])
{ {
int i; int i;
for (i = 0; i < aec->num_partitions; i++) { for (i = 0; i < NR_PART; i++) {
int j; int j;
int xPos = (i + aec->xfBufBlockPos) * PART_LEN1; int xPos = (i + aec->xfBufBlockPos) * PART_LEN1;
int pos = i * PART_LEN1; int pos = i * PART_LEN1;
// Check for wrap // Check for wrap
if (i + aec->xfBufBlockPos >= aec->num_partitions) { if (i + aec->xfBufBlockPos >= NR_PART) {
xPos -= aec->num_partitions*(PART_LEN1); xPos -= NR_PART*(PART_LEN1);
} }
for (j = 0; j < PART_LEN1; j++) { for (j = 0; j < PART_LEN1; j++) {
@ -311,25 +301,22 @@ static void FilterFar(AecCore* aec, float yf[2][PART_LEN1])
static void ScaleErrorSignal(AecCore* aec, float ef[2][PART_LEN1]) static void ScaleErrorSignal(AecCore* aec, float ef[2][PART_LEN1])
{ {
const float mu = aec->extended_filter_enabled ? kExtendedMu : aec->normal_mu;
const float error_threshold = aec->extended_filter_enabled ?
kExtendedErrorThreshold : aec->normal_error_threshold;
int i; int i;
float abs_ef; float absEf;
for (i = 0; i < (PART_LEN1); i++) { for (i = 0; i < (PART_LEN1); i++) {
ef[0][i] /= (aec->xPow[i] + 1e-10f); ef[0][i] /= (aec->xPow[i] + 1e-10f);
ef[1][i] /= (aec->xPow[i] + 1e-10f); ef[1][i] /= (aec->xPow[i] + 1e-10f);
abs_ef = sqrtf(ef[0][i] * ef[0][i] + ef[1][i] * ef[1][i]); absEf = sqrtf(ef[0][i] * ef[0][i] + ef[1][i] * ef[1][i]);
if (abs_ef > error_threshold) { if (absEf > aec->errThresh) {
abs_ef = error_threshold / (abs_ef + 1e-10f); absEf = aec->errThresh / (absEf + 1e-10f);
ef[0][i] *= abs_ef; ef[0][i] *= absEf;
ef[1][i] *= abs_ef; ef[1][i] *= absEf;
} }
// Stepsize factor // Stepsize factor
ef[0][i] *= mu; ef[0][i] *= aec->mu;
ef[1][i] *= mu; ef[1][i] *= aec->mu;
} }
} }
@ -338,35 +325,35 @@ static void ScaleErrorSignal(AecCore* aec, float ef[2][PART_LEN1])
//static void FilterAdaptationUnconstrained(AecCore* aec, float *fft, //static void FilterAdaptationUnconstrained(AecCore* aec, float *fft,
// float ef[2][PART_LEN1]) { // float ef[2][PART_LEN1]) {
// int i, j; // int i, j;
// for (i = 0; i < aec->num_partitions; i++) { // for (i = 0; i < NR_PART; i++) {
// int xPos = (i + aec->xfBufBlockPos)*(PART_LEN1); // int xPos = (i + aec->xfBufBlockPos)*(PART_LEN1);
// int pos; // int pos;
// // Check for wrap // // Check for wrap
// if (i + aec->xfBufBlockPos >= aec->num_partitions) { // if (i + aec->xfBufBlockPos >= NR_PART) {
// xPos -= aec->num_partitions * PART_LEN1; // xPos -= NR_PART * PART_LEN1;
// } // }
// //
// pos = i * PART_LEN1; // pos = i * PART_LEN1;
// //
// for (j = 0; j < PART_LEN1; j++) { // for (j = 0; j < PART_LEN1; j++) {
// aec->wfBuf[0][pos + j] += MulRe(aec->xfBuf[0][xPos + j], // aec->wfBuf[pos + j][0] += MulRe(aec->xfBuf[xPos + j][0],
// -aec->xfBuf[1][xPos + j], // -aec->xfBuf[xPos + j][1],
// ef[0][j], ef[1][j]); // ef[j][0], ef[j][1]);
// aec->wfBuf[1][pos + j] += MulIm(aec->xfBuf[0][xPos + j], // aec->wfBuf[pos + j][1] += MulIm(aec->xfBuf[xPos + j][0],
// -aec->xfBuf[1][xPos + j], // -aec->xfBuf[xPos + j][1],
// ef[0][j], ef[1][j]); // ef[j][0], ef[j][1]);
// } // }
// } // }
//} //}
static void FilterAdaptation(AecCore* aec, float *fft, float ef[2][PART_LEN1]) { static void FilterAdaptation(AecCore* aec, float *fft, float ef[2][PART_LEN1]) {
int i, j; int i, j;
for (i = 0; i < aec->num_partitions; i++) { for (i = 0; i < NR_PART; i++) {
int xPos = (i + aec->xfBufBlockPos)*(PART_LEN1); int xPos = (i + aec->xfBufBlockPos)*(PART_LEN1);
int pos; int pos;
// Check for wrap // Check for wrap
if (i + aec->xfBufBlockPos >= aec->num_partitions) { if (i + aec->xfBufBlockPos >= NR_PART) {
xPos -= aec->num_partitions * PART_LEN1; xPos -= NR_PART * PART_LEN1;
} }
pos = i * PART_LEN1; pos = i * PART_LEN1;
@ -440,12 +427,12 @@ int WebRtcAec_InitAec(AecCore* aec, int sampFreq)
aec->sampFreq = sampFreq; aec->sampFreq = sampFreq;
if (sampFreq == 8000) { if (sampFreq == 8000) {
aec->normal_mu = 0.6f; aec->mu = 0.6f;
aec->normal_error_threshold = 2e-6f; aec->errThresh = 2e-6f;
} }
else { else {
aec->normal_mu = 0.5f; aec->mu = 0.5f;
aec->normal_error_threshold = 1.5e-6f; aec->errThresh = 1.5e-6f;
} }
if (WebRtc_InitBuffer(aec->nearFrBuf) == -1) { if (WebRtc_InitBuffer(aec->nearFrBuf) == -1) {
@ -487,9 +474,6 @@ int WebRtcAec_InitAec(AecCore* aec, int sampFreq)
aec->delay_logging_enabled = 0; aec->delay_logging_enabled = 0;
memset(aec->delay_histogram, 0, sizeof(aec->delay_histogram)); memset(aec->delay_histogram, 0, sizeof(aec->delay_histogram));
aec->extended_filter_enabled = 0;
aec->num_partitions = kNormalNumPartitions;
// Default target suppression mode. // Default target suppression mode.
aec->nlp_mode = 1; aec->nlp_mode = 1;
@ -499,7 +483,7 @@ int WebRtcAec_InitAec(AecCore* aec, int sampFreq)
aec->mult = (short)aec->sampFreq / 16000; aec->mult = (short)aec->sampFreq / 16000;
} }
else { else {
aec->mult = (short)aec->sampFreq / 8000; aec->mult = (short)aec->sampFreq / 8000;
} }
aec->farBufWritePos = 0; aec->farBufWritePos = 0;
@ -530,14 +514,11 @@ int WebRtcAec_InitAec(AecCore* aec, int sampFreq)
aec->xfBufBlockPos = 0; aec->xfBufBlockPos = 0;
// TODO: Investigate need for these initializations. Deleting them doesn't // TODO: Investigate need for these initializations. Deleting them doesn't
// change the output at all and yields 0.4% overall speedup. // change the output at all and yields 0.4% overall speedup.
memset(aec->xfBuf, 0, sizeof(complex_t) * kExtendedNumPartitions * memset(aec->xfBuf, 0, sizeof(complex_t) * NR_PART * PART_LEN1);
PART_LEN1); memset(aec->wfBuf, 0, sizeof(complex_t) * NR_PART * PART_LEN1);
memset(aec->wfBuf, 0, sizeof(complex_t) * kExtendedNumPartitions *
PART_LEN1);
memset(aec->sde, 0, sizeof(complex_t) * PART_LEN1); memset(aec->sde, 0, sizeof(complex_t) * PART_LEN1);
memset(aec->sxd, 0, sizeof(complex_t) * PART_LEN1); memset(aec->sxd, 0, sizeof(complex_t) * PART_LEN1);
memset(aec->xfwBuf, 0, sizeof(complex_t) * kExtendedNumPartitions * memset(aec->xfwBuf, 0, sizeof(complex_t) * NR_PART * PART_LEN1);
PART_LEN1);
memset(aec->se, 0, sizeof(float) * PART_LEN1); memset(aec->se, 0, sizeof(float) * PART_LEN1);
// To prevent numerical instability in the first block. // To prevent numerical instability in the first block.
@ -753,11 +734,13 @@ int WebRtcAec_GetDelayMetricsCore(AecCore* self, int* median, int* std) {
} }
int WebRtcAec_echo_state(AecCore* self) { int WebRtcAec_echo_state(AecCore* self) {
assert(self != NULL);
return self->echoState; return self->echoState;
} }
void WebRtcAec_GetEchoStats(AecCore* self, Stats* erl, Stats* erle, void WebRtcAec_GetEchoStats(AecCore* self, Stats* erl, Stats* erle,
Stats* a_nlp) { Stats* a_nlp) {
assert(self != NULL);
assert(erl != NULL); assert(erl != NULL);
assert(erle != NULL); assert(erle != NULL);
assert(a_nlp != NULL); assert(a_nlp != NULL);
@ -768,12 +751,14 @@ void WebRtcAec_GetEchoStats(AecCore* self, Stats* erl, Stats* erle,
#ifdef WEBRTC_AEC_DEBUG_DUMP #ifdef WEBRTC_AEC_DEBUG_DUMP
void* WebRtcAec_far_time_buf(AecCore* self) { void* WebRtcAec_far_time_buf(AecCore* self) {
assert(self != NULL);
return self->far_time_buf; return self->far_time_buf;
} }
#endif #endif
void WebRtcAec_SetConfigCore(AecCore* self, int nlp_mode, int metrics_mode, void WebRtcAec_SetConfigCore(AecCore* self, int nlp_mode, int metrics_mode,
int delay_logging) { int delay_logging) {
assert(self != NULL);
assert(nlp_mode >= 0 && nlp_mode < 3); assert(nlp_mode >= 0 && nlp_mode < 3);
self->nlp_mode = nlp_mode; self->nlp_mode = nlp_mode;
self->metricsMode = metrics_mode; self->metricsMode = metrics_mode;
@ -786,20 +771,13 @@ void WebRtcAec_SetConfigCore(AecCore* self, int nlp_mode, int metrics_mode,
} }
} }
void WebRtcAec_enable_delay_correction(AecCore* self, int enable) {
self->extended_filter_enabled = enable;
self->num_partitions = enable ? kExtendedNumPartitions : kNormalNumPartitions;
}
int WebRtcAec_delay_correction_enabled(AecCore* self) {
return self->extended_filter_enabled;
}
int WebRtcAec_system_delay(AecCore* self) { int WebRtcAec_system_delay(AecCore* self) {
assert(self != NULL);
return self->system_delay; return self->system_delay;
} }
void WebRtcAec_SetSystemDelay(AecCore* self, int delay) { void WebRtcAec_SetSystemDelay(AecCore* self, int delay) {
assert(self != NULL);
assert(delay >= 0); assert(delay >= 0);
self->system_delay = delay; self->system_delay = delay;
} }
@ -875,8 +853,7 @@ static void ProcessBlock(AecCore* aec) {
for (i = 0; i < PART_LEN1; i++) { for (i = 0; i < PART_LEN1; i++) {
far_spectrum = (xf_ptr[i] * xf_ptr[i]) + far_spectrum = (xf_ptr[i] * xf_ptr[i]) +
(xf_ptr[PART_LEN1 + i] * xf_ptr[PART_LEN1 + i]); (xf_ptr[PART_LEN1 + i] * xf_ptr[PART_LEN1 + i]);
aec->xPow[i] = gPow[0] * aec->xPow[i] + gPow[1] * aec->num_partitions * aec->xPow[i] = gPow[0] * aec->xPow[i] + gPow[1] * NR_PART * far_spectrum;
far_spectrum;
// Calculate absolute spectra // Calculate absolute spectra
abs_far_spectrum[i] = sqrtf(far_spectrum); abs_far_spectrum[i] = sqrtf(far_spectrum);
@ -936,7 +913,7 @@ static void ProcessBlock(AecCore* aec) {
// Update the xfBuf block position. // Update the xfBuf block position.
aec->xfBufBlockPos--; aec->xfBufBlockPos--;
if (aec->xfBufBlockPos == -1) { if (aec->xfBufBlockPos == -1) {
aec->xfBufBlockPos = aec->num_partitions - 1; aec->xfBufBlockPos = NR_PART - 1;
} }
// Buffer xf // Buffer xf
@ -1037,21 +1014,18 @@ static void NonLinearProcessing(AecCore* aec, short *output, short *outputH)
float cohde[PART_LEN1], cohxd[PART_LEN1]; float cohde[PART_LEN1], cohxd[PART_LEN1];
float hNlDeAvg, hNlXdAvg; float hNlDeAvg, hNlXdAvg;
float hNl[PART_LEN1]; float hNl[PART_LEN1];
float hNlPref[kPrefBandSize]; float hNlPref[PREF_BAND_SIZE];
float hNlFb = 0, hNlFbLow = 0; float hNlFb = 0, hNlFbLow = 0;
const float prefBandQuant = 0.75f, prefBandQuantLow = 0.5f; const float prefBandQuant = 0.75f, prefBandQuantLow = 0.5f;
const int prefBandSize = kPrefBandSize / aec->mult; const int prefBandSize = PREF_BAND_SIZE / aec->mult;
const int minPrefBand = 4 / aec->mult; const int minPrefBand = 4 / aec->mult;
// Near and error power sums // Near and error power sums
float sdSum = 0, seSum = 0; float sdSum = 0, seSum = 0;
// Power estimate smoothing coefficients. // Power estimate smoothing coefficients
const float *ptrGCoh = aec->extended_filter_enabled ? const float gCoh[2][2] = {{0.9f, 0.1f}, {0.93f, 0.07f}};
kExtendedSmoothingCoefficients[aec->mult - 1] : const float *ptrGCoh = gCoh[aec->mult - 1];
kNormalSmoothingCoefficients[aec->mult - 1];
const float* min_overdrive = aec->extended_filter_enabled ?
kExtendedMinOverDrive : kNormalMinOverDrive;
// Filter energy // Filter energy
float wfEnMax = 0, wfEn = 0; float wfEnMax = 0, wfEn = 0;
@ -1074,7 +1048,7 @@ static void NonLinearProcessing(AecCore* aec, short *output, short *outputH)
if (aec->delayEstCtr == 0) { if (aec->delayEstCtr == 0) {
wfEnMax = 0; wfEnMax = 0;
aec->delayIdx = 0; aec->delayIdx = 0;
for (i = 0; i < aec->num_partitions; i++) { for (i = 0; i < NR_PART; i++) {
pos = i * PART_LEN1; pos = i * PART_LEN1;
wfEn = 0; wfEn = 0;
for (j = 0; j < PART_LEN1; j++) { for (j = 0; j < PART_LEN1; j++) {
@ -1215,7 +1189,7 @@ static void NonLinearProcessing(AecCore* aec, short *output, short *outputH)
if (aec->hNlXdAvgMin == 1) { if (aec->hNlXdAvgMin == 1) {
aec->echoState = 0; aec->echoState = 0;
aec->overDrive = min_overdrive[aec->nlp_mode]; aec->overDrive = kMinOverDrive[aec->nlp_mode];
if (aec->stNearState == 1) { if (aec->stNearState == 1) {
memcpy(hNl, cohde, sizeof(hNl)); memcpy(hNl, cohde, sizeof(hNl));
@ -1271,7 +1245,7 @@ static void NonLinearProcessing(AecCore* aec, short *output, short *outputH)
aec->hNlMinCtr = 0; aec->hNlMinCtr = 0;
aec->overDrive = WEBRTC_SPL_MAX(kTargetSupp[aec->nlp_mode] / aec->overDrive = WEBRTC_SPL_MAX(kTargetSupp[aec->nlp_mode] /
((float)log(aec->hNlFbMin + 1e-10f) + 1e-10f), ((float)log(aec->hNlFbMin + 1e-10f) + 1e-10f),
min_overdrive[aec->nlp_mode]); kMinOverDrive[aec->nlp_mode]);
} }
// Smooth the overdrive. // Smooth the overdrive.
@ -1491,6 +1465,7 @@ static void InitStats(Stats* stats) {
} }
static void InitMetrics(AecCore* self) { static void InitMetrics(AecCore* self) {
assert(self != NULL);
self->stateCounter = 0; self->stateCounter = 0;
InitLevel(&self->farlevel); InitLevel(&self->farlevel);
InitLevel(&self->nearlevel); InitLevel(&self->nearlevel);
@ -1712,4 +1687,3 @@ static void TimeToFrequency(float time_data[PART_LEN2],
freq_data[1][i] = time_data[2 * i + 1]; freq_data[1][i] = time_data[2 * i + 1];
} }
} }

