This change makes the transparent mode classifier consider filter
convergence also for microphone signals with very low volume. This is
in order to prevent entering transparent mode when there is low, but
audible, echo.
Furthermore, the suppression gain during transparent mode is restored
to avoid leaks when the echo is too low to be reliably detected by
filter convergence.
Bug: webrtc:10232, chromium:1140452
Change-Id: Idd4f40c4aee7c20baa444afaa5ec604eb65bcfd0
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/189786
Reviewed-by: Per Åhgren <peah@webrtc.org>
Commit-Queue: Gustaf Ullberg <gustaf@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#32461}
This change introduces a new Hidden Markov Model based classifier for
AEC3's 'transparent mode'. Transparent mode is used with
headsets/headphones where the speaker signal does not leak into the
microphone signal.
The current classifier suffers from two problems:
1. It sometimes takes a long time to enter transparent mode.
2. Sometimes transparent mode is left (and it once again takes a long
time to re-enter).
Both problems have a severe effect on AEC transparency.
The new classifier enters transparent mode quicker and is less likely
to exit transparent mode when there is no echo. This improves the
audio experience when using headset/headphones.
Another (minor) benefit of this change is that when transparent mode
is disabled no classifier is run (or even created) saving some memory
and CPU cycles.
Bug: webrtc:10232
Change-Id: I509af0e22b59463aeaead53c78c35be1e97fe8c3
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/184500
Commit-Queue: Gustaf Ullberg <gustaf@webrtc.org>
Reviewed-by: Per Åhgren <peah@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#32182}
