SignalInfo in new AudioMetaData

Hi,
We're using sox_io backend for torchaudio. New version of the AudioMetaData class returned by the info method doesn't have all the data I need for analysis.
info method in the deprecated sox backend used to return sox_signalinfo_t and sox_encodinginfo_t that had much more detailed information about encoding, bit depth etc.
What should I use to get the info about audio file's encoding, bit depth and other details?

[vincentqb]

(since this is part of the BC-breaking change discussed in #903, cc @mthrok)

[mthrok]

Hi @RimaOlson

Currently there is no way to get the other information using "sox_io" backend.
We made this decision because, exposing the internal of libsox makes it very difficult to maintain the backends across the different OS.

Having said that, could you give information of what type of attributes you are using as a form of feature request?
From there I can come up with a way to make it available again.

@mthrok we use following attributes to ensure proper encoding of audio files before feeding them into our models:

sox_signalinfo_t.length
sox_encodinginfo_t.encoding
sox_encodinginfo_t.bits_per_sample

[mthrok]

@mthrok we use following attributes to ensure proper encoding of audio files before feeding them into our models:

sox_signalinfo_t.length
sox_encodinginfo_t.encoding
sox_encodinginfo_t.bits_per_sample

Thanks @RimaOlson

1. Could you tell, which value do you use for encoding?
   For the available values of encoding, you can see https://fossies.org/dox/sox-14.4.2/sox_8h.html#ab0a659b0258d1989c751ba5415e6a4cd

2. length is the number of samples (across channels) so you should be able to get the number with num_channels * num_frames. Let me know if this is not the case.

1. encoding-> 1 (SOX_ENCODING_SIGN2) and encoding-> 9 (SOX_ENCODING_ULAW )
2. So num_frames is the same as sox_signalinfo_t.length for mono files?

[mthrok]

@RimaOlson

1. encoding-> 1 (SOX_ENCODING_SIGN2) and encoding-> 9 (SOX_ENCODING_ULAW )

Thanks

2. So num_frames is the same as sox_signalinfo_t.length for mono files?

Yes.

[mthrok]

@RimaOlson One more question about encoding. What formats (extensions) does your application expect?
I think there are formats other than WAV that are mapped to SOX_ENCODING_SIGN2, and would like to get a clearer vision if possible.

@mthrok we use wav files

[mthrok]

Note to self: this is the corresponding encoding + bit depth part in PySoundFile

https://github.com/bastibe/python-soundfile/blob/744efb4b01abc72498a96b09115b42a4cabd85e4/soundfile.py#L38-L94

[mthrok]

Note2: It is ideal that the mapping we create from sox_encoding_t to our custom type is generic enough to handle ffmpeg's ffmpeg's AVCodecID enum.

[mthrok]

Hi @RimaOlson

Can you illustrate what your if statement looks like when you use sox_encodinginfo_t.encoding?
I am wondering what the new data structure should look like. Strongly typed ENUM type or just string would do.

@mthrok it will be something like this:
 

   switcher = {
        0: "SOX_ENCODING_UNKNOWN ",
        1: "SOX_ENCODING_SIGN2",
        2: "SOX_ENCODING_UNSIGNED",
    …
    }

QQ: does torchaudio.load automatically transcode to linear by default?