15
webrtc/modules/audio_processing/aec/aec_core.h
View File

@ -70,38 +70,23 @@ void WebRtcAec_ProcessFrame(AecCore* aec,
// Returns the number of elements moved, and adjusts |system_delay| by the // Returns the number of elements moved, and adjusts |system_delay| by the
// corresponding amount in ms. // corresponding amount in ms.
int WebRtcAec_MoveFarReadPtr(AecCore* aec, int elements); int WebRtcAec_MoveFarReadPtr(AecCore* aec, int elements);
// Calculates the median and standard deviation among the delay estimates // Calculates the median and standard deviation among the delay estimates
// collected since the last call to this function. // collected since the last call to this function.
int WebRtcAec_GetDelayMetricsCore(AecCore* self, int* median, int* std); int WebRtcAec_GetDelayMetricsCore(AecCore* self, int* median, int* std);
// Returns the echo state (1: echo, 0: no echo). // Returns the echo state (1: echo, 0: no echo).
int WebRtcAec_echo_state(AecCore* self); int WebRtcAec_echo_state(AecCore* self);
// Gets statistics of the echo metrics ERL, ERLE, A_NLP. // Gets statistics of the echo metrics ERL, ERLE, A_NLP.
void WebRtcAec_GetEchoStats(AecCore* self, Stats* erl, Stats* erle, void WebRtcAec_GetEchoStats(AecCore* self, Stats* erl, Stats* erle,
Stats* a_nlp); Stats* a_nlp);
#ifdef WEBRTC_AEC_DEBUG_DUMP #ifdef WEBRTC_AEC_DEBUG_DUMP
void* WebRtcAec_far_time_buf(AecCore* self); void* WebRtcAec_far_time_buf(AecCore* self);
#endif #endif
// Sets local configuration modes. // Sets local configuration modes.
void WebRtcAec_SetConfigCore(AecCore* self, int nlp_mode, int metrics_mode, void WebRtcAec_SetConfigCore(AecCore* self, int nlp_mode, int metrics_mode,
int delay_logging); int delay_logging);
// We now interpret delay correction to mean an extended filter length feature.
// We reuse the delay correction infrastructure to avoid changes through to
// libjingle. See details along with |DelayCorrection| in
// echo_cancellation_impl.h. Non-zero enables, zero disables.
void WebRtcAec_enable_delay_correction(AecCore* self, int enable);
// Returns non-zero if delay correction is enabled and zero if disabled.
int WebRtcAec_delay_correction_enabled(AecCore* self);
// Returns the current |system_delay|, i.e., the buffered difference between // Returns the current |system_delay|, i.e., the buffered difference between
// far-end and near-end. // far-end and near-end.
int WebRtcAec_system_delay(AecCore* self); int WebRtcAec_system_delay(AecCore* self);
// Sets the |system_delay| to |value|. Note that if the value is changed // Sets the |system_delay| to |value|. Note that if the value is changed
// improperly, there can be a performance regression. So it should be used with // improperly, there can be a performance regression. So it should be used with
// care. // care.

27
webrtc/modules/audio_processing/aec/aec_core_internal.h
View File

@ -19,15 +19,8 @@
#include "webrtc/modules/audio_processing/utility/ring_buffer.h" #include "webrtc/modules/audio_processing/utility/ring_buffer.h"
#include "webrtc/typedefs.h" #include "webrtc/typedefs.h"
// Number of partitions for the extended filter mode. The first one is an enum #define NR_PART 12 // Number of partitions in filter.
// to be used in array declarations, as it represents the maximum filter length. #define PREF_BAND_SIZE 24
enum { kExtendedNumPartitions = 32 };
static const int kNormalNumPartitions = 12;
// Extended filter adaptation parameters.
// TODO(ajm): No narrowband tuning yet.
static const float kExtendedMu = 0.4f;
static const float kExtendedErrorThreshold = 1.0e-6f;
typedef struct PowerLevel { typedef struct PowerLevel {
float sfrsum; float sfrsum;
@ -63,12 +56,11 @@ struct AecCore {
float dInitMinPow[PART_LEN1]; float dInitMinPow[PART_LEN1];
float *noisePow; float *noisePow;
float xfBuf[2][kExtendedNumPartitions * PART_LEN1]; // farend fft buffer float xfBuf[2][NR_PART * PART_LEN1]; // farend fft buffer
float wfBuf[2][kExtendedNumPartitions * PART_LEN1]; // filter fft float wfBuf[2][NR_PART * PART_LEN1]; // filter fft
complex_t sde[PART_LEN1]; // cross-psd of nearend and error complex_t sde[PART_LEN1]; // cross-psd of nearend and error
complex_t sxd[PART_LEN1]; // cross-psd of farend and nearend complex_t sxd[PART_LEN1]; // cross-psd of farend and nearend
// Farend windowed fft buffer. complex_t xfwBuf[NR_PART * PART_LEN1]; // farend windowed fft buffer
complex_t xfwBuf[kExtendedNumPartitions * PART_LEN1];
float sx[PART_LEN1], sd[PART_LEN1], se[PART_LEN1]; // far, near, error psd float sx[PART_LEN1], sd[PART_LEN1], se[PART_LEN1]; // far, near, error psd
float hNs[PART_LEN1]; float hNs[PART_LEN1];
@ -93,8 +85,8 @@ struct AecCore {
int sampFreq; int sampFreq;
uint32_t seed; uint32_t seed;
float normal_mu; // stepsize float mu; // stepsize
float normal_error_threshold; // error threshold float errThresh; // error threshold
int noiseEstCtr; int noiseEstCtr;
@ -120,11 +112,6 @@ struct AecCore {
void* delay_estimator_farend; void* delay_estimator_farend;
void* delay_estimator; void* delay_estimator;
// 1 = extended filter mode enabled, 0 = disabled.
int extended_filter_enabled;
// Runtime selection of number of filter partitions.
int num_partitions;
#ifdef WEBRTC_AEC_DEBUG_DUMP #ifdef WEBRTC_AEC_DEBUG_DUMP
RingBuffer* far_time_buf; RingBuffer* far_time_buf;
FILE *farFile; FILE *farFile;

