ESB / Open ASR English average WER - release-reported composition
Multimodal
Average word error rate across the English test sets of the ESB / Hugging Face Open ASR Leaderboard suite (the leaderboard averages AMI, Earnings22, GigaSpeech, LibriSpeech clean and other, SPGISpeech, TED-LIUM and VoxPopuli) as reported by a model release. The release lists neither the averaged datasets nor the text normalizer, so the value is diagnostic rather than an official leaderboard average; lower is better.
Top models (lower is better)
| Model | Score |
|---|---|
| Phi-4-multimodal-instruct | 5.1 |
| LFM3-Audio-2B | 6.2 |
| LFM3-Omni-2B | 6.4 |
| LFM2.5-Audio-1.5B | 7.5 |