Fix normalization of noise estimate in NoiseSuppressor

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..7b25241e903c573c85a94752d6d197e90e3b5567 100644
--- a/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc
+++ b/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc
@@ -29,7 +29,7 @@ const int kWindowSizeMs = 16;
 const int kChunkSizeMs = 10;  // Size provided by APM.
 const float kClipFreqKhz = 0.2f;
 const float kKbdAlpha = 1.5f;
-const float kLambdaBot = -1.0f;      // Extreme values in bisection
+const float kLambdaBot = -1.f;      // Extreme values in bisection
 const float kLambdaTop = -1e-5f;      // search for lamda.
 const float kVoiceProbabilityThreshold = 0.02f;
 // Number of chunks after voice activity which is still considered speech.
@@ -37,6 +37,7 @@ 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 kRho = 0.0004f;  // Default production and interpretation SNR.
+const float kPowerNormalizationFactor = 1.f / (1 << 30);
 
 // Returns dot product of vectors |a| and |b| with size |length|.
 float DotProduct(const float* a, const float* b, size_t length) {
@@ -54,7 +55,8 @@ void MapToErbBands(const float* pow,
                    float* result) {
   for (size_t i = 0; i < filter_bank.size(); ++i) {
     RTC_DCHECK_GT(filter_bank[i].size(), 0u);
-    result[i] = DotProduct(filter_bank[i].data(), pow, filter_bank[i].size());
+    result[i] = kPowerNormalizationFactor *
+                DotProduct(filter_bank[i].data(), pow, filter_bank[i].size());
   }
 }
 
@@ -140,8 +142,8 @@ void IntelligibilityEnhancer::ProcessAudioBlock(
   MapToErbBands(noise_power.data(), capture_filter_bank_,
                 filtered_noise_pow_.data());
   SolveForGainsGivenLambda(kLambdaTop, start_freq_, gains_eq_.data());
-  const float power_target =
-      std::accumulate(clear_power.data(), clear_power.data() + freqs_, 0.f);
+  const float power_target = std::accumulate(

[peah-webrtc, 2016/03/31 20:50:57]

Is this related to the normalization of the noise?

[aluebs-webrtc, 2016/04/01 01:52:42]

Yes, I need to accumulate over the normalized signal. But also, it makes more
sense to accumulate on the ERB bands for the target power, since later the gains
are applied and the power to compare is calculated on that domain.

+      filtered_clear_pow_.data(), filtered_clear_pow_.data() + bank_size_, 0.f);
   const float power_top =
       DotProduct(gains_eq_.data(), filtered_clear_pow_.data(), bank_size_);
   SolveForGainsGivenLambda(kLambdaBot, start_freq_, gains_eq_.data());