Tweak kDecayRate in the IntelligibilityEnhancer

webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc

Index: webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc
diff --git a/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc b/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc
index 04d36545216879d464065a62d773f3abb8d74dbd..334b7714cf2e345efceddfdc4588ba2a09d4698c 100644
--- a/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc
+++ b/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc
@@ -34,8 +34,7 @@ const float kLambdaTop = -1e-5f;      // search for lamda.
 const float kVoiceProbabilityThreshold = 0.02f;
 // Number of chunks after voice activity which is still considered speech.
 const size_t kSpeechOffsetDelay = 80;
-const float kDecayRate = 0.98f;              // Power estimation decay rate.
-const float kMaxRelativeGainChange = 0.04f;  // Maximum relative change in gain.
+const float kDecayRate = 0.994f;              // Power estimation decay rate.

[peah-webrtc, 2016/04/01 04:43:50]

I cannot really say that I yet understand the underlying algorithm but it seems
a bit strange to lock the map the relative_change_limit in GainApplier to the
decay rate.

I'm not saying it is wrong to do so, but they seem quite independent, and are
used in different places. Furthermore, this is a change from how it was before.

[hlundin-webrtc, 2016/04/01 08:36:38]

Acknowledged.

[aluebs-webrtc, 2016/04/01 17:33:35]

Well, the decay rate of the PSD estimations doesn't necessarily have to be
linked to the maximum gain change, but it kind of makes sense, since both are
linked to the speed the algorithm can adapt. And I find it elegant to only have
one constant to tweak the adaptation speed. But of course, I can go back to
having 2 different constants if you think it adds value.

 const float kRho = 0.0004f;  // Default production and interpretation SNR.
 
 // Returns dot product of vectors |a| and |b| with size |length|.
@@ -78,7 +77,7 @@ IntelligibilityEnhancer::IntelligibilityEnhancer(int sample_rate_hz,
       capture_filter_bank_(CreateErbBank(num_noise_bins)),
       render_filter_bank_(CreateErbBank(freqs_)),
       gains_eq_(bank_size_),
-      gain_applier_(freqs_, kMaxRelativeGainChange),
+      gain_applier_(freqs_, 1.f - kDecayRate),
       audio_s16_(chunk_length_),
       chunks_since_voice_(kSpeechOffsetDelay),
       is_speech_(false),