Large English Speech Recognition Corpus With 30,000+ Hours

The People's Speech Dataset contains over 30,000 hours of transcribed English speech, licensed for academic and commercial use under CC-BY-SA and CC-BY 4.0. It was created by MLCommons to train speech-to-text systems and features a diverse set of speakers.

Use Cases

• Train automatic speech recognition models on over 30,000 hours of transcribed English audio.
• Analyze speaker diversity patterns within the large corpus of crowdsourced and machine-generated speech.
• Fine-tune speech-to-text systems using the permissively licensed CC-BY and CC-BY-SA audio and transcript pairs.
• Benchmark model performance on a monolingual English speech recognition task with US regional data.

Strengths

• Contains over 30,000 hours of transcribed speech, making it one of the world's largest English speech corpora.
• Available under permissive licenses (CC-BY-SA and CC-BY 4.0) for both academic and commercial usage.
• Includes speech from a diverse set of speakers, as noted in the dataset summary.
• Annotations were created through both crowdsourced and machine-generated methods, providing varied data sources.

Limitations

• The dataset is monolingual (English only), limiting its applicability for multilingual speech recognition tasks.
• Specific details on audio quality, speaker demographics, or transcription accuracy are not provided in the input.
• The input does not specify the time range of the collected speech data, which may affect model relevance.

Provenance

Source

MLCommons

Collection Method

Combination of crowdsourced and machine-generated annotations, sourced from original datasets.

Freshness

The dataset was last updated on 2024-08-25.

Geography

Primarily US region, as indicated by tags.