Add docs for audio processing

[stevhliu]

This PR adds documentation for the Audio feature. It describes:

• The difference between loading path and audio, as well as use-cases/best practices for each of them.
• Resampling audio files with cast_column, and then calling ds[0]["audio"] to automatically decode and resample to the desired sampling rate.
• Resampling with map.

Preview here, let me know if I'm missing anything!

[lhoestq]

Cool thanks :)
also pinging @anton-l @patrickvonplaten @albertvillanova

[lhoestq]

Nice ! love it this way. I guess you can set this PR to "ready for review" ?

[patrickvonplaten]

That's a great document - thanks a lot for putting all of this together!

I left some tips on how the preprocessing transformers code example could be a bit simplified.

In short 99% of the use cases when Audio datasets is used for transformers is either:

a) A pretrained speech model is fine-tuned

or:

b) A fine-tuned speech model is evaluated / used in inference

For both a) and b) the feature_extractor is always defined. So we should always advocate to use AutoFeatureExtractor.from_pretrained(...) here IMO.

For a) the tokenizer is not defined and has to be created as described in the docs currently. For b) the tokenizer is also defined so that one can directly use Wav2Vec2Processor.from_pretrained(...)

Hope that helps a bit :-)

[lhoestq]

Looks all good to me now :)
Let us know if you have more comments or if it's ready to merge

[anton-l]

LGTM, great reference the transformers examples!

[lhoestq]

I guess we can merge this one now :)