Automatic Speech Recognition on CHiME-4
4.38WER (Real Condition)Mega-ASR w/ router
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Mega-ASR w/ routerModel type=Our model, Routing=dynamic routing LoRA2026.05 | 4.38 | 5.62 | 5 | |
| Mega-ASRModel type=Our model2026.05 | 4.41 | 6.04 | 5.23 | |
| Qwen3-ASRModel type=Open-source2026.05 | 4.66 | 6.11 | 5.39 | |
| Step-Audio-2-miniModel type=Open-source2026.05 | 5.35 | 7.06 | 6.2 | |
| GPT-4o-trans.Model type=Closed-source2026.05 | 5.36 | 7.57 | 6.47 | |
| Whisper-L-v3Model type=Open-source2026.05 | 5.65 | 8.39 | 7.02 | |
| Kimi-AudioModel type=Open-source2026.05 | 5.66 | 7.46 | 6.56 | |
| Voxtral-MiniModel type=Open-source2026.05 | 6.01 | 9.04 | 7.53 | |
| Gemini3-FlashModel type=Closed-source2026.05 | 6.58 | 5.67 | 6.125 | |
| Parakeet-v3Model type=Open-source2026.05 | 6.61 | 8.82 | 7.72 | |
| Qwen2.5-OmniModel type=Open-source2026.05 | 6.62 | 8.13 | 7.37 | |
| Canary-1B-v2Model type=Open-source2026.05 | 7.19 | 9.73 | 8.46 | |
| Doubao-LLM ASRModel type=Closed-source2026.05 | 9.95 | 11.62 | 10.79 |