People's Speech: 30,000+ Hours of Transcribed English Audio

MLCommons provides the People's Speech dataset, a collection of over 30,000 hours of transcribed English speech. This corpus is designed for training large-scale speech-to-text systems and is released under permissive licenses for both academic and commercial applications.

Use Cases

• Training automatic speech recognition (ASR) models using audio recordings and text transcriptions
• Acoustic modeling to account for diverse speaker profiles
• Benchmarking speech-to-text accuracy against machine-generated and crowdsourced labels

Strengths

• 30,000+ hours of transcribed audio
• Permissive CC-BY and CC-BY-SA licensing for commercial use
• Diverse speaker representation

Limitations

• Potential transcription noise due to machine-generated annotations
• Monolingual English focus limits multilingual application

Provenance

Source

MLCommons

Collection Method

Gathered from various sources with machine-generated and crowdsourced annotations

Freshness

Last updated November 2024