Issue metadata
Sign in to add a comment
|
105.2% 'improvement' in browser_tests at 503499:510069 |
||||||||||||||||||||
Issue descriptionExcept it's not an improvement. The gain varies by a lot after the alert. Locally, SOX splits the recorded signal into more pieces than the reference signals causing the test to crash.
,
Oct 30 2017
The gain varies a lot both after the change and before. The test is brittle and often fails for me locally when I try to reproduce. Attaching prel results of investigation: * recorded audio before and after alert (I can't notice any big difference) * logs from 10 runs each before and after the change * waveform diagrams of the result before the alert (shows wildly varying gains during speech)
,
Oct 30 2017
I hacked Chrome to always start aecdumps. Result of digging: the fast gain change and resulting bad experience seems to be the fault of webrtc. AECdump attached.
,
Oct 30 2017
I think I have found the cause for the test brittleness, at least for my attempts of reproducing it locally: the AEC is on, and it changes the output in an unintended way. The test makes two tabs, records sound from one of them, and sends to the other. The APM of the first tab is fed the playout of the second. Since that playout is the processed input, there is strong correlation, and the AEC tries to compensate. For reference: here is how I run the test ./out/debug/browser_tests --gtest_filter=WebRtcAudioQualityBrowserTest.MANUAL_TestAutoGainControlOnLowAudio --single-process 2> log_file Here is how I run it offline with AEC switched off (the output is reasonable): ~/webrtc/src/out/debug/audioproc_f -dump_input /tmp/aecdump_dir/after_55Am40.aecdump -o running_offline.wav --aec 0 --agc 1 --agc_mode 2 --agc_limiter 1 --experimental_agc 1 --hpf 1 --ns 1 --ns_level 2 --ts 1 --ie 0
,
Nov 1 2017
The following revision refers to this bug: https://chromium.googlesource.com/chromium/src.git/+/5a81a99f7d5f6307e0117f210826790dcb220708 commit 5a81a99f7d5f6307e0117f210826790dcb220708 Author: Alex Loiko <aleloi@chromium.org> Date: Wed Nov 01 00:45:24 2017 AGC test shouldn't enable AEC. The test WebRtcAudioQualityBrowserTest.MANUAL_TestAutoGainControlOnLowAudio sets up a WebRTC audio-only one-way connection between two Chrome tabs. The first tab receives recorded audio from a fake microphone and sends it to the second tab. The second tab plays the audio which is then recorded. The WebRTC connection is negotiated with constraints "{audio: true}", which enable the AEC. The AEC affects audio levels, and changes in it potentially cause alerts such as this one: https://chromeperf.appspot.com/group_report?bug_id=779057 On some test settings (with '--single-process'), the AEC of one tab sees the playout stream (reverse) of the other tab, causing it to cancel the 'echo'. That leads to heavy speech distortion and often test failure (see the bug chromium:779057 ). This change disables AEC in the AGC tests. Bug: chromium:779057 Change-Id: Ie774e2faba41f057e47fb2ed4ee402217da8aaf3 Reviewed-on: https://chromium-review.googlesource.com/746641 Reviewed-by: Patrik Höglund <phoglund@chromium.org> Commit-Queue: Alex Loiko <aleloi@chromium.org> Cr-Commit-Position: refs/heads/master@{#513028} [modify] https://crrev.com/5a81a99f7d5f6307e0117f210826790dcb220708/chrome/browser/media/webrtc/webrtc_audio_quality_browsertest.cc
,
Nov 7 2017
|
|||||||||||||||||||||
►
Sign in to add a comment |
|||||||||||||||||||||
Comment 1 by 42576172...@developer.gserviceaccount.com
, Oct 27 2017