YODAS: 369,510 Hours of YouTube Speech and Captions

369,510 hours of speech audio and text captions sourced from YouTube, released by the espnet team in 2024. The dataset pairs audio utterances with either user-uploaded (manual) or system-generated (automatic) captions.

Use Cases

• Training automatic speech recognition (ASR) models using audio utterances and caption pairs
• Benchmarking speech-to-text alignment using the manual vs. automatic caption subsets
• Large-scale acoustic pre-training for speech foundation models

Strengths

• 369,510 hours of speech data
• Includes both manual and automatic caption subsets
• CC BY 3.0 license for open usage

Limitations

• Manual captions are user-uploaded and may not be human-verified
• Audio quality is subject to YouTube compression and varied recording environments
• Specific column names and file formats are not detailed in the primary metadata

Provenance

Source

YouTube via espnet

Collection Method

Scraped and extracted from YouTube

Freshness

Last updated June 2024.

Geography

Global