Division by zero in NoiseSuppression.
This change handles a special case in NoiseSuppression. The special
case was found by the AudioProcessing fuzzer.
A const copy of the capture audio stream is sent to
NoiseSuppression::AnalyzeCaptureAudio. Then audio undergoes processing
by e.g. the echo canceller. Then it's processed by
NoiseSuppression::ProcessCaptureAudio.
The special case is when the following conditions are all satisfied:
* All stream samples are constantly zero in the call to
AnalyzeCaptureAudio
* a processing component modifies it to be nonzero before the call to
ProcessCaptureAudio
* The array NoiseSuppressionC::magnPrevAnalyze is filled with
zeros. This holds after initialization.
In this case, there is a division by zero in WebRtcNs_ProcessCore. The
resulting NaN values pollute the output signal. They are only detected
several submodules later in the process chain. The NaN values cause
the EchoDetector to crash in debug mode.
There is special handling of the case when the signal is constant zero
in ProcessCore. This change avoids zero division by handling this
issue the same way.
Bug: chromium:803810 chromium:804634
Change-Id: I6d698dd0cd27e6d550b42085124300ce58533125
Reviewed-on: https://webrtc-review.googlesource.com/41282
Commit-Queue: Alex Loiko <aleloi@webrtc.org>
Reviewed-by: Henrik Lundin <henrik.lundin@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#21745}diff --git a/modules/audio_processing/ns/ns_core.c b/modules/audio_processing/ns/ns_core.c
index 3345b24..c87713a 100644
--- a/modules/audio_processing/ns/ns_core.c
+++ b/modules/audio_processing/ns/ns_core.c
@@ -1081,6 +1081,7 @@
// Depending on the duration of the inactive signal it takes a
// considerable amount of time for the system to learn what is noise and
// what is speech.
+ self->signalEnergy = 0;
return;
}
@@ -1239,7 +1240,7 @@
Windowing(self->window, self->dataBuf, self->anaLen, winData);
energy1 = Energy(winData, self->anaLen);
- if (energy1 == 0.0) {
+ if (energy1 == 0.0 || self->signalEnergy == 0) {
// Synthesize the special case of zero input.
// Read out fully processed segment.
for (i = self->windShift; i < self->blockLen + self->windShift; i++) {
@@ -1379,6 +1380,7 @@
sumMagnAnalyze += self->magnPrevAnalyze[i];
sumMagnProcess += self->magnPrevProcess[i];
}
+ RTC_DCHECK_GT(sumMagnAnalyze, 0);
avgProbSpeechHB *= sumMagnProcess / sumMagnAnalyze;
// Average filter gain from low band.
// Average over second half (i.e., 4->8kHz) of frequencies spectrum.
diff --git a/modules/audio_processing/ns/ns_core.h b/modules/audio_processing/ns/ns_core.h
index 97f76ba..990d363 100644
--- a/modules/audio_processing/ns/ns_core.h
+++ b/modules/audio_processing/ns/ns_core.h
@@ -110,7 +110,6 @@
float speechProb[HALF_ANAL_BLOCKL]; // Final speech/noise prob: prior + LRT.
// Buffering data for HB.
float dataBufHB[NUM_HIGH_BANDS_MAX][ANAL_BLOCKL_MAX];
-
} NoiseSuppressionC;
#ifdef __cplusplus
