Update smoothed bitrate.

webrtc/common_audio/smoothing_filter.cc

Index: webrtc/common_audio/smoothing_filter.cc
diff --git a/webrtc/common_audio/smoothing_filter.cc b/webrtc/common_audio/smoothing_filter.cc
index ff79ab8799e9de332ff324152e97c43947837e6a..f3dafaf3f1a6b52acc11ff0fdcc64cd1afc8a1b8 100644
--- a/webrtc/common_audio/smoothing_filter.cc
+++ b/webrtc/common_audio/smoothing_filter.cc
@@ -50,13 +50,19 @@ void SmoothingFilterImpl::AddSample(float sample) {
   //   float alpha = pow(base, last_update_time_ms_ - now_ms);
   //   filtered_ = alpha * filtered_ + (1 - alpha) * sample;
   filter_.Apply(static_cast<float>(last_sample_time_ms_ - now_ms), sample);
+  last_sample_ = sample;
   last_sample_time_ms_ = now_ms;
 }
 
 rtc::Optional<float> SmoothingFilterImpl::GetAverage() const {
   float value = filter_.filtered();
-  return value == rtc::ExpFilter::kValueUndefined ? rtc::Optional<float>()
-                                                  : rtc::Optional<float>(value);
+  if (value == rtc::ExpFilter::kValueUndefined)
+    return rtc::Optional<float>();
+  int64_t now_ms = clock_->TimeInMilliseconds();
+  return rtc::Optional<float>(

[minyue-webrtc, 2016/12/01 11:32:50]

I like the idea of making the query of average time dependent. But is this
mathematically rigorous? It feels to me that we use exponential filter when we
add sample and use another filter when we query the average

+      value -
+      (value - last_sample_) *
+          (1 - exp((last_sample_time_ms_ - now_ms) / time_constant_ms_)));
 }
 
 void SmoothingFilterImpl::SetTimeConstantMs(int time_constant_ms) {