Bug 228140 – [WebAudio] webm; properly trim frames according to the codec delay information

• Component: Web Audio
• Status: RESOLVED FIXED

[Jean-Yves Avenard [:jya]]

The webm metadata contains codec delay information, that specify how many audio frames should be trimmed at the start. This is used to properly deal with the Opus and Vorbis encoder delay behaviour that make them output silence for the first few packets.

Opus encoder delay is typically 45ms so we must drop that much from the start; otherwise that silence is audible and quite noticeable when you play the sound repeatedly.

[Radar WebKit Bug Importer]

<rdar://problem/80883882>

[Jean-Yves Avenard [:jya]]

Created attachment 434497 [details]
Patch

[Jean-Yves Avenard [:jya]]

Created attachment 434614 [details]
Patch

For the proper amount of frames to be returned, we need to explicitly tell the AudioConverter that we've reached EOS; this is done by returning noErr and set the number of decoded frames to 0

[Jean-Yves Avenard [:jya]]

Created attachment 434615 [details]
Patch

[Jean-Yves Avenard [:jya]]

Created attachment 434733 [details]
Patch

Fix style

[Eric Carlson]

Comment on attachment 434733 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=434733&action=review

> Source/WebCore/ChangeLog:11
> +        vorbis decoding return the right number of frames.

s/return/returns/

> Source/WebCore/platform/audio/cocoa/AudioFileReaderCocoa.cpp:266
> +    if (static_cast<char*>(data->mBuffers[0].mData) + data->mBuffers[0].mDataByteSize > userData->m_data + userData->m_dataSize)
> +        return kAudioConverterErr_UnspecifiedError;

This is serious enough that it might be worth logging an error.

> Source/WebCore/platform/audio/cocoa/AudioFileReaderCocoa.cpp:322
> +    PAL::AudioConverterSetProperty(converter, kAudioConverterPrimeInfo, sizeof(AudioConverterPrimeInfo), &primeInfo);

Nit: sizeof(primeInfo)

> Source/WebCore/platform/audio/cocoa/AudioFileReaderCocoa.cpp:377
> +        } while (status != kNoMoreDataErr && status != noErr);

It seems a little weird to use "noErr" to signal eos, why not define a specific error like we did for kNoMoreDataErr?

> Source/WebCore/platform/graphics/cocoa/AudioTrackPrivateWebM.cpp:93
> +    return MediaTime(m_track.codec_delay.value(), 1000000000);

It would be helpful to people that don't know about WebM who read this to use a named constant for 1000000000.

> Source/WebCore/platform/graphics/cocoa/SourceBufferParserWebM.cpp:1036
> +            m_callOnClientThreadCallback([this, protectedThis = makeRef(*this), trackID = trackData->track().track_uid.value(), padding = MediaTime(blockGroup.discard_padding.value(), 1000000000)]() {

Ditto.

[Jean-Yves Avenard [:jya]]

Created attachment 434814 [details]
Patch

Apply comments

[EWS]

ChangeLog entry in LayoutTests/ChangeLog contains OOPS!.

[Jean-Yves Avenard [:jya]]

Comment on attachment 434733 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=434733&action=review

> Source/WebCore/platform/audio/cocoa/AudioFileReaderCocoa.cpp:378
>      }

Unfortunately, this is required for the AudioConverter to work properly. We must return a number of packet of 0 and noErr for it to detect EOS. If we don't, any pending frames in the opus decoder will not be output and we'll be one packet (960 frames) short. This is an undocumented behaviour unfortunately and I only found out once I could check their code.
The AudioConverter will return the error the callback method returned. So we have to deal with noErr

[Jean-Yves Avenard [:jya]]

Created attachment 434816 [details]
Patch

Add reviewer to Changelogs

[EWS]

Committed r280584 (240206@main): <https://commits.webkit.org/240206@main>

All reviewed patches have been landed. Closing bug and clearing flags on attachment 434816 [details].