ASR Error Correction on ASR Error Correction Evaluation Set (test)
16.07WEROur full model
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Our full modelArchitecture=S2S-based approach2026.01 | 16.07 | 78.22 | 76.96 | |
| TA promptPrompting strategy=Task-Activating (TA) prompt, N-best ASR hypothesis count=5-best, Base Model=InstructGPT2026.01 | 16.62 | 77.99 | 75.98 | |
| Alpaca promptPrompting strategy=Alpaca prompt, N-best ASR hypothesis count=5-best, Base Model=InstructGPT2026.01 | 17.01 | 76.97 | 75.44 | |
| Original ASR transcript2026.01 | 17.18 | 76.56 | 75.29 | |
| Alpaca promptPrompting strategy=Alpaca prompt, N-best ASR hypothesis count=1-best, Base Model=InstructGPT2026.01 | 17.18 | 76.56 | 75.29 | |
| TA promptPrompting strategy=Task-Activating (TA) prompt, N-best ASR hypothesis count=1-best, Base Model=InstructGPT2026.01 | 17.18 | 76.57 | 75.3 |