Backend "sox_io" is not one of available backends: ['soundfile'] even after set up of 'soundfile' on torchaudio

Question

I am working on some speech-recognition project, but I got error when I'be tried to load an audio.

RuntimeError: Backend "sox_io" is not one of available backends: ['soundfile'].

I've imported libries load_dataset, load_metric and Audio (per some tutorial)

from datasets import load_dataset, load_metric, Audio

That looks as 'sox_io' is for Linux and 'soundfile' for Windows. I am using Windows so I've changed my settings by:

torchaudio.set_audio_backend('soundfile')

But issue is still there - what I have to do to solve problem?

The above exception was the direct cause of the following exception:

ImportError: To support decoding 'mp3' audio files, please install 'sox'.

Answer 1

As of 0.12, mp3 is not supported by soundfile backend in TorchAudio. So there is no native way to load mp3 on Windows.

Recently pysoundfile added support for mp3, and there is a PR to add support to TorchAudio, https://github.com/pytorch/audio/pull/2712.