Automatic Speech Recognition on ASGDTS (full)
17.5WERMichaud
Evaluation results
| Method | Links | |
|---|---|---|
| MichaudMethod=QLoRA r=200, Training Data=∼870h + ASGDTS, Evaluation Protocol=Contaminated evaluation (ASGDTS or same-distribution data in training)2026.05 | 17.5 | |
| Ours: Full FT (Run 16)Method=Full fine-tune, Training Data=1,367h, Evaluation Protocol=Honest evaluation (no ASGDTS in training)2026.05 | 25.6 | |
| Ours: LoRA (Run 8, CP-1500)Method=LoRA r=160, Training Data=1,011h, Evaluation Protocol=Honest evaluation (no ASGDTS in training)2026.05 | 26.28 | |
| Flurin17Method=LoRA, Training Data=STT4SG + SDS-200, Evaluation Protocol=Honest evaluation (no ASGDTS in training)2026.05 | 28 | |
| Whisper large-v3 baselineMethod=Zero-shot, Training Data=—, Evaluation Protocol=Honest evaluation (no ASGDTS in training)2026.05 | 28.56 |