Add Librispeech ASR

[patrickvonplaten]

This PR adds the librispeech asr dataset: https://www.tensorflow.org/datasets/catalog/librispeech

There are 2 configs: "clean" and "other" whereas there are two "train" datasets for "clean", hence the name "train.100" and "train.360".

As suggested by @lhoestq, due to the enormous size of the dataset in .arrow format, the speech files are not directly prepared to a float32-array, but instead just the path to the array file is stored.

[lhoestq]

Nice thank you !

I added a few comments

[lhoestq]

Awesome thank you !

The dummy data are quite big but it was expected given that the raw files are flac files.
Given that the script doesn't even read the flac files I think we can remove them. Or maybe use empty flac files (see here for example). What do you think ?

We'll find a better solution to be able to have bigger dummy_data (max 1MB instead of a few KB, maybe using git LFS.

[patrickvonplaten]

Awesome thank you !

The dummy data are quite big but it was expected given that the raw files are flac files.
Given that the script doesn't even read the flac files I think we can remove them. Or maybe use empty flac files (see here for example). What do you think ?

We'll find a better solution to be able to have bigger dummy_data (max 1MB instead of a few KB, maybe using git LFS.

Hmm, I already made the dummy data as small as possible (a single flac filie per split only). I'd like to keep them at least to have complete dummy data and don't think 500KB for all datasets together is a problem (the long-range summarization datasets are similarly heavy). The moment we allow dummy data to be loaded directly for testing, we need the flac files IMO.

But I agree that longterm, we need a better solution for the dummy data (maybe stop hosting it on github to not make the repo too heavy)