54
webrtc/modules/audio_processing/aec/aec_core_sse2.c
View File

@ -34,14 +34,13 @@ __inline static float MulIm(float aRe, float aIm, float bRe, float bIm)
static void FilterFarSSE2(AecCore* aec, float yf[2][PART_LEN1]) static void FilterFarSSE2(AecCore* aec, float yf[2][PART_LEN1])
{ {
int i; int i;
const int num_partitions = aec->num_partitions; for (i = 0; i < NR_PART; i++) {
for (i = 0; i < num_partitions; i++) {
int j; int j;
int xPos = (i + aec->xfBufBlockPos) * PART_LEN1; int xPos = (i + aec->xfBufBlockPos) * PART_LEN1;
int pos = i * PART_LEN1; int pos = i * PART_LEN1;
// Check for wrap // Check for wrap
if (i + aec->xfBufBlockPos >= num_partitions) { if (i + aec->xfBufBlockPos >= NR_PART) {
xPos -= num_partitions*(PART_LEN1); xPos -= NR_PART*(PART_LEN1);
} }
// vectorized code (four at once) // vectorized code (four at once)
@ -76,11 +75,8 @@ static void FilterFarSSE2(AecCore* aec, float yf[2][PART_LEN1])
static void ScaleErrorSignalSSE2(AecCore* aec, float ef[2][PART_LEN1]) static void ScaleErrorSignalSSE2(AecCore* aec, float ef[2][PART_LEN1])
{ {
const __m128 k1e_10f = _mm_set1_ps(1e-10f); const __m128 k1e_10f = _mm_set1_ps(1e-10f);
const __m128 kMu = aec->extended_filter_enabled ? const __m128 kThresh = _mm_set1_ps(aec->errThresh);
_mm_set1_ps(kExtendedMu) : _mm_set1_ps(aec->normal_mu); const __m128 kMu = _mm_set1_ps(aec->mu);
const __m128 kThresh = aec->extended_filter_enabled ?
_mm_set1_ps(kExtendedErrorThreshold) :
_mm_set1_ps(aec->normal_error_threshold);
int i; int i;
// vectorized code (four at once) // vectorized code (four at once)
@ -114,39 +110,32 @@ static void ScaleErrorSignalSSE2(AecCore* aec, float ef[2][PART_LEN1])
_mm_storeu_ps(&ef[1][i], ef_im); _mm_storeu_ps(&ef[1][i], ef_im);
} }
// scalar code for the remaining items. // scalar code for the remaining items.
{ for (; i < (PART_LEN1); i++) {
const float mu = aec->extended_filter_enabled ? float absEf;
kExtendedMu : aec->normal_mu; ef[0][i] /= (aec->xPow[i] + 1e-10f);
const float error_threshold = aec->extended_filter_enabled ? ef[1][i] /= (aec->xPow[i] + 1e-10f);
kExtendedErrorThreshold : aec->normal_error_threshold; absEf = sqrtf(ef[0][i] * ef[0][i] + ef[1][i] * ef[1][i]);
for (; i < (PART_LEN1); i++) {
float abs_ef;
ef[0][i] /= (aec->xPow[i] + 1e-10f);
ef[1][i] /= (aec->xPow[i] + 1e-10f);
abs_ef = sqrtf(ef[0][i] * ef[0][i] + ef[1][i] * ef[1][i]);
if (abs_ef > error_threshold) { if (absEf > aec->errThresh) {
abs_ef = error_threshold / (abs_ef + 1e-10f); absEf = aec->errThresh / (absEf + 1e-10f);
ef[0][i] *= abs_ef; ef[0][i] *= absEf;
ef[1][i] *= abs_ef; ef[1][i] *= absEf;
}
// Stepsize factor
ef[0][i] *= mu;
ef[1][i] *= mu;
} }
// Stepsize factor
ef[0][i] *= aec->mu;
ef[1][i] *= aec->mu;
} }
} }
static void FilterAdaptationSSE2(AecCore* aec, float *fft, float ef[2][PART_LEN1]) { static void FilterAdaptationSSE2(AecCore* aec, float *fft, float ef[2][PART_LEN1]) {
int i, j; int i, j;
const int num_partitions = aec->num_partitions; for (i = 0; i < NR_PART; i++) {
for (i = 0; i < num_partitions; i++) {
int xPos = (i + aec->xfBufBlockPos)*(PART_LEN1); int xPos = (i + aec->xfBufBlockPos)*(PART_LEN1);
int pos = i * PART_LEN1; int pos = i * PART_LEN1;
// Check for wrap // Check for wrap
if (i + aec->xfBufBlockPos >= num_partitions) { if (i + aec->xfBufBlockPos >= NR_PART) {
xPos -= num_partitions * PART_LEN1; xPos -= NR_PART * PART_LEN1;
} }
// Process the whole array... // Process the whole array...
@ -424,4 +413,3 @@ void WebRtcAec_InitAec_SSE2(void) {
WebRtcAec_FilterAdaptation = FilterAdaptationSSE2; WebRtcAec_FilterAdaptation = FilterAdaptationSSE2;
WebRtcAec_OverdriveAndSuppress = OverdriveAndSuppressSSE2; WebRtcAec_OverdriveAndSuppress = OverdriveAndSuppressSSE2;
} }

