Support remote file systems for Audio

[OllieBroadhurst]

Feature request

Hi there!

It would be super cool if Audio(), and potentially other features, could read files from a remote file system.

Motivation

Large amounts of data is often stored in buckets. load_from_disk is able to retrieve data from cloud storage but to my knowledge actually copies the datasets across first, so if you're working off a system with smaller disk specs (like a VM), you can run out of space very quickly.

Your contribution

Something like this (for Google Cloud Platform in this instance):

from datasets import Dataset, Audio
import gcsfs

fs = gcsfs.GCSFileSystem()

list_of_audio_fp = {'audio': ['1', '2', '3']}
ds = Dataset.from_dict(list_of_audio_fp)

ds = ds.cast_column("audio", Audio(sampling_rate=16000, fs=fs))

Under the hood:

import librosa
from io import BytesIO

def load_audio(fp, sampling_rate=None, fs=None):
    if fs is not None:
        with fs.open(fp, 'rb') as f:
            arr, sr = librosa.load(BytesIO(f), sr=sampling_rate)
    else:
        # Perform existing io operations

Written from memory so some things could be wrong.

[OllieBroadhurst]

Just seen #5281