LibriSpeech ASR Corpus
OpenSLR (Vassil Panayotov, Daniel Povey, and colleagues)
1,000 hours of volunteers reading public-domain audiobooks aloud, the speech recognition yardstick.
English read speech from LibriVox audiobook recordings, segmented and aligned to the matching Project Gutenberg texts. Split into 'clean' and 'other' (harder) training and test sets, it is the most common benchmark for automatic speech recognition, and most papers report word error rates on test-clean and test-other.
Access
- Cost
- FreeFree download.
- License
- Open licenseCC BY 4.0
- Sign-up
- No sign-up
- How you get it
- Bulk downloadFLAC, TXT
What's in it
- One row is
- One audio utterance (a few seconds) with its transcript and speaker ID
- Covers
- English-speaking volunteer readers worldwide
- Years
- Released 2015
- Updated
- One-time release
- Size
- About 1,000 hours at 16 kHz
Quality
Clean, carefully aligned, and universally used; read audiobook speech is far easier than conversation.
- Read speech, not spontaneous conversation
- Readers skew toward American English
- Top systems are near saturation on test-clean
Connects to
- Project Gutenberg: Source texts for the transcripts
- Switchboard-1 Release 2: Spontaneous telephone conversation, a harder recognition test
- Mozilla Common Voice: Crowdsourced read speech in many more languages and accents
Questions it could answer
- How much does word error rate rise between test-clean and test-other?
- Do recognition errors vary with speaker gender in LibriSpeech?