539
webrtc/modules/audio_processing/aec/echo_cancellation.c
View File

@ -27,61 +27,6 @@
#include "webrtc/modules/audio_processing/utility/ring_buffer.h" #include "webrtc/modules/audio_processing/utility/ring_buffer.h"
#include "webrtc/typedefs.h" #include "webrtc/typedefs.h"
// Measured delays [ms]
// Device Chrome GTP
// MacBook Air 10
// MacBook Retina 10 100
// MacPro 30?
//
// Win7 Desktop 70 80?
// Win7 T430s 110
// Win8 T420s 70
//
// Daisy 50
// Pixel (w/ preproc?) 240
// Pixel (w/o preproc?) 110 110
// The extended filter mode gives us the flexibility to ignore the system's
// reported delays. We do this for platforms which we believe provide results
// which are incompatible with the AEC's expectations. Based on measurements
// (some provided above) we set a conservative (i.e. lower than measured)
// fixed delay.
//
// WEBRTC_UNTRUSTED_DELAY will only have an impact when |extended_filter_mode|
// is enabled. See the note along with |DelayCorrection| in
// echo_cancellation_impl.h for more details on the mode.
//
// Justification:
// Chromium/Mac: Here, the true latency is so low (~10-20 ms), that it plays
// havoc with the AEC's buffering. To avoid this, we set a fixed delay of 20 ms
// and then compensate by rewinding by 10 ms (in wideband) through
// kDelayDiffOffsetSamples. This trick does not seem to work for larger rewind
// values, but fortunately this is sufficient.
//
// Chromium/Linux(ChromeOS): The values we get on this platform don't correspond
// well to reality. The variance doesn't match the AEC's buffer changes, and the
// bulk values tend to be too low. However, the range across different hardware
// appears to be too large to choose a single value.
//
// GTP/Linux(ChromeOS): TBD, but for the moment we will trust the values.
#if defined(WEBRTC_CHROMIUM_BUILD) && defined(WEBRTC_MAC)
#define WEBRTC_UNTRUSTED_DELAY
#endif
#if defined(WEBRTC_MAC)
static const int kFixedDelayMs = 20;
static const int kDelayDiffOffsetSamples = -160;
#elif defined(WEBRTC_WIN)
static const int kFixedDelayMs = 50;
static const int kDelayDiffOffsetSamples = 0;
#else
// Essentially ChromeOS.
static const int kFixedDelayMs = 50;
static const int kDelayDiffOffsetSamples = 0;
#endif
static const int kMinTrustedDelayMs = 20;
static const int kMaxTrustedDelayMs = 500;
// Maximum length of resampled signal. Must be an integer multiple of frames // Maximum length of resampled signal. Must be an integer multiple of frames
// (ceil(1/(1 + MIN_SKEW)*2) + 1)*FRAME_LEN // (ceil(1/(1 + MIN_SKEW)*2) + 1)*FRAME_LEN
// The factor of 2 handles wb, and the + 1 is as a safety margin // The factor of 2 handles wb, and the + 1 is as a safety margin
@ -98,14 +43,7 @@ int webrtc_aec_instance_count = 0;
// Estimates delay to set the position of the far-end buffer read pointer // Estimates delay to set the position of the far-end buffer read pointer
// (controlled by knownDelay) // (controlled by knownDelay)
static void EstBufDelayNormal(aecpc_t *aecInst); static int EstBufDelay(aecpc_t *aecInst);
static void EstBufDelayExtended(aecpc_t *aecInst);
static int ProcessNormal(aecpc_t* self, const int16_t* near,
const int16_t* near_high, int16_t* out, int16_t* out_high,
int16_t num_samples, int16_t reported_delay_ms, int32_t skew);
static void ProcessExtended(aecpc_t* self, const int16_t* near,
const int16_t* near_high, int16_t* out, int16_t* out_high,
int16_t num_samples, int16_t reported_delay_ms, int32_t skew);
int32_t WebRtcAec_Create(void **aecInst) int32_t WebRtcAec_Create(void **aecInst)
{ {
@ -197,6 +135,10 @@ int32_t WebRtcAec_Init(void *aecInst, int32_t sampFreq, int32_t scSampFreq)
aecpc_t *aecpc = aecInst; aecpc_t *aecpc = aecInst;
AecConfig aecConfig; AecConfig aecConfig;
if (aecpc == NULL) {
return -1;
}
if (sampFreq != 8000 && sampFreq != 16000 && sampFreq != 32000) { if (sampFreq != 8000 && sampFreq != 16000 && sampFreq != 32000) {
aecpc->lastError = AEC_BAD_PARAMETER_ERROR; aecpc->lastError = AEC_BAD_PARAMETER_ERROR;
return -1; return -1;
@ -235,31 +177,31 @@ int32_t WebRtcAec_Init(void *aecInst, int32_t sampFreq, int32_t scSampFreq)
aecpc->splitSampFreq = sampFreq; aecpc->splitSampFreq = sampFreq;
} }
aecpc->skewFrCtr = 0;
aecpc->activity = 0;
aecpc->delayCtr = 0; aecpc->delayCtr = 0;
aecpc->sampFactor = (aecpc->scSampFreq * 1.0f) / aecpc->splitSampFreq;
// Sampling frequency multiplier (SWB is processed as 160 frame size).
aecpc->rate_factor = aecpc->splitSampFreq / 8000;
aecpc->sum = 0; aecpc->sum = 0;
aecpc->counter = 0; aecpc->counter = 0;
aecpc->checkBuffSize = 1; aecpc->checkBuffSize = 1;
aecpc->firstVal = 0; aecpc->firstVal = 0;
aecpc->startup_phase = 1; aecpc->ECstartup = 1;
aecpc->bufSizeStart = 0; aecpc->bufSizeStart = 0;
aecpc->checkBufSizeCtr = 0; aecpc->checkBufSizeCtr = 0;
aecpc->msInSndCardBuf = 0; aecpc->filtDelay = 0;
aecpc->filtDelay = -1; // -1 indicates an initialized state.
aecpc->timeForDelayChange = 0; aecpc->timeForDelayChange = 0;
aecpc->knownDelay = 0; aecpc->knownDelay = 0;
aecpc->lastDelayDiff = 0; aecpc->lastDelayDiff = 0;
aecpc->skewFrCtr = 0; aecpc->skew = 0;
aecpc->resample = kAecFalse; aecpc->resample = kAecFalse;
aecpc->highSkewCtr = 0; aecpc->highSkewCtr = 0;
aecpc->skew = 0; aecpc->sampFactor = (aecpc->scSampFreq * 1.0f) / aecpc->splitSampFreq;
aecpc->farend_started = 0; // Sampling frequency multiplier (SWB is processed as 160 frame size).
aecpc->rate_factor = aecpc->splitSampFreq / 8000;
// Default settings. // Default settings.
aecConfig.nlpMode = kAecNlpModerate; aecConfig.nlpMode = kAecNlpModerate;
@ -297,6 +239,10 @@ int32_t WebRtcAec_BufferFarend(void *aecInst, const int16_t *farend,
float skew; float skew;
int i = 0; int i = 0;
if (aecpc == NULL) {
return -1;
}
if (farend == NULL) { if (farend == NULL) {
aecpc->lastError = AEC_NULL_POINTER_ERROR; aecpc->lastError = AEC_NULL_POINTER_ERROR;
return -1; return -1;
@ -322,7 +268,6 @@ int32_t WebRtcAec_BufferFarend(void *aecInst, const int16_t *farend,
farend_ptr = (const int16_t*) newFarend; farend_ptr = (const int16_t*) newFarend;
} }
aecpc->farend_started = 1;
WebRtcAec_SetSystemDelay(aecpc->aec, WebRtcAec_system_delay(aecpc->aec) + WebRtcAec_SetSystemDelay(aecpc->aec, WebRtcAec_system_delay(aecpc->aec) +
newNrOfSamples); newNrOfSamples);
@ -366,6 +311,17 @@ int32_t WebRtcAec_Process(void *aecInst, const int16_t *nearend,
{ {
aecpc_t *aecpc = aecInst; aecpc_t *aecpc = aecInst;
int32_t retVal = 0; int32_t retVal = 0;
short i;
short nBlocks10ms;
short nFrames;
// Limit resampling to doubling/halving of signal
const float minSkewEst = -0.5f;
const float maxSkewEst = 1.0f;
if (aecpc == NULL) {
return -1;
}
if (nearend == NULL) { if (nearend == NULL) {
aecpc->lastError = AEC_NULL_POINTER_ERROR; aecpc->lastError = AEC_NULL_POINTER_ERROR;
return -1; return -1;
@ -398,21 +354,144 @@ int32_t WebRtcAec_Process(void *aecInst, const int16_t *nearend,
aecpc->lastError = AEC_BAD_PARAMETER_WARNING; aecpc->lastError = AEC_BAD_PARAMETER_WARNING;
retVal = -1; retVal = -1;
} }
else if (msInSndCardBuf > kMaxTrustedDelayMs) { else if (msInSndCardBuf > 500) {
// The clamping is now done in ProcessExtended/Normal(). msInSndCardBuf = 500;
aecpc->lastError = AEC_BAD_PARAMETER_WARNING; aecpc->lastError = AEC_BAD_PARAMETER_WARNING;
retVal = -1; retVal = -1;
} }
// TODO(andrew): we need to investigate if this +10 is really wanted.
msInSndCardBuf += 10;
aecpc->msInSndCardBuf = msInSndCardBuf;
// This returns the value of aec->extended_filter_enabled. if (aecpc->skewMode == kAecTrue) {
if (WebRtcAec_delay_correction_enabled(aecpc->aec)) { if (aecpc->skewFrCtr < 25) {
ProcessExtended(aecpc, nearend, nearendH, out, outH, nrOfSamples, aecpc->skewFrCtr++;
msInSndCardBuf, skew); }
else {
