tree 62e3a650f611923489b1feb30d077ab67eb4db11
parent 600bdb4adc89a0f326c4d3ec37b840378e989f26
author Alex Loiko <aleloi@webrtc.org> 1516883273 +0100
committer Commit Bot <commit-bot@chromium.org> 1516887004 +0000

NaNs in Echo Canceller.

A coherence vector cohxd is computed in
WebRtcAec_ComputeCoherence. The coherence values should theoretically
be 0 <= x <= 1. Due to the way they are computed that is not always
the case.

The coherence values are used to update an error signal
estimate hNl in webrtc::EchoSuppression. 'hNl[i]' should contain an
error magnitude for frequency 'i'.

The error magnitudes are used as a basis for exponentiation. If a
magnitude is negative, the result is NaN.

The NaNs will then spread to the output signal.

This change caps the hNl values at 0. I considered capping the
coherence values at 1. The coherence values are calculated differently
for MIPS, NEON and SSE. Therefore it's simpler to cap the hNl values
instead.

The issue was found by the AudioProcessing fuzzer.

Bug: chromium:804634
Change-Id: I8ebaa441d77c3f79d9c194a850cb2b9eed1c2024
Reviewed-on: https://webrtc-review.googlesource.com/43740
Commit-Queue: Alex Loiko <aleloi@webrtc.org>
Reviewed-by: Per Åhgren <peah@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#21761}