Allow all supported audio formats in LoadAudio

[christian-byrne]

This changes LoadAudio to allow you to use all audio files supported by torchaudio.

This is the function torchaudio (2.3.1) is using to load audio files:

    def load(
        uri: Union[BinaryIO, str, os.PathLike],
        frame_offset: int = 0,
        num_frames: int = -1,
        normalize: bool = True,
        channels_first: bool = True,
        format: Optional[str] = None,
        buffer_size: int = 4096,
        backend: Optional[str] = None,
    ) -> Tuple[torch.Tensor, int]:
        """Load audio data from source.

        By default (``normalize=True``, ``channels_first=True``), this function returns Tensor with
        ``float32`` dtype, and the shape of `[channel, time]`.

        Note:
            The formats this function can handle depend on the availability of backends.
            Please use the following functions to fetch the supported formats.

            - FFmpeg: :py:func:`torchaudio.utils.ffmpeg_utils.get_audio_decoders`
            - Sox: :py:func:`torchaudio.utils.sox_utils.list_read_formats`
            - SoundFile: Refer to `the official document <https://pysoundfile.readthedocs.io/>`__.
        rest of docstring...
        """

Here is an alternative to hardocoding the supported formats:

• soundfile: soundfile.available_formats()
• sox: torchaudio.utils.sox_utils.list_read_formats()
• ffmpeg: torchaudio.utils.ffmpeg_utils.get_audio_decoders, but this returns a list of the codecs not the actual file extensions. To get the extensions you can use a subprocess to ffmpeg -formats and then parse the output.

Those are the techniques I used to generate the hardcoded lists. Sometimes the audio player widget doesnt support the format but everything else will still work (e.g., aiff).

[mcmonkey4eva]

Is this really the best way to do this? Is there no variable to read or something to get this list dynamically

[christian-byrne]

The methods I mentioned in the PR description can generate the lists dynamically. Here are the trade-offs as far as I can tell:

• For soundfile, you have to import soundfile, which may or may not match the version used by torchaudio.
• For ffmpeg, you need a way to map codecs to file extensions if using torchaudio.utils.ffmpeg_utils.get_audio_decoders(), or use subprocess.

The reason I hard-coded the lists was becuase I didn't think the supported formats in each library were volatile enough to warrant that overhead. What do you think?

[mcmonkey4eva]

maybe it'd be better to just not list at all, and instead just try to load blindly regardless of file ext, and just error if it errors

[huchenlei]

I don't think blindly load regardless of ext is good idea, as video/image file will also be recognized and blended into the selection choices.

[christian-byrne]

torchaudio can also load and extract audio from video files if ffmpeg is available, which is why video formats are included in the ffmpeg list.

I have been using the LoadAudio node with videos and can confirm it works.

[christian-byrne]

Implemented by #4054.