retVal = WebRtcAec_GetSkew(aecpc->resampler, skew, &aecpc->skew);
if (retVal == -1) {
aecpc->skew = 0;
aecpc->lastError = AEC_BAD_PARAMETER_WARNING;
}
aecpc->skew /= aecpc->sampFactor*nrOfSamples;
if (aecpc->skew < 1.0e-3 && aecpc->skew > -1.0e-3) {
aecpc->resample = kAecFalse;
}
else {
aecpc->resample = kAecTrue;
}
if (aecpc->skew < minSkewEst) {
aecpc->skew = minSkewEst;
}
else if (aecpc->skew > maxSkewEst) {
aecpc->skew = maxSkewEst;
}
#ifdef WEBRTC_AEC_DEBUG_DUMP
(void)fwrite(&aecpc->skew, sizeof(aecpc->skew), 1, aecpc->skewFile);
#endif
}
}
nFrames = nrOfSamples / FRAME_LEN;
nBlocks10ms = nFrames / aecpc->rate_factor;
if (aecpc->ECstartup) {
if (nearend != out) {
// Only needed if they don't already point to the same place.
memcpy(out, nearend, sizeof(short) * nrOfSamples);
}
// The AEC is in the start up mode
// AEC is disabled until the system delay is OK
// Mechanism to ensure that the system delay is reasonably stable.
if (aecpc->checkBuffSize) {
aecpc->checkBufSizeCtr++;
// Before we fill up the far-end buffer we require the system delay
// to be stable (+/-8 ms) compared to the first value. This
// comparison is made during the following 6 consecutive 10 ms
// blocks. If it seems to be stable then we start to fill up the
// far-end buffer.
if (aecpc->counter == 0) {
aecpc->firstVal = aecpc->msInSndCardBuf;
aecpc->sum = 0;
}
if (abs(aecpc->firstVal - aecpc->msInSndCardBuf) <
WEBRTC_SPL_MAX(0.2 * aecpc->msInSndCardBuf, sampMsNb)) {
aecpc->sum += aecpc->msInSndCardBuf;
aecpc->counter++;
}
else {
aecpc->counter = 0;
}
if (aecpc->counter * nBlocks10ms >= 6) {
// The far-end buffer size is determined in partitions of
// PART_LEN samples. Use 75% of the average value of the system
// delay as buffer size to start with.
aecpc->bufSizeStart = WEBRTC_SPL_MIN((3 * aecpc->sum *
aecpc->rate_factor * 8) / (4 * aecpc->counter * PART_LEN),
kMaxBufSizeStart);
// Buffer size has now been determined.
aecpc->checkBuffSize = 0;
}
if (aecpc->checkBufSizeCtr * nBlocks10ms > 50) {
// For really bad systems, don't disable the echo canceller for
// more than 0.5 sec.
aecpc->bufSizeStart = WEBRTC_SPL_MIN((aecpc->msInSndCardBuf *
aecpc->rate_factor * 3) / 40, kMaxBufSizeStart);
aecpc->checkBuffSize = 0;
}
}
// If |checkBuffSize| changed in the if-statement above.
if (!aecpc->checkBuffSize) {
// The system delay is now reasonably stable (or has been unstable
// for too long). When the far-end buffer is filled with
// approximately the same amount of data as reported by the system
// we end the startup phase.
int overhead_elements =
WebRtcAec_system_delay(aecpc->aec) / PART_LEN -
aecpc->bufSizeStart;
if (overhead_elements == 0) {
// Enable the AEC
aecpc->ECstartup = 0;
} else if (overhead_elements > 0) {
// TODO(bjornv): Do we need a check on how much we actually
// moved the read pointer? It should always be possible to move
// the pointer |overhead_elements| since we have only added data
// to the buffer and no delay compensation nor AEC processing
// has been done.
WebRtcAec_MoveFarReadPtr(aecpc->aec, overhead_elements);
// Enable the AEC
aecpc->ECstartup = 0;
}
}
} else { } else {
if (ProcessNormal(aecpc, nearend, nearendH, out, outH, nrOfSamples, // AEC is enabled.
msInSndCardBuf, skew) != 0) {
retVal = -1; EstBufDelay(aecpc);
}
// Note that 1 frame is supported for NB and 2 frames for WB.
for (i = 0; i < nFrames; i++) {
// Call the AEC.
WebRtcAec_ProcessFrame(aecpc->aec,
&nearend[FRAME_LEN * i],
&nearendH[FRAME_LEN * i],
aecpc->knownDelay,
&out[FRAME_LEN * i],
&outH[FRAME_LEN * i]);
// TODO(bjornv): Re-structure such that we don't have to pass
// |aecpc->knownDelay| as input. Change name to something like
// |system_buffer_diff|.
}
} }
#ifdef WEBRTC_AEC_DEBUG_DUMP #ifdef WEBRTC_AEC_DEBUG_DUMP
@ -430,6 +509,11 @@ int32_t WebRtcAec_Process(void *aecInst, const int16_t *nearend,
int WebRtcAec_set_config(void* handle, AecConfig config) { int WebRtcAec_set_config(void* handle, AecConfig config) {
aecpc_t* self = (aecpc_t*)handle; aecpc_t* self = (aecpc_t*)handle;
if (handle == NULL ) {
return -1;
}
if (self->initFlag != initCheck) { if (self->initFlag != initCheck) {
self->lastError = AEC_UNINITIALIZED_ERROR; self->lastError = AEC_UNINITIALIZED_ERROR;
return -1; return -1;
@ -464,6 +548,10 @@ int WebRtcAec_set_config(void* handle, AecConfig config) {
int WebRtcAec_get_echo_status(void* handle, int* status) { int WebRtcAec_get_echo_status(void* handle, int* status) {
aecpc_t* self = (aecpc_t*)handle; aecpc_t* self = (aecpc_t*)handle;
if (handle == NULL ) {
return -1;
}
if (status == NULL ) { if (status == NULL ) {
self->lastError = AEC_NULL_POINTER_ERROR; self->lastError = AEC_NULL_POINTER_ERROR;
return -1; return -1;
@ -577,6 +665,10 @@ int WebRtcAec_GetMetrics(void* handle, AecMetrics* metrics) {
int WebRtcAec_GetDelayMetrics(void* handle, int* median, int* std) { int WebRtcAec_GetDelayMetrics(void* handle, int* median, int* std) {
aecpc_t* self = handle; aecpc_t* self = handle;
if (handle == NULL) {
return -1;
}
if (median == NULL) { if (median == NULL) {
self->lastError = AEC_NULL_POINTER_ERROR; self->lastError = AEC_NULL_POINTER_ERROR;
return -1; return -1;
@ -601,6 +693,11 @@ int WebRtcAec_GetDelayMetrics(void* handle, int* median, int* std) {
int32_t WebRtcAec_get_error_code(void *aecInst) int32_t WebRtcAec_get_error_code(void *aecInst)
{ {
aecpc_t *aecpc = aecInst; aecpc_t *aecpc = aecInst;
if (aecpc == NULL) {
return -1;
}
return aecpc->lastError; return aecpc->lastError;
} }
@ -611,220 +708,7 @@ AecCore* WebRtcAec_aec_core(void* handle) {
return ((aecpc_t*) handle)->aec; return ((aecpc_t*) handle)->aec;
} }
static int ProcessNormal(aecpc_t *aecpc, const int16_t *nearend, static int EstBufDelay(aecpc_t* aecpc) {
const int16_t *nearendH, int16_t *out, int16_t *outH,
int16_t nrOfSamples, int16_t msInSndCardBuf,
int32_t skew) {
int retVal = 0;
short i;
short nBlocks10ms;
short nFrames;
// Limit resampling to doubling/halving of signal
const float minSkewEst = -0.5f;
const float maxSkewEst = 1.0f;
msInSndCardBuf = msInSndCardBuf > kMaxTrustedDelayMs ?
kMaxTrustedDelayMs : msInSndCardBuf;
// TODO(andrew): we need to investigate if this +10 is really wanted.
msInSndCardBuf += 10;
aecpc->msInSndCardBuf = msInSndCardBuf;
if (aecpc->skewMode == kAecTrue) {
if (aecpc->skewFrCtr < 25) {
aecpc->skewFrCtr++;
}
else {
retVal = WebRtcAec_GetSkew(aecpc->resampler, skew, &aecpc->skew);
if (retVal == -1) {
aecpc->skew = 0;
aecpc->lastError = AEC_BAD_PARAMETER_WARNING;
}
aecpc->skew /= aecpc->sampFactor*nrOfSamples;
if (aecpc->skew < 1.0e-3 && aecpc->skew > -1.0e-3) {
aecpc->resample = kAecFalse;
}
else {
aecpc->resample = kAecTrue;
}
if (aecpc->skew < minSkewEst) {
aecpc->skew = minSkewEst;
}
else if (aecpc->skew > maxSkewEst) {
aecpc->skew = maxSkewEst;
}
#ifdef WEBRTC_AEC_DEBUG_DUMP
(void)fwrite(&aecpc->skew, sizeof(aecpc->skew), 1, aecpc->skewFile);
#endif
}
}
nFrames = nrOfSamples / FRAME_LEN;
nBlocks10ms = nFrames / aecpc->rate_factor;
if (aecpc->startup_phase) {
// Only needed if they don't already point to the same place.
if (nearend != out) {
memcpy(out, nearend, sizeof(short) * nrOfSamples);
}
if (nearendH != outH) {
memcpy(outH, nearendH, sizeof(short) * nrOfSamples);
}
// The AEC is in the start up mode
// AEC is disabled until the system delay is OK
// Mechanism to ensure that the system delay is reasonably stable.
if (aecpc->checkBuffSize) {
aecpc->checkBufSizeCtr++;
// Before we fill up the far-end buffer we require the system delay
// to be stable (+/-8 ms) compared to the first value. This
// comparison is made during the following 6 consecutive 10 ms
// blocks. If it seems to be stable then we start to fill up the
// far-end buffer.
if (aecpc->counter == 0) {
aecpc->firstVal = aecpc->msInSndCardBuf;
