Encoding: support wav, flac etc. #630
Conversation
facebook-github-bot
added
the
CLA Signed
| @@ -92,14 +92,13 @@ AudioEncoder::AudioEncoder( | |||
| validateSampleRate(*avCodec, sampleRate); | |||
| avCodecContext_->sample_rate = sampleRate; | |||
|
|
|||
| // Note: This is the format of the **input** waveform. This doesn't determine | |||
| // the output. | |||
There was a problem hiding this comment.
My original comment was wrong: that's not the format of the input waveform. It's the format of the input AVFrame that we pass to avcodec_send_frame(). And it needs to be a format that the codec supports.
| avCodecContext_->frame_size > 0, | ||
| "frame_size is ", | ||
| avCodecContext_->frame_size, | ||
| ". Cannot encode. This should probably never happen?"); |
There was a problem hiding this comment.
This won't always be non-zero, see below.
There was a problem hiding this comment.
Interesting - it might be worth noting why from the docs (https://ffmpeg.org/doxygen/6.0/structAVCodecContext.html#aec57f0d859a6df8b479cd93ca3a44a33, which I admit to not understanding) when we turn 0 into our default.
| # Check that decode(encode(samples)) == samples on lossless formats | ||
|
|
||
| if get_ffmpeg_major_version() == 4 and output_format == "wav": | ||
| pytest.skip("Swresample with FFmpeg 4 doesn't work on wav files") |
There was a problem hiding this comment.
FYI, we're hitting this:
torchcodec/src/torchcodec/_core/FFMPEGCommon.cpp
Lines 159 to 166 in 12cdaa8
This PR adds support for encoding formats where the encoder doesn't natively support FLTP, which is the sample format of the input waveform. We use
swresampleto convert the input FLTP AVFrames into a sample format that the encoder support.