Emilia
Amphion (Chinese University of Hong Kong, Shenzhen)
100,000+ hours of in-the-wild multilingual speech, processed by an open pipeline.
A multilingual speech dataset for large-scale speech generation, built from web audio by the open Emilia-Pipe preprocessing tool. One row is one speech segment with its audio and transcript. The Emilia-Large expansion reaches over 200,000 hours.
Access
- Cost
- FreeFree but gated; accept terms on the Hub.
- License
- Provider's own termsCC BY 4.0 per the Hub tag; the card says the original 101k-hour set is CC BY-NC 4.0 while Emilia-Large combines other parts
- Sign-up
- Free account
- How you get it
- API and Bulk downloadAudio, JSON
What's in it
- One row is
- One row per speech segment
- Covers
- Global, with Chinese, English, Japanese, and other languages
- Years
- Not stated
- Updated
- Irregular
- Size
- Original 101,000 hours; Emilia-Large over 200,000 hours
Quality
Large and openly processed, but the recordings were scraped from the web and the card shows mixed license terms.
- License differs between the original and large versions
- Transcripts come from automatic recognition
- Scraped speech with unclear speaker consent
Connects to
- YODAS v3: Another web-crawled multilingual speech corpus, with 1.1 million hours
- FLEURS: Smaller, cleaner multilingual speech benchmark for evaluating models trained on corpora like this
Questions it could answer
- How do the share of hours and speaker counts vary by language?
- How much do automatic transcripts differ from a clean reference?