Request to Enable Dataset Viewer

#8
by sanchit-gandhi HF staff - opened
SpeechColab org

Hey @lhoestq ! Could we please enable the dataset viewer for the GigaSpeech dataset, given it's one of the most popular ASR datasets for researchers and practitioners. Many thanks 🙌

cc @albertvillanova @severo

I added the exception :) Will let you know once it's deployed

Would it be useful to have a script (or a new method in datasets) that converts to WebDataset ?
It could reuse the code from https://github.com/huggingface/datasets/pull/5127

I refreshed it. It should be available soon.
Capture d’écran 2024-01-08 à 18.35.26.png

SpeechColab org

Amazing, thanks both! From an audio perspective, this conversion method would only be interesting if it offers a significant performance advantage for streaming. Since WebDatasets is not really otherwise used in audio, it wouldn't be worth it just to have 'compatibility' with training scripts that use WebDatasets, since these are few and far between.

Datasets with scripts won't work anymore unless users pass trust_remote_code=True in the next major release of datasets.
Let's encourage people to uploads datasets in AudioFolder, Parquet or WebDataset formats when possible, otherwise the audio datasets will be subject to security vulnerabilities in the future.

In terms of performance, WebDataset is superior to datasets scripts that use iter_archive because it offers a faster streaming start (it yields directly without having to load all the metadata in memory first), though it terms of throughput they are the same (both based on TAR streaming). Also converting to WebDataset would data loading parallelism.

Alternatively we could add TAR support for AudioFolder and have one metadata file per TAR archive but this will require some work.

Hi @sanchit-gandhi . Apart from just using a Python script, do you think other format could be useful for audio datasets? Thanks for your insights

Sign up or log in to comment