Speech-to-Text Translation on FLEURS supervised training languages (test)
40.87BLEU (en→zh)POTSA (Our Model)
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| POTSA (Our Model)Bias Compensation=true, OT Alignment=true2025.11 | 40.87 | 