libsox does not detect mp3 when extension is missing

[orlando-labs]

🐛 Bug

Sox IO backend doesn't allow to load file with explicit filetype while doesn't offer auto detection.

To Reproduce

Steps to reproduce the behavior:

>>> torchaudio.set_audio_backend("sox")
>>> torchaudio.load("/path/to/extensionless/file", filetype='mp3')
(tensor([[0., 0., 0.,  ..., 0., 0., 0.]]), 16000)
>>> torchaudio.set_audio_backend("sox_io")
>>> torchaudio.load("/path/to/extensionless/file", filetype='mp3')
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: load() got an unexpected keyword argument 'filetype'
>>> torchaudio.load("/path/to/extensionless/file")
formats: can't determine type of file `/path/to/extensionless/file'

Expected behavior

Expected an autodetect feature or allowance to pass an explicit format argument

Environment

torch/torchaudio build from source. master of v0.7.0 affected

[mthrok]

Hi @orlando-labs

Thanks for the feedback. I will look into it.

[mthrok]

Looks like this happens on specific formats, like mp3, but not on wav or flac.

[mthrok]

Hi @orlando-labs

Do you use other formats than mp3 that I should be looking into?

[mthrok]

Investigation Note:

On torchaudio side, error happens here, which was called here

This happens because sox_open_read had some internal error and returns null pointer.

The actual implementation of sox_open_read is found in open_read and the filetype member is set at

https://github.com/dmkrepo/libsox/blob/b9dd1a86e71bbd62221904e3e59dfaa9e5e72046/src/formats.c#L559

but will be reset if an error happens.

https://github.com/dmkrepo/libsox/blob/b9dd1a86e71bbd62221904e3e59dfaa9e5e72046/src/formats.c#L603

the error happens at

https://github.com/dmkrepo/libsox/blob/b9dd1a86e71bbd62221904e3e59dfaa9e5e72046/src/formats.c#L539

This is because auto_detect_format fails.

Interestingly, there is no mention to mp3 in auto_detect_format. And MP3 detection sorely relies on file extension.

Note: In formats.h, MP3 is added to the list of supported formats.
https://github.com/dmkrepo/libsox/blob/b9dd1a86e71bbd62221904e3e59dfaa9e5e72046/src/formats.h#L92-L93

Another way to fix is to add mp3 detection mechanism

• https://stackoverflow.com/a/11360513
• http://www.mp3-tech.org/programmer/frame_header.html

[orlando-labs]

Hi @orlando-labs

Do you use other formats than mp3 that I should be looking into?

Hi, currently mp3 only

[mthrok]

@orlando-labs

I landed #1042, which adds header-based mp3 detection.
It will become available starting from tomorrow's nightly build (which happens at around 7:30 AM in Eastern Time)
I believe it works as long as the mp3 file is well-formatted, but can you try and tell if it works?

[mthrok]

@vincentqb

(flagging as potentially BC-breaking if sox_io becomes the default in 0.8)

Oops I accidentally deleted the comment, but switching to sox_io IS BC-breaking so no need to flag something additionally.

[mthrok]

@orlando-labs

Did you have a chance to try out the nightly build?

[orlando-labs]

I'll try it in a day

[orlando-labs]

tried master torchaudio==0.8.0a0+3616961
and nightly torchaudio==0.8.0a0+cc04bbb
unfortunately, with no positive effect

Python 3.8.3 (default, Aug 31 2020, 16:03:14) 
[GCC 8.3.1 20191121 (Red Hat 8.3.1-5)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torchaudio

>>> torchaudio.set_audio_backend("sox_io")

>>> torchaudio.load("/dataset/audio//50/pz/50pz2wwojr60uetqsyp40mmb48mi", format="mp3")
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: load() got an unexpected keyword argument 'format'

>>> torchaudio.load("/dataset/audio//50/pz/50pz2wwojr60uetqsyp40mmb48mi")
formats: can't determine type of file `/bigdata/audio//50/pz/50pz2wwojr60uetqsyp40mmb48mi'
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib64/python3.8/site-packages/torchaudio-0.8.0a0+3616961-py3.8-linux-x86_64.egg/torchaudio/backend/sox_io_backend.py", line 105, in load
    signal = torch.ops.torchaudio.sox_io_load_audio_file(
RuntimeError: Error loading audio file: failed to open file.

>>> torchaudio.set_audio_backend("sox")
/usr/local/lib64/python3.8/site-packages/torchaudio-0.8.0a0+3616961-py3.8-linux-x86_64.egg/torchaudio/backend/utils.py:59: UserWarning: "sox" backend is depredated and will be removed in 0.9.0. Please use "sox_io" backend.
  warnings.warn(

>>> torchaudio.load("/dataset/audio/50/pz/50pz2wwojr60uetqsyp40mmb48mi", filetype="mp3")
(tensor([[0.0000, 0.0000, 0.0000,  ..., 0.0018, 0.0062, 0.0067]]), 16000)

The file itself:

file /dataset/audio/50/pz/50pz2wwojr60uetqsyp40mmb48mi
/dataset/audio/50/pz/50pz2wwojr60uetqsyp40mmb48mi: Audio file with ID3 version 2.4.0, contains:MPEG ADTS, layer III, v2,  24 kbps, 16 kHz, Monaural```