Add KS task to SUPERB

[anton-l]

Add the KS (keyword spotting) task as described in the SUPERB paper.

• s3prl instructions
• s3prl implementation
• TFDS implementation

Some notable quirks:

• The dataset is originally single-archive (train+val+test all in one), but the test set has a "canonical" distribution in a separate archive, which is also used here (see _split_ks_files()).
• The _background_noise_/_silence_ audio files are much longer than others, so they require some sort of slicing for downstream training. I decided to leave the implementation of that up to the users, since TFDS and s3prl take different approaches (either slicing wavs deterministically, or subsampling randomly at runtime)

Related to #2619.

[lewtun]

thanks a lot for implementing this @anton-l !!

i won't have time to review this while i'm away, so happy for @albertvillanova and @patrickvonplaten to decide when to merge :)

[patrickvonplaten]

Looks very clean to me!

[albertvillanova]

Hi @anton-l, thanks a lot for the addition of this SUPERB task. You did an awesome job! ^^

Just some comments and suggested changes before we merge it into master.

[anton-l]

@albertvillanova thanks! Everything should be ready now :)

[albertvillanova]

Just one minor missing "id" and that's all! :)

[albertvillanova]

Thank you!!

[albertvillanova]

The background_noise/silence audio files are much longer than others, so they require some sort of slicing for downstream training. I decided to leave the implementation of that up to the users, since TFDS and s3prl take different approaches (either slicing wavs deterministically, or subsampling randomly at runtime)

@anton-l I was thinking that maybe we could give some hints in the dataset card (in a Usage section); something similar as for diarization: https://github.com/huggingface/datasets/blob/master/datasets/superb/README.md#example-of-usage
Note that for diarization it is not yet finished: we have to test it and then provide an end-to-end example: https://github.com/huggingface/datasets/pull/2661/files#r680224909

[anton-l]

@albertvillanova yeah, I'm not sure how to best implement it in pure datasets yet. It's something like this, where sample_noise() needs to be called from a pytorch batch collator or other framework-specific variant:

def map_to_array(example):
    import soundfile as sf

    speech_array, sample_rate = sf.read(example["file"])
    example["speech"] = speech_array
    example["sample_rate"] = sample_rate
    return example


def sample_noise(example):
    # Use a version of this function in a stateless way to extract random 1 sec slices of background noise
    # on each epoch
    from random import randint

    # _silence_ audios are longer than 1 sec
    if example["label"] == "_silence_":
        random_offset = randint(0, len(example["speech"]) - example["sample_rate"] - 1)
        example["speech"] = example["speech"][random_offset : random_offset + example["sample_rate"]]

    return example

[albertvillanova]

I see... Yes, not trivial indeed. Maybe for the moment you could add those functions above to the README (as it is the case for now in diarization)? What do you think?