internal block size of the AEC differ from the frame
size in the AEC output.
Before this CL, this buffering was done using ringbuffers
as well as secondary internal AEC buffers that were stored
on the state. The internal buffers were redundant, and the
ringbuffers were so short that the benefit of using
ringbuffers were lost.
This CL addresses the above issues by replacing the
ringbuffers by linear buffers. This has the main advantage
of cleaner code but it should significantly less
computational complex.
Furthermore, as the complexity of the function where the
conversion to external and internal AEC frame sizes is done
increased significantly with the changes in this CL, the
CL also include refactoring the near-end buffer handling
to increase readability and reduce code repetition.
After the changes in this CL it is very clear that the
former buffering of the output was incorrectly done for
the first frames. This CL corrects that but in doing that
it breaks the bitexactness with the former code.
The bitexactness is, however, only broken for the first
1000 samples and it has been verified that for a test suite
the CL maintains bitexactness in the AEC output
after the first 1000 samples.
This CL is chained to the CL https://codereview.webrtc.org/2311833002/ and will be
followed by more CLs that refactor the other buffers
inside the AEC.
BUG=webrtc:5298, webrtc:6018
Review-Url: https://codereview.webrtc.org/2321483002
Cr-Commit-Position: refs/heads/master@{#14184}
functionalities doing sample-rate conversions:
-The implicit resampling done in the AudioBuffer CopyTo,
CopyFrom, InterleaveTo and DeinterleaveFrom methods.
-The multi-band splitting scheme.
The selection of rates in these have been difficult and
complicated, partly due to that the APM API which allows
for activating the APM submodules without notifying
the APM.
This CL adds functionality that for each capture frame
polls all submodules for whether they are active or not
and compares this against a cached result.
Furthermore, new functionality is added that based on the
results of the comparison do a reinitialization of the APM.
This has several advantages
-The code deciding on whether to analysis and synthesis is
needed for the bandsplitting can be much simplified and
centralized.
-The selection of the processing rate can be done such as
to avoid the implicit resampling that was in some cases
unnecessarily done.
-The optimization for whether an output copy is needed
that was done to improve performance due to the implicit
resampling is no longer needed, which simplifies the
code and makes it less error-prone in the sense that
is no longer neccessary to keep track of whether any
module has changed the signal.
Finally, it should be noted that the polling of the state
for all the submodules was done previously as well, but in
a less obvious and distributed manner.
BUG=webrtc:6181, webrtc:6220, webrtc:5298, webrtc:6296, webrtc:6298, webrtc:6297
Review-Url: https://codereview.webrtc.org/2304123002
Cr-Commit-Position: refs/heads/master@{#14175}
There were a series of changes in the calculation of echo metrics. There changes made the existing unittests lose, e.g., EXPECT_EQ become EXPECT_NEAR. It is good time to protect the echo calculation more strictly.
The change is not simply generating a new reference file and change EXPECT_NEAR to EXPECT_EQ. It strengthens the test as well. Main changes are
1. the old test only sample a metric at the end of processing, while the new test takes metrics during the call with a certain time interval. This gives a much stronger protection.
2. added protection of a newly added metric, called divergent_filter_fraction.
3. as said, use EXPECT_EQ (actually ASSERT_EQ) instead of EXPECT_NEAR as much as possible, even for float point values. This may be too restrictive. But it can be good to be restrictive at the beginning.
BUG=
Review-Url: https://codereview.webrtc.org/1969403003
Cr-Commit-Position: refs/heads/master@{#12871}
Added conditional updating of the statistics and the delay estimate so that
updates are only done when the farend is non-stationary.
The reason for this is that all the values that go into the updating of the
statistics, and that in turn are also used to update the delay, are frozen
when the farend signal is non-stationary. Therefore, when the farend signal
is silent (stationary), the last estimates present before the silent (stationary)
period began are used to continue to update the statistics. This is a problem as
the updating is done in a manner that assumes that the estimates continue
to be updated.
This CL conditions the updating based on stationarity instead of silence
as both are treated in the same manner in the delay agnostic AEC.
This makes sense theoretically as the delay agnostic AEC operates on
analyzing power deviations (in bands) from a slowly updated average power and
therefore for a stationary signal will have no such deviations to base its analysis
on.
BUG=webrtc:5875, chromium:576624
NOTRY=True
Review-Url: https://codereview.webrtc.org/1967033002
Cr-Commit-Position: refs/heads/master@{#12700}
The first approach landed here: https://codereview.webrtc.org/1773173002
But it was partially reverted, because it affected the AEC performance, here: https://codereview.webrtc.org/1867483003/
The main difference of this approach is that it doesn't use the 3-band splitting filter in the reverse stream, which seems to be the culprit of the AEC regression.
Also, the 2-band splitting filter has been used for the 32kHz case for a long time without any problem, and this is expanded in the CL to cover the 48kHz case as well.
BUG=webrtc:5725
TBR=tina.legrand@webrtc.org
Review URL: https://codereview.webrtc.org/1865633005
Cr-Commit-Position: refs/heads/master@{#12451}
were added in https://codereview.webrtc.org/1773173002.
The reason for the revert is that for some scenarios
that CL causes problems in the coherence estimate used
in the AEC, which in turn causes echo leakage.
The reason for not reverting the actual CL is that
it would cause subsequent CLs to be reverted as well.
Therefore the choice was made to in this CK
instead revert the effects of that CL.
With the changes in this CL, the behavior is bitexact
to what it was before the CL mentioned above.
TBR=aluebs@webrtc.org
BUG=webrtc:5725
Review URL: https://codereview.webrtc.org/1867483003
Cr-Commit-Position: refs/heads/master@{#12259}
Because of the Finch experiment, this will not affect Chrome's behaviour at all.
The SNRs in AudioProcessingTest.Formats were only increased to the next multiple of 5.
BUG=webrtc:3146
R=andrew@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/43359004
Cr-Commit-Position: refs/heads/master@{#9263}
To more easily determine if for example the AEC is not working properly one could monitor how often the estimated delay is out of bounds. With out of bounds we mean either being negative or too large, where both cases will break the AEC.
A new delay metric is added telling the user how often poor delay values were estimated. This is measured in percentage since last time the metrics were calculated.
All APIs have been updated with a third parameter with EchoCancellation::GetDelayMetrics() giving the option to exclude the new metric not to break existing code.
The new metric has been added to audio_processing_unittests with an additional protobuf member, and reference files accordingly updated.
voe_auto_test has not been updated to display the new metric.
BUG=4246
TESTED=audioproc on files
R=aluebs@webrtc.org, andrew@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/39739004
Cr-Commit-Position: refs/heads/master@{#8230}
git-svn-id: http://webrtc.googlecode.com/svn/trunk@8230 4adac7df-926f-26a2-2b94-8c16560cd09d
