admin/webrtc_m130
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Calculating ERLE in AEC more properly.

Browse Source 
The audio level of the AEC's output level was calculated before overlapping add, and therefore, a compensation was needed. The compensation is multiplying the level by 2 since, before overlapping add, the level is roughly halved due to windowing.

This had to be that way because the level was calculated in frequency domain and the signal after overlapping add has only its time domain representation.

The level calculation has been updated to work on time domain signal and therefore the problem is not there any longer.

This CL is to put the calculation of the AEC output level after overlapping add and remove the compensation.

BUG=
R=peah@webrtc.org

Review URL: https://codereview.webrtc.org/1644133002 .

Cr-Commit-Position: refs/heads/master@{#11810}
This commit is contained in:
minyuel 2016-02-29 13:09:01 +01:00
parent fb45d170c0
commit 944744b25c
2 changed files with 8 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
webrtc/modules/audio_processing/aec/aec_core.cc
View File

@ -689,12 +689,11 @@ static void UpdateMetrics(AecCore* aec) {
// ERLE
// subtract noise power
suppressedEcho = 2 * (aec->nlpoutlevel.averagelevel -
safety * aec->nlpoutlevel.minlevel);
suppressedEcho = aec->nlpoutlevel.averagelevel -
safety * aec->nlpoutlevel.minlevel;
dtmp = 10 * static_cast<float>(log10(aec->nearlevel.averagelevel /
(2 * aec->nlpoutlevel.averagelevel) +
1e-10f));
aec->nlpoutlevel.averagelevel + 1e-10f));
dtmp2 = 10 * static_cast<float>(log10(echo / suppressedEcho + 1e-10f));
dtmp = dtmp2;
@ -1139,16 +1138,6 @@ static void EchoSuppression(AecCore* aec,
// Inverse error fft.
ScaledInverseFft(efw, fft, 2.0f, 1);
// 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, CalculatePower(fft, PART_LEN2));
}
// Overlap and add to obtain output.
for (i = 0; i < PART_LEN; i++) {
output[i] = (fft[i] * WebRtcAec_sqrtHanning[i] +
@ -1358,6 +1347,7 @@ static void ProcessBlock(AecCore* aec) {
EchoSuppression(aec, farend_ptr, echo_subtractor_output, output, outputH_ptr);
if (aec->metricsMode == 1) {
UpdateLevel(&aec->nlpoutlevel, CalculatePower(output, PART_LEN));
UpdateMetrics(aec);
}

8
webrtc/modules/audio_processing/test/audio_processing_unittest.cc
View File

@ -204,10 +204,10 @@ int16_t MaxAudioFrame(const AudioFrame& frame) {
#if defined(WEBRTC_AUDIOPROC_FLOAT_PROFILE)
void TestStats(const AudioProcessing::Statistic& test,
const audioproc::Test::Statistic& reference) {
EXPECT_NEAR(reference.instant(), test.instant, 1);
EXPECT_EQ(reference.average(), test.average);
EXPECT_EQ(reference.maximum(), test.maximum);
EXPECT_NEAR(reference.minimum(), test.minimum, 1);
EXPECT_NEAR(reference.instant(), test.instant, 2);
EXPECT_NEAR(reference.average(), test.average, 2);
EXPECT_NEAR(reference.maximum(), test.maximum, 2);
EXPECT_NEAR(reference.minimum(), test.minimum, 2);
}
void WriteStatsMessage(const AudioProcessing::Statistic& output,