Add task template for automatic speech recognition

[lewtun]

This PR adds a task template for automatic speech recognition. In this task, the input is a path to an audio file which the model consumes to produce a transcription.

Usage:

from datasets import load_dataset
from datasets.tasks import AutomaticSpeechRecognition

ds = load_dataset("timit_asr", split="train[:10]")
# Dataset({
#     features: ['file', 'text', 'phonetic_detail', 'word_detail', 'dialect_region', 'sentence_type', 'speaker_id', 'id'],
#     num_rows: 10
# })

task = AutomaticSpeechRecognition(audio_file_column="file", transcription_column="text")
ds.prepare_for_task(task)
# Dataset({
#     features: ['audio_file', 'transcription'],
#     num_rows: 10
# })

[lhoestq]

Thanks for adding this template :)

Note that in the future we'll have an Audio feature type that will probably have additional parameters (like ClassLabel) such as the sampling rate or the audio format.

[lewtun]

good to know!

[lhoestq]

don't hesitate to move it outside of the BaseDatasetTest class, as mentioned in #2529

[lewtun]

yep, will do this here as well!

[SBrandeis]

Thank you @lewtun !

[SBrandeis]

Maybe audio_file_path would be more explicit on what the column represent ?

[SBrandeis]

Also, paths are not portable between machines a priori. This is probably good enough for now, but at some point, we'll need to replace the Value("string") with an Audio or Signal feature!

[lewtun]

@SBrandeis @lhoestq i've integrated your suggestions, so this is ready for another review :)

[lhoestq]

Merging if it's good for you @lewtun :)