Make several audio datasets streamable

[lhoestq]

Needs #3129 to be merged first

Make those audio datasets streamable:

• common_voice
• openslr
• vivos
• librispeech_asr (still has some issues to read FLAC) actually it's ok
• multilingual_librispeech (yet to be converted) TODO in a separate PR

[albertvillanova]

Reading FLAC (for librispeech_asr) works OK for me (soundfile version: 0.10.3):

In [2]: ds = load_dataset("datasets/librispeech_asr/librispeech_asr.py", "clean", streaming=True, split="train.100")

In [3]: item = next(iter(ds))

In [4]: item.keys()
Out[4]: dict_keys(['file', 'audio', 'text', 'speaker_id', 'chapter_id', 'id'])

In [5]: item["file"]
Out[5]: '374-180298-0000.flac'

In [6]: item["audio"].keys()
Out[6]: dict_keys(['path', 'array', 'sampling_rate'])

In [7]: item["audio"]["sampling_rate"]
Out[7]: 16000

In [8]: item["audio"]["path"]
Out[8]: '374-180298-0000.flac'

In [9]: item["audio"]["array"].shape
Out[9]: (232480,)

[lhoestq]

Oh cool ! I think this might have come from an issue with my local soundfile installation then

[lhoestq]

I'll do multilingual_librispeech in a separate PR since it requires the data to be in another format (in particular separate the train/dev/test splits in different files)

[patrickvonplaten]

@lhoestq @albertvillanova - think it would have been nice to have added a big message at the top stating that this is a breaking change and ping transformers people a bit more here.

[patrickvonplaten]

this is breaking no?