English Audio Transcripts with 3.4 Million Hours of Speech

OLMoASR-Pool contains approximately 3.4 million hours of audio and 18.8 million unique transcripts collected from the public internet. It was created by AllenAI to train English speech recognition models and includes a variety of speaking styles, accents, and audio setups.

Use Cases

• Train speech recognition models on 3.4 million hours of audio to transcribe diverse English accents and speaking styles.
• Analyze the relationship between audio characteristics and transcript accuracy across different audio setups.
• Fine-tune language models using the 18.8 million unique transcripts for tasks like text generation or summarization.

Strengths

• Contains approximately 3.4 million hours of audio data.
• Includes 18.8 million unique transcript IDs.
• Encompasses a variety of speaking styles, accents, and audio setups.

Limitations

• Specific column definitions, file formats, and audio quality metrics are not provided.
• The dataset's geographic and temporal coverage is not specified, which may limit analysis of regional or time-based trends.

Provenance

Source

AllenAI, collected from the public internet.

Collection Method

Web-scale collection from the public internet.

Freshness

Last updated on March 20, 2026.