Fine tune the IntelligibilityEnhancer
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}
diff --git a/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc b/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc
index a10d10a..243bd12 100644
--- a/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc
+++ b/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc
@@ -31,11 +31,11 @@
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
diff --git a/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer_unittest.cc b/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer_unittest.cc
index 30035ab..852357e 100644
--- a/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer_unittest.cc
+++ b/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer_unittest.cc
@@ -409,7 +409,7 @@
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);
diff --git a/webrtc/modules/audio_processing/intelligibility/intelligibility_utils.cc b/webrtc/modules/audio_processing/intelligibility/intelligibility_utils.cc
index 3675f66..dfa912b 100644
--- a/webrtc/modules/audio_processing/intelligibility/intelligibility_utils.cc
+++ b/webrtc/modules/audio_processing/intelligibility/intelligibility_utils.cc
@@ -23,7 +23,7 @@
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|.
