Unsupervised Peoples Speech: 1 Million Hours of English Audio

MLCommons provides over one million hours of English audio extracted from Archive.org for unsupervised speech research. The collection features a diverse set of speakers and is available under CC-BY and CC-BY-SA licenses for academic and commercial applications. It was last updated in February 2025 to support large-scale speech model development.

Use Cases

• Self-supervised pre-training of Automatic Speech Recognition (ASR) models using raw audio files
• Audio Language Identification to distinguish English dialects or accents
• Unsupervised feature extraction for audio classification tasks

Strengths

• Over 1,000,000 hours of audio
• Permissive CC-BY and CC-BY-SA licensing for commercial use
• Diverse speaker representation from public archive sources

Limitations

• Lack of ground-truth transcripts necessitates unsupervised or self-supervised approaches
• Variable recording quality and background noise inherent in public archive sources

Provenance

Source

Archive.org

Collection Method

Extracted from public archives

Freshness

Last updated February 2025; source material from Archive.org is static once collected.

Geography

United States