Add ASR task for SUPERB

[lewtun]

This PR starts building up the SUPERB benchmark by including the ASR task as described in the SUPERB paper and s3prl instructions.

Usage:

from datasets import load_dataset 

asr = load_dataset("superb", "asr")
# DatasetDict({
#     train: Dataset({
#         features: ['file', 'text', 'speaker_id', 'chapter_id', 'id'],
#         num_rows: 28539
#     })
#     validation: Dataset({
#         features: ['file', 'text', 'speaker_id', 'chapter_id', 'id'],
#         num_rows: 2703
#     })
#     test: Dataset({
#         features: ['file', 'text', 'speaker_id', 'chapter_id', 'id'],
#         num_rows: 2620
#     })
# })

I've used the GLUE benchmark as a guide for filling out the README.

To move fast during the evaluation PoC I propose to merge one task at a time, so we can continue building the training / evaluation framework in parallel.

Note: codewise this PR is ready for review - I'll add the missing YAML tags once #2620 is merged :)

[lewtun]

Wait until #2620 is merged before pushing the README tags in this PR

[albertvillanova]

#2620 is already merged into master. You can merge master into this branch.

[albertvillanova]

Thanks!

One question: aren't you adding task_templates to the _info method (and to the dataset_infos.json?

[lewtun]

Thanks!

One question: aren't you adding task_templates to the _info method (and to the dataset_infos.json?

great catch! i've now added the asr task template (along with a mapping from superb task -> template) and updated the dataset_infos.json :)

[albertvillanova]

Good!

I have a suggested refactoring... Tell me what you think! :)

[lewtun]

Good!

I have a suggested refactoring... Tell me what you think! :)

your approach is much more elegant - i've included your suggestions 🙏

[albertvillanova]

Thank you.