People's Speech

30,000-hour English ASR corpus assembled from diverse internet sources including radio broadcasts, court hearings, and conferences.

• Task: asr
• Languages: en
• Hours: 30000
• Domain: diverse (radio, court, conference, podcast)
• License: CC BY-SA 4.0
• Homepage: https://huggingface.co/datasets/MLCommons/peoples_speech
• Paper: https://arxiv.org/abs/2111.09344

Recommendation

A large permissively-licensed English corpus that adds domain diversity beyond audiobooks. Automatic labels vary in quality — apply the data-cleaning pipeline to filter low-confidence segments before training. Good complement to LibriSpeech/GigaSpeech in a multi-dataset training mix.

Getting the data

Obtain from the dataset homepage.

Available via HuggingFace datasets. The full 30,000 h "dirty" split has more noise; use the "clean" split or apply quality filtering for best ASR training results.

Suggested processing

A recommended VoxKitchen pipeline ships in the repository at voxkitchen/templates/pipelines/data-cleaning.yaml — run it with vkit docker run.