The improvement is mainly to extrapolate missing samples so that when querying the output, it assumes the input to continue even if no actual new samples are added.
The new implementation does not rely on base/exp_filter any longer. This is because it would be a bit cumbersome. base/exp_filter does pre-extrapolate, i.e., it assumes the all missing samples since the last sample equals the new sample.
BUG=webrtc:6443
Review-Url: https://codereview.webrtc.org/2551363002
Cr-Commit-Position: refs/heads/master@{#15575}
301fc4a712