aecpc->sum = 0;
}
if (abs(aecpc->firstVal - aecpc->msInSndCardBuf) <
WEBRTC_SPL_MAX(0.2 * aecpc->msInSndCardBuf, sampMsNb)) {
aecpc->sum += aecpc->msInSndCardBuf;
aecpc->counter++;
}
else {
aecpc->counter = 0;
}
if (aecpc->counter * nBlocks10ms >= 6) {
// The far-end buffer size is determined in partitions of
// PART_LEN samples. Use 75% of the average value of the system
// delay as buffer size to start with.
aecpc->bufSizeStart = WEBRTC_SPL_MIN((3 * aecpc->sum *
aecpc->rate_factor * 8) / (4 * aecpc->counter * PART_LEN),
kMaxBufSizeStart);
// Buffer size has now been determined.
aecpc->checkBuffSize = 0;
}
if (aecpc->checkBufSizeCtr * nBlocks10ms > 50) {
// For really bad systems, don't disable the echo canceller for
// more than 0.5 sec.
aecpc->bufSizeStart = WEBRTC_SPL_MIN((aecpc->msInSndCardBuf *
aecpc->rate_factor * 3) / 40, kMaxBufSizeStart);
aecpc->checkBuffSize = 0;
}
}
// If |checkBuffSize| changed in the if-statement above.
if (!aecpc->checkBuffSize) {
// The system delay is now reasonably stable (or has been unstable
// for too long). When the far-end buffer is filled with
// approximately the same amount of data as reported by the system
// we end the startup phase.
int overhead_elements =
WebRtcAec_system_delay(aecpc->aec) / PART_LEN - aecpc->bufSizeStart;
if (overhead_elements == 0) {
// Enable the AEC
aecpc->startup_phase = 0;
} else if (overhead_elements > 0) {
// TODO(bjornv): Do we need a check on how much we actually
// moved the read pointer? It should always be possible to move
// the pointer |overhead_elements| since we have only added data
// to the buffer and no delay compensation nor AEC processing
// has been done.
WebRtcAec_MoveFarReadPtr(aecpc->aec, overhead_elements);
// Enable the AEC
aecpc->startup_phase = 0;
}
}
} else {
// AEC is enabled.
EstBufDelayNormal(aecpc);
// Note that 1 frame is supported for NB and 2 frames for WB.
for (i = 0; i < nFrames; i++) {
// Call the AEC.
WebRtcAec_ProcessFrame(aecpc->aec,
&nearend[FRAME_LEN * i],
&nearendH[FRAME_LEN * i],
aecpc->knownDelay,
&out[FRAME_LEN * i],
&outH[FRAME_LEN * i]);
// TODO(bjornv): Re-structure such that we don't have to pass
// |aecpc->knownDelay| as input. Change name to something like
// |system_buffer_diff|.
}
}
return retVal;
}
static void ProcessExtended(aecpc_t* self, const int16_t* near,
const int16_t* near_high, int16_t* out, int16_t* out_high,
int16_t num_samples, int16_t reported_delay_ms, int32_t skew) {
int i;
const int num_frames = num_samples / FRAME_LEN;
#if defined(WEBRTC_UNTRUSTED_DELAY)
const int delay_diff_offset = kDelayDiffOffsetSamples;
reported_delay_ms = kFixedDelayMs;
#else
// This is the usual mode where we trust the reported system delay values.
const int delay_diff_offset = 0;
// Due to the longer filter, we no longer add 10 ms to the reported delay
// to reduce chance of non-causality. Instead we apply a minimum here to avoid
// issues with the read pointer jumping around needlessly.
reported_delay_ms = reported_delay_ms < kMinTrustedDelayMs ?
kMinTrustedDelayMs : reported_delay_ms;
// If the reported delay appears to be bogus, we attempt to recover by using
// the measured fixed delay values. We use >= here because higher layers
// may already clamp to this maximum value, and we would otherwise not
// detect it here.
reported_delay_ms = reported_delay_ms >= kMaxTrustedDelayMs ?
kFixedDelayMs : reported_delay_ms;
#endif
self->msInSndCardBuf = reported_delay_ms;
if (!self->farend_started) {
// Only needed if they don't already point to the same place.
if (near != out) {
memcpy(out, near, sizeof(short) * num_samples);
}
if (near_high != out_high) {
memcpy(out_high, near_high, sizeof(short) * num_samples);
}
return;
}
if (self->startup_phase) {
// In the extended mode, there isn't a startup "phase", just a special
// action on the first frame. In the trusted delay case, we'll take the
// current reported delay, unless it's less then our conservative
// measurement.
int startup_size_ms = reported_delay_ms < kFixedDelayMs ?
kFixedDelayMs : reported_delay_ms;
int overhead_elements = (WebRtcAec_system_delay(self->aec) -
startup_size_ms / 2 * self->rate_factor * 8) / PART_LEN;
WebRtcAec_MoveFarReadPtr(self->aec, overhead_elements);
self->startup_phase = 0;
}
EstBufDelayExtended(self);
for (i = 0; i < num_frames; ++i) {
// |delay_diff_offset| gives us the option to manually rewind the delay on
// very low delay platforms which can't be expressed purely through
// |reported_delay_ms|.
WebRtcAec_ProcessFrame(self->aec, &near[FRAME_LEN * i],
&near_high[FRAME_LEN * i], self->knownDelay + delay_diff_offset,
&out[FRAME_LEN * i], &out_high[FRAME_LEN * i]);
}
}
static void EstBufDelayNormal(aecpc_t* aecpc) {
int nSampSndCard = aecpc->msInSndCardBuf * sampMsNb * aecpc->rate_factor; int nSampSndCard = aecpc->msInSndCardBuf * sampMsNb * aecpc->rate_factor;
int current_delay = nSampSndCard - WebRtcAec_system_delay(aecpc->aec); int current_delay = nSampSndCard - WebRtcAec_system_delay(aecpc->aec);
int delay_difference = 0; int delay_difference = 0;
@ -848,11 +732,8 @@ static void EstBufDelayNormal(aecpc_t* aecpc) {
current_delay += WebRtcAec_MoveFarReadPtr(aecpc->aec, 1) * PART_LEN; current_delay += WebRtcAec_MoveFarReadPtr(aecpc->aec, 1) * PART_LEN;
} }
// We use -1 to signal an initialized state in the "extended" implementation;
// compensate for that.
aecpc->filtDelay = aecpc->filtDelay < 0 ? 0 : aecpc->filtDelay;
aecpc->filtDelay = WEBRTC_SPL_MAX(0, (short) (0.8 * aecpc->filtDelay + aecpc->filtDelay = WEBRTC_SPL_MAX(0, (short) (0.8 * aecpc->filtDelay +
0.2 * current_delay)); 0.2 * current_delay));
delay_difference = aecpc->filtDelay - aecpc->knownDelay; delay_difference = aecpc->filtDelay - aecpc->knownDelay;
if (delay_difference > 224) { if (delay_difference > 224) {
@ -875,58 +756,6 @@ static void EstBufDelayNormal(aecpc_t* aecpc) {
if (aecpc->timeForDelayChange > 25) { if (aecpc->timeForDelayChange > 25) {
aecpc->knownDelay = WEBRTC_SPL_MAX((int) aecpc->filtDelay - 160, 0); aecpc->knownDelay = WEBRTC_SPL_MAX((int) aecpc->filtDelay - 160, 0);
} }
}
return 0;
static void EstBufDelayExtended(aecpc_t* self) {
int reported_delay = self->msInSndCardBuf * sampMsNb * self->rate_factor;
int current_delay = reported_delay - WebRtcAec_system_delay(self->aec);
int delay_difference = 0;
// Before we proceed with the delay estimate filtering we:
// 1) Compensate for the frame that will be read.
// 2) Compensate for drift resampling.
// 3) Compensate for non-causality if needed, since the estimated delay can't
// be negative.
// 1) Compensating for the frame(s) that will be read/processed.
current_delay += FRAME_LEN * self->rate_factor;
// 2) Account for resampling frame delay.
if (self->skewMode == kAecTrue && self->resample == kAecTrue) {
current_delay -= kResamplingDelay;
}
// 3) Compensate for non-causality, if needed, by flushing two blocks.
if (current_delay < PART_LEN) {
current_delay += WebRtcAec_MoveFarReadPtr(self->aec, 2) * PART_LEN;
}
if (self->filtDelay == -1) {
self->filtDelay = WEBRTC_SPL_MAX(0, 0.5 * current_delay);
} else {
self->filtDelay = WEBRTC_SPL_MAX(0, (short) (0.95 * self->filtDelay +
0.05 * current_delay));
}
delay_difference = self->filtDelay - self->knownDelay;
if (delay_difference > 384) {
if (self->lastDelayDiff < 128) {
self->timeForDelayChange = 0;
} else {
self->timeForDelayChange++;
}
} else if (delay_difference < 128 && self->knownDelay > 0) {
if (self->lastDelayDiff > 384) {
self->timeForDelayChange = 0;
} else {
self->timeForDelayChange++;
}
} else {
self->timeForDelayChange = 0;
}
self->lastDelayDiff = delay_difference;
if (self->timeForDelayChange > 25) {
self->knownDelay = WEBRTC_SPL_MAX((int) self->filtDelay - 256, 0);
}
} }

