LibriSpeech test-clean
Multimodal
The test-clean split of LibriSpeech, 2,620 utterances of read English audiobook speech from LibriVox drawn from the corpus's lower-difficulty speakers. Scores are corpus word error rate in percent after text normalization (the Hugging Face Open ASR Leaderboard applies the Whisper English normalizer); lower is better.
Top models (lower is better)
| Model | Score |
|---|---|
| LFM2.5-Audio-1.5B | 1.9 |