English Speech Audio Dataset with 1 Million Hours

1 million hours of English audio-text data was collected from the public internet by AllenAI. The dataset includes a variety of speaking styles, accents, and audio setups, supporting the training of the OLMoASR speech recognition models.

Use Cases

• Train English automatic speech recognition models on 1 million hours of audio with diverse accents.
• Fine-tune speech models on varied audio setups and speaking styles to improve robustness.
• Analyze the distribution of speaking styles and accents present in web-scale audio data.

Strengths

• Contains approximately 1 million hours of audio, providing substantial scale for model training.
• Encompasses a variety of speaking styles, accents, and audio setups, increasing data diversity.

Limitations

• Specific column structure, sample data, and file formats are not provided, complicating initial assessment.
• The dataset's curation method and potential biases from web-scale collection are not detailed.

Provenance

Source

AllenAI, collected from the public internet.

Collection Method

Web-scale collection, curated from the OLMoASR-Pool.

Freshness

Last updated March 2026.