Let a dataset with an Audio column be flattened

[LeSingh1]

Dataset.flatten() fails on any dataset that holds a decodable Audio column:

from datasets import Dataset, Features, Audio, Value

features = Features({"audio": Audio(), "text": Value("string")})
ds = Dataset.from_dict({"audio": [None], "text": ["a"]}, features=features)
ds.flatten()

ValueError: Cannot flatten a decoded Audio feature.

The same dataset with an Image, Video or Pdf column flattens without complaint:

feature	Features.flatten()	Dataset.flatten()
Image()	{'audio': Image(...), 'text': ...}	['audio', 'text']
Video()	leaves it alone	['audio', 'text']
Pdf()	leaves it alone	['audio', 'text']
Audio()	ValueError	ValueError

So a single audio column makes the whole dataset impossible to flatten — even when the user only wants to expand an unrelated struct column, and even though flattening does not decode anything.

Cause

Image.flatten, Video.flatten and Pdf.flatten all share the same docstring and shape:

"If in the decodable state, return the feature itself, otherwise flatten the feature into a dictionary."

Audio.flatten is the odd one out:

def flatten(self):
    """If in the decodable state, raise an error, otherwise flatten the feature into a dictionary."""
    if self.decode:
        raise ValueError("Cannot flatten a decoded Audio feature.")

Features.flatten calls subfeature.flatten() on every feature that has the method, so the raise propagates out of Features.flatten() and Dataset.flatten().

Fix

Return the feature itself when it is decodable, exactly like the other three media features. Audio(decode=False) still flattens to {"bytes": Value("binary"), "path": Value("string")} as before.

No test asserted the old error — grep -rn "Cannot flatten" src/ tests/ matched only the line being removed — so nothing existing had to be adjusted.

Tests

tests/features/test_audio.py::test_audio_feature_flatten checks Audio().flatten(), Audio(decode=False).flatten(), Features.flatten() and Dataset.flatten(). It fails on main with the ValueError above, and needs no audio backend so it runs without torchcodec.

tests/features/, tests/test_formatting.py, tests/test_table.py, tests/test_arrow_dataset.py and tests/test_dataset_dict.py: 1134 passed, 0 failures.