6
webrtc/modules/audio_processing/aec/echo_cancellation_internal.h
View File

@ -20,6 +20,8 @@ typedef struct {
int splitSampFreq; int splitSampFreq;
int scSampFreq; int scSampFreq;
float sampFactor; // scSampRate / sampFreq float sampFactor; // scSampRate / sampFreq
short autoOnOff;
short activity;
short skewMode; short skewMode;
int bufSizeStart; int bufSizeStart;
int knownDelay; int knownDelay;
@ -37,7 +39,7 @@ typedef struct {
short msInSndCardBuf; short msInSndCardBuf;
short filtDelay; // Filtered delay estimate. short filtDelay; // Filtered delay estimate.
int timeForDelayChange; int timeForDelayChange;
int startup_phase; int ECstartup;
int checkBuffSize; int checkBuffSize;
short lastDelayDiff; short lastDelayDiff;
@ -60,8 +62,6 @@ typedef struct {
int lastError; int lastError;
int farend_started;
AecCore* aec; AecCore* aec;
} aecpc_t; } aecpc_t;

6
webrtc/modules/audio_processing/aec/system_delay_unittest.cc
View File

@ -128,7 +128,7 @@ void SystemDelayTest::RunStableStartup() {
for (; process_time_ms < kStableConvergenceMs; process_time_ms += 10) { for (; process_time_ms < kStableConvergenceMs; process_time_ms += 10) {
RenderAndCapture(kDeviceBufMs); RenderAndCapture(kDeviceBufMs);
buffer_size += samples_per_frame_; buffer_size += samples_per_frame_;
if (self_->startup_phase == 0) { if (self_->ECstartup == 0) {
// We have left the startup phase. // We have left the startup phase.
break; break;
} }
@ -222,7 +222,7 @@ TEST_F(SystemDelayTest, CorrectDelayAfterUnstableStartup) {
RenderAndCapture(reported_delay_ms); RenderAndCapture(reported_delay_ms);
buffer_size += samples_per_frame_; buffer_size += samples_per_frame_;
buffer_offset_ms = -buffer_offset_ms; buffer_offset_ms = -buffer_offset_ms;
if (self_->startup_phase == 0) { if (self_->ECstartup == 0) {
// We have left the startup phase. // We have left the startup phase.
break; break;
} }
@ -268,7 +268,7 @@ TEST_F(SystemDelayTest, CorrectDelayAfterStableBufferBuildUp) {
for (; process_time_ms <= kMaxConvergenceMs; process_time_ms += 10) { for (; process_time_ms <= kMaxConvergenceMs; process_time_ms += 10) {
RenderAndCapture(kDeviceBufMs); RenderAndCapture(kDeviceBufMs);
buffer_size += samples_per_frame_; buffer_size += samples_per_frame_;
if (self_->startup_phase == 0) { if (self_->ECstartup == 0) {
// We have left the startup phase. // We have left the startup phase.
break; break;
} }

