Hypothesis-mode word alignment on Buckeye 5h (test)
93.1Reference Match RateCanary-Qwen
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Canary-QwenModel Type=Speech LLM, Align method=Gradients2026.07 | 93.1 | 11.4 | 95 | 17.1 | |
| Canary-QwenModel Type=Speech LLM, Align method=Self-att.2026.07 | 93.1 | 11.4 | 212 | 3.9 | |
| Phi-4-MMModel Type=Speech LLM, Align method=Gradients2026.07 | 92.8 | 11.7 | 102 | 19 | |
| Phi-4-MMModel Type=Speech LLM, Align method=Self-att.2026.07 | 92.8 | 11.7 | 80 | 16 | |
| Crisper-WhisperModel Type=AED, Align method=Official2026.07 | 92.6 | 12.2 | 46 | 43.5 | |
| Parakeet CTCModel Type=CTC, Align method=Gradients2026.07 | 92.3 | 11.9 | 116 | 14 | |
| Parakeet CTCModel Type=CTC, Align method=Posteriors2026.07 | 92.3 | 11.9 | 99 | 11.1 | |
| Crisper-WhisperModel Type=AED, Align method=Gradients2026.07 | 92 | 12.7 | 56 | 33.8 | |
| Crisper-WhisperModel Type=AED, Align method=Cross-att.2026.07 | 92 | 12.7 | 44 | 59.6 | |
| Parakeet TDTModel Type=Transd., Align method=Gradients2026.07 | 91.8 | 12.8 | 155 | 7.8 | |
| Parakeet TDTModel Type=Transd., Align method=Posteriors2026.07 | 91.8 | 12.8 | 86 | 17.1 | |
| Parakeet RNN-TModel Type=Transd., Align method=Gradients2026.07 | 89.8 | 13.6 | 156 | 7.6 | |
| Parakeet RNN-TModel Type=Transd., Align method=Posteriors2026.07 | 89.8 | 13.6 | 92 | 13.2 | |
| OWSM-CTCModel Type=CTC, Align method=Gradients2026.07 | 89.2 | 14.3 | 142 | 9.9 | |
| OWSM-CTCModel Type=CTC, Align method=Posteriors2026.07 | 89.2 | 14.3 | 114 | 5.4 | |
| VoxtralModel Type=Speech LLM, Align method=Gradients2026.07 | 88.3 | 15.3 | 87 | 25.3 | |
| VoxtralModel Type=Speech LLM, Align method=Self-att.2026.07 | 88.3 | 15.3 | 67 | 37.4 | |
| Whisper-large-v3Model Type=AED, Align method=Gradients2026.07 | 88.1 | 15.3 | 66 | 40.4 | |
| Whisper-large-v3Model Type=AED, Align method=Cross-att.2026.07 | 88.1 | 15.3 | 66 | 40.4 | |
| FastConformer (streaming)Model Type=Transd., Align method=Gradients2026.07 | 87.1 | 15.9 | 206 | 5.9 | |
| FastConformer (streaming)Model Type=Transd., Align method=Posteriors2026.07 | 87.1 | 15.9 | 318 | 0.4 | |
| FastConformer (streaming)Model Type=CTC, Align method=Gradients2026.07 | 86.5 | 16.8 | 164 | 9.7 | |
| FastConformer (streaming)Model Type=CTC, Align method=Posteriors2026.07 | 86.5 | 16.8 | 363 | 0.2 | |
| OWLS-1BModel Type=AED, Align method=Gradients2026.07 | 85.7 | 20.8 | 1,024 | 3.8 | |
| OWLS-1BModel Type=AED, Align method=Cross-att.2026.07 | 85.7 | 20.8 | 1,312 | 7.8 | |
| Whisper-baseModel Type=AED, Align method=Gradients2026.07 | 85.5 | 18 | 85 | 25.5 | |
| Whisper-baseModel Type=AED, Align method=Cross-att.2026.07 | 85.5 | 18 | 77 | 35.8 | |
| Emformer (streaming)Model Type=Transd., Align method=Gradients2026.07 | 72.1 | 30.5 | 155 | 7.1 | |
| Emformer (streaming)Model Type=Transd., Align method=Posteriors2026.07 | 72.1 | 30.5 | 391 | 0 |