Add a reliability term to the echo detector.

webrtc/modules/audio_processing/residual_echo_detector.cc

Index: webrtc/modules/audio_processing/residual_echo_detector.cc
diff --git a/webrtc/modules/audio_processing/residual_echo_detector.cc b/webrtc/modules/audio_processing/residual_echo_detector.cc
index 66b34ff10541c02550cf1f8ee65af319b3965aae..f72e3aedd8d54f04f157b589ebb0d22622820118 100644
--- a/webrtc/modules/audio_processing/residual_echo_detector.cc
+++ b/webrtc/modules/audio_processing/residual_echo_detector.cc
@@ -25,6 +25,7 @@ float Power(rtc::ArrayView<const float> input) {
 constexpr size_t kLookbackFrames = 650;
 // TODO(ivoc): Verify the size of this buffer.
 constexpr size_t kRenderBufferSize = 30;
+constexpr float kAlpha = 0.001f;
 
 }  // namespace
 
@@ -100,6 +101,8 @@ void ResidualEchoDetector::AnalyzeCaptureAudio(
     echo_likelihood_ = std::max(
         echo_likelihood_, covariances_[delay].normalized_cross_correlation());
   }
+  reliability_ = (1.0f - kAlpha) * reliability_ + kAlpha * 1.0f;

[hlundin-webrtc, 2016/11/16 09:46:30]

How long until this converges to close enough to 1.0? I'm wondering if it might
make sense to enclose those two lines in 
if (reliability_ < 1.0 - some_epsilon) {
  ...
}

[ivoc, 2016/11/16 12:30:38]

After about a minute the reliability is at 0.9975 (16 seconds to 0.8). I'm not
sure if a branch would be more or less efficient than 5 floating point
operations though.

[hlundin-webrtc, 2016/11/17 08:14:45]

Neither am I...

+  echo_likelihood_ *= reliability_;
   int echo_percentage = static_cast<int>(echo_likelihood_ * 100);
   RTC_HISTOGRAM_COUNTS("WebRTC.Audio.ResidualEchoDetector.EchoLikelihood",
                        echo_percentage, 0, 100, 100 /* number of bins */);