16
webrtc/modules/audio_processing/echo_cancellation_impl.cc
View File

@ -13,14 +13,12 @@
#include <assert.h> #include <assert.h>
#include <string.h> #include <string.h>
extern "C" {
#include "webrtc/modules/audio_processing/aec/aec_core.h"
}
#include "webrtc/modules/audio_processing/aec/include/echo_cancellation.h"
#include "webrtc/modules/audio_processing/audio_buffer.h" #include "webrtc/modules/audio_processing/audio_buffer.h"
#include "webrtc/modules/audio_processing/audio_processing_impl.h" #include "webrtc/modules/audio_processing/audio_processing_impl.h"
#include "webrtc/system_wrappers/interface/critical_section_wrapper.h" #include "webrtc/system_wrappers/interface/critical_section_wrapper.h"
#include "webrtc/modules/audio_processing/aec/include/echo_cancellation.h"
namespace webrtc { namespace webrtc {
typedef void Handle; typedef void Handle;
@ -71,8 +69,7 @@ EchoCancellationImpl::EchoCancellationImpl(const AudioProcessingImpl* apm)
stream_drift_samples_(0), stream_drift_samples_(0),
was_stream_drift_set_(false), was_stream_drift_set_(false),
stream_has_echo_(false), stream_has_echo_(false),
delay_logging_enabled_(false), delay_logging_enabled_(false) {}
delay_correction_enabled_(false) {}
EchoCancellationImpl::~EchoCancellationImpl() {} EchoCancellationImpl::~EchoCancellationImpl() {}
@ -341,11 +338,6 @@ int EchoCancellationImpl::Initialize() {
return apm_->kNoError; return apm_->kNoError;
} }
void EchoCancellationImpl::SetExtraOptions(const Config& config) {
delay_correction_enabled_ = config.Get<DelayCorrection>().enabled;
Configure();
}
void* EchoCancellationImpl::CreateHandle() const { void* EchoCancellationImpl::CreateHandle() const {
Handle* handle = NULL; Handle* handle = NULL;
if (WebRtcAec_Create(&handle) != apm_->kNoError) { if (WebRtcAec_Create(&handle) != apm_->kNoError) {
@ -377,8 +369,6 @@ int EchoCancellationImpl::ConfigureHandle(void* handle) const {
config.skewMode = drift_compensation_enabled_; config.skewMode = drift_compensation_enabled_;
config.delay_logging = delay_logging_enabled_; config.delay_logging = delay_logging_enabled_;
WebRtcAec_enable_delay_correction(WebRtcAec_aec_core(
static_cast<Handle*>(handle)), delay_correction_enabled_ ? 1 : 0);
return WebRtcAec_set_config(static_cast<Handle*>(handle), config); return WebRtcAec_set_config(static_cast<Handle*>(handle), config);
} }

26
webrtc/modules/audio_processing/echo_cancellation_impl.h
View File

@ -15,30 +15,6 @@
namespace webrtc { namespace webrtc {
// Use to enable the delay correction feature. This now engages an extended
// filter mode in the AEC, along with robustness measures around the reported
// system delays. It comes with a significant increase in AEC complexity, but is
// much more robust to unreliable reported delays.
//
// Detailed changes to the algorithm:
// - The filter length is changed from 48 to 128 ms. This comes with tuning of
// several parameters: i) filter adaptation stepsize and error threshold;
// ii) non-linear processing smoothing and overdrive.
// - Option to ignore the reported delays on platforms which we deem
// sufficiently unreliable. See WEBRTC_UNTRUSTED_DELAY in echo_cancellation.c.
// - Faster startup times by removing the excessive "startup phase" processing
// of reported delays.
// - Much more conservative adjustments to the far-end read pointer. We smooth
// the delay difference more heavily, and back off from the difference more.
// Adjustments force a readaptation of the filter, so they should be avoided
// except when really necessary.
struct DelayCorrection {
DelayCorrection() : enabled(false) {}
DelayCorrection(bool enabled) : enabled(enabled) {}
bool enabled;
};
class AudioProcessingImpl; class AudioProcessingImpl;
class AudioBuffer; class AudioBuffer;
@ -58,7 +34,6 @@ class EchoCancellationImpl : public EchoCancellationImplWrapper {
// ProcessingComponent implementation. // ProcessingComponent implementation.
virtual int Initialize() OVERRIDE; virtual int Initialize() OVERRIDE;
virtual void SetExtraOptions(const Config& config) OVERRIDE;
private: private:
// EchoCancellation implementation. // EchoCancellation implementation.
@ -95,7 +70,6 @@ class EchoCancellationImpl : public EchoCancellationImplWrapper {
bool was_stream_drift_set_; bool was_stream_drift_set_;
bool stream_has_echo_; bool stream_has_echo_;
bool delay_logging_enabled_; bool delay_logging_enabled_;
bool delay_correction_enabled_;
}; };
} // namespace webrtc } // namespace webrtc

51
webrtc/modules/audio_processing/echo_cancellation_impl_unittest.cc
View File

@ -1,51 +0,0 @@
/*
* Copyright (c) 2013 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.
*/
#include "testing/gtest/include/gtest/gtest.h"
extern "C" {
#include "webrtc/modules/audio_processing/aec/aec_core.h"
}
#include "webrtc/modules/audio_processing/echo_cancellation_impl.h"
#include "webrtc/modules/audio_processing/include/audio_processing.h"
#include "webrtc/system_wrappers/interface/scoped_ptr.h"
namespace webrtc {
TEST(EchoCancellationInternalTest, DelayCorrection) {
scoped_ptr<AudioProcessing> ap(AudioProcessing::Create(0));
EXPECT_TRUE(ap->echo_cancellation()->aec_core() == NULL);
EXPECT_EQ(ap->kNoError, ap->echo_cancellation()->Enable(true));
EXPECT_TRUE(ap->echo_cancellation()->is_enabled());
AecCore* aec_core = ap->echo_cancellation()->aec_core();
ASSERT_TRUE(aec_core != NULL);
// Disabled by default.
EXPECT_EQ(0, WebRtcAec_delay_correction_enabled(aec_core));
Config config;
config.Set<DelayCorrection>(new DelayCorrection(true));
ap->SetExtraOptions(config);
EXPECT_EQ(1, WebRtcAec_delay_correction_enabled(aec_core));
// Retains setting after initialization.
EXPECT_EQ(ap->kNoError, ap->Initialize());
EXPECT_EQ(1, WebRtcAec_delay_correction_enabled(aec_core));
config.Set<DelayCorrection>(new DelayCorrection(false));
ap->SetExtraOptions(config);
EXPECT_EQ(0, WebRtcAec_delay_correction_enabled(aec_core));
// Retains setting after initialization.
EXPECT_EQ(ap->kNoError, ap->Initialize());
EXPECT_EQ(0, WebRtcAec_delay_correction_enabled(aec_core));
}
} // namespace webrtc

1
webrtc/modules/modules.gyp
View File

@ -145,7 +145,6 @@
'audio_coding/neteq4/mock/mock_payload_splitter.h', 'audio_coding/neteq4/mock/mock_payload_splitter.h',
'audio_processing/aec/system_delay_unittest.cc', 'audio_processing/aec/system_delay_unittest.cc',
'audio_processing/aec/echo_cancellation_unittest.cc', 'audio_processing/aec/echo_cancellation_unittest.cc',
'audio_processing/echo_cancellation_impl_unittest.cc',
'audio_processing/test/audio_processing_unittest.cc', 'audio_processing/test/audio_processing_unittest.cc',
'audio_processing/utility/delay_estimator_unittest.cc', 'audio_processing/utility/delay_estimator_unittest.cc',
'audio_processing/utility/ring_buffer_unittest.cc', 'audio_processing/utility/ring_buffer_unittest.cc',