Add new variance update option and unittests for intelligibility

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 3029e21619a981917f377afb78075bfd66def952..c258886f598294df613c7f1800e12dba1ba7a649 100644
--- a/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc
+++ b/webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc
@@ -27,26 +27,24 @@
 #include "webrtc/common_audio/vad/include/webrtc_vad.h"
 #include "webrtc/common_audio/window_generator.h"
 
-using std::complex;
-using std::max;
-using std::min;
-
 namespace webrtc {
 
-const int IntelligibilityEnhancer::kErbResolution = 2;
-const int IntelligibilityEnhancer::kWindowSizeMs = 2;
-const int IntelligibilityEnhancer::kChunkSizeMs = 10;  // Size provided by APM.
-const int IntelligibilityEnhancer::kAnalyzeRate = 800;
-const int IntelligibilityEnhancer::kVarianceRate = 2;
-const float IntelligibilityEnhancer::kClipFreq = 200.0f;
-const float IntelligibilityEnhancer::kConfigRho = 0.02f;
-const float IntelligibilityEnhancer::kKbdAlpha = 1.5f;
+namespace {
+
+const int kErbResolution = 2;
+const int kWindowSizeMs = 2;
+const int kChunkSizeMs = 10;  // Size provided by APM.
+const float kClipFreq = 200.0f;
+const float kConfigRho = 0.02f;  // Default production and interpretation SNR.
+const float kKbdAlpha = 1.5f;
+const float kLambdaBot = -1.0;      // Extreme values in bisection
+const float kLambdaTop = -10e-18f;  // search for lamda.
 
-// To disable gain update smoothing, set gain limit to be VERY high.
-// TODO(ekmeyerson): Add option to disable gain smoothing altogether
-// to avoid the extra computation.
-const float IntelligibilityEnhancer::kGainChangeLimit = 0.0125f;
+}  // namespace
 
+using std::complex;
+using std::max;
+using std::min;
 using VarianceType = intelligibility::VarianceArray::StepType;
 
 IntelligibilityEnhancer::TransformCallback::TransformCallback(
@@ -203,8 +201,6 @@ void IntelligibilityEnhancer::DispatchAudio(
 
 void IntelligibilityEnhancer::ProcessClearBlock(const complex<float>* in_block,
                                                 complex<float>* out_block) {
-  float power_target;
-
   if (block_count_ < 2) {
     memset(out_block, 0, freqs_ * sizeof(*out_block));
     ++block_count_;
@@ -216,8 +212,8 @@ void IntelligibilityEnhancer::ProcessClearBlock(const complex<float>* in_block,
   // based on experiments with different cutoffs.
   if (has_voice_low_ || true) {
     clear_variance_.Step(in_block, false);
-    power_target = std::accumulate(clear_variance_.variance(),
-                                   clear_variance_.variance() + freqs_, 0.0f);
+    const float power_target = std::accumulate(
+        clear_variance_.variance(), clear_variance_.variance() + freqs_, 0.0f);
 
     if (block_count_ % analysis_rate_ == analysis_rate_ - 1) {
       AnalyzeClearBlock(power_target);
@@ -239,26 +235,32 @@ void IntelligibilityEnhancer::AnalyzeClearBlock(float power_target) {
   FilterVariance(clear_variance_.variance(), filtered_clear_var_.get());
   FilterVariance(noise_variance_.variance(), filtered_noise_var_.get());
 
-  // Bisection search for optimal |lambda|
-
-  float lambda_bot = -1.0f, lambda_top = -10e-18f, lambda;
-  float power_bot, power_top, power;
-  SolveForGainsGivenLambda(lambda_top, start_freq_, gains_eq_.get());
-  power_top =
+  SolveForGainsGivenLambda(kLambdaTop, start_freq_, gains_eq_.get());
+  const float power_top =
       DotProduct(gains_eq_.get(), filtered_clear_var_.get(), bank_size_);
-  SolveForGainsGivenLambda(lambda_bot, start_freq_, gains_eq_.get());
-  power_bot =
+  SolveForGainsGivenLambda(kLambdaBot, start_freq_, gains_eq_.get());
+  const float power_bot =
       DotProduct(gains_eq_.get(), filtered_clear_var_.get(), bank_size_);
-  DCHECK(power_target >= power_bot && power_target <= power_top);
+  if (power_target >= power_bot && power_target <= power_top) {
+    SolveForLambda(power_target, power_bot, power_top);
+    UpdateErbGains();
+  }  // Else experiencing variance underflow, so do nothing.
+}
 
+void IntelligibilityEnhancer::SolveForLambda(float power_target,
+                                             float power_bot,
+                                             float power_top) {
+  float lambda_bot = kLambdaBot;
+  float lambda_top = kLambdaTop;
   float power_ratio = 2.0f;  // Ratio of achieved power to target power.
   const float kConvergeThresh = 0.001f;  // TODO(ekmeyerson): Find best values
   const int kMaxIters = 100;             // for these, based on experiments.
   int iters = 0;
   while (fabs(power_ratio - 1.0f) > kConvergeThresh && iters <= kMaxIters) {

[Andrew MacDonald, 2015/07/09 03:22:46]

std::fabs

[ekm, 2015/07/09 18:19:22]

Done.

-    lambda = lambda_bot + (lambda_top - lambda_bot) / 2.0f;
+    const float lambda = lambda_bot + (lambda_top - lambda_bot) / 2.0f;
     SolveForGainsGivenLambda(lambda, start_freq_, gains_eq_.get());
-    power = DotProduct(gains_eq_.get(), filtered_clear_var_.get(), bank_size_);
+    const float power =
+        DotProduct(gains_eq_.get(), filtered_clear_var_.get(), bank_size_);
     if (power < power_target) {
       lambda_bot = lambda;
     } else {
@@ -267,7 +269,9 @@ void IntelligibilityEnhancer::AnalyzeClearBlock(float power_target) {
     power_ratio = fabs(power / power_target);

[Andrew MacDonald, 2015/07/09 03:22:46]

To be sure we avoid a divide, outside the loop:
const float reciprocal_power_target = 1.f / power_target;
...
power_rate = std::fabs(power * reciprocal_power_target);

[ekm, 2015/07/09 18:19:22]

Done. Nice!

     ++iters;
   }
+}
 
+void IntelligibilityEnhancer::UpdateErbGains() {
   // (ERB gain) = filterbank' * (freq gain)
   float* gains = gain_applier_.target();
   for (int i = 0; i < freqs_; ++i) {