Sounds better according to a MUSHRA listening test.
The computational complexity is unaffected.
An empirically estimated gain was added to compensate for the attenuation introduced by the algorithm.
There are some TODOs, which I will address in follow up CLs.
It was tested in Hangouts without headphones and highest volume, to make sure it doesn't affect the AEC.
Review URL: https://codereview.webrtc.org/1378973003
Cr-Commit-Position: refs/heads/master@{#10308}
45daf7b26f