zqz/platform-external-webrtc
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fine tune the IntelligibilityEnhancer

Browse Source 
Label less chunks as speech, adapt slower and be more conservative with the maximum gain it can apply.

Review-Url: https://codereview.webrtc.org/2087623003
Cr-Commit-Position: refs/heads/master@{#13242}
This commit is contained in:
aluebs
2016-06-21 11:30:25 -07:00
committed by Commit bot
parent 435264a183
commit 7bd5f253bc
3 changed files with 6 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc
View File

@ -31,11 +31,11 @@ const float kClipFreqKhz = 0.2f;
const float kKbdAlpha = 1.5f;
const float kLambdaBot = -1.f; // Extreme values in bisection
const float kLambdaTop = -1e-5f; // search for lamda.
const float kVoiceProbabilityThreshold = 0.02f;
const float kVoiceProbabilityThreshold = 0.5f;
// Number of chunks after voice activity which is still considered speech.
const size_t kSpeechOffsetDelay = 80;
const float kDecayRate = 0.994f; // Power estimation decay rate.
const float kMaxRelativeGainChange = 0.006f;
const size_t kSpeechOffsetDelay = 10;
const float kDecayRate = 0.995f; // Power estimation decay rate.
const float kMaxRelativeGainChange = 0.005f;
const float kRho = 0.0004f; // Default production and interpretation SNR.
const float kPowerNormalizationFactor = 1.f / (1 << 30);
const float kMaxActiveSNR = 128.f; // 21dB

2
webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer_unittest.cc
View File

@ -409,7 +409,7 @@ TEST_F(IntelligibilityEnhancerTest, TestSolveForGains) {
TEST_F(IntelligibilityEnhancerTest, TestNoiseGainHasExpectedResult) {
const int kGainDB = 6;
const float kGainFactor = std::pow(10.f, kGainDB / 20.f);
const float kTolerance = 0.003f;
const float kTolerance = 0.007f;
std::vector<float> noise(kNumNoiseBins);
std::vector<float> noise_psd(kNumNoiseBins);
std::generate(noise.begin(), noise.end(), float_rand);

2
webrtc/modules/audio_processing/intelligibility/intelligibility_utils.cc
View File

@ -23,7 +23,7 @@ namespace intelligibility {
namespace {
const float kMinFactor = 0.01f;
const float kMaxFactor = 1000.f;
const float kMaxFactor = 100.f;
// Return |current| changed towards |target|, with the relative change being at
// most |limit|.