YODAS v3
ESPnet (Carnegie Mellon University and collaborators)
Over 1.1 million hours of CC-BY web audio in more than 100 languages.
A web-crawled speech dataset of videos originally released under a CC-BY-3.0 license, with audio, transcripts, and metadata. One row is one audio segment. It supports speech recognition, text-to-speech, and audio representation learning, and uses videos distinct from earlier YODAS versions.
Access
- Cost
- FreeNo charge; very large audio files.
- License
- Open licenseCC BY 3.0
- Sign-up
- No sign-up
- How you get it
- API and Bulk downloadAudio, Parquet
What's in it
- One row is
- One row per audio segment with transcript
- Covers
- Global, 100+ languages
- Years
- Not stated
- Updated
- Irregular
- Size
- Over 1.1 million hours of audio
Quality
The CC-BY filter gives clearer rights, but transcripts are generated automatically and labels come from uploader metadata.
- Transcripts and language tags are automatic
- Uploader-chosen license is not always reliable
- Audio quality is uneven
Connects to
- YouTube-Commons: Transcript-only counterpart from the same CC-BY YouTube pool
- Emilia: Competing large web speech corpus
- Common Voice: Volunteer-read speech, cleaner but smaller
Questions it could answer
- Which languages have enough hours to train a recognizer from scratch?
- How does transcript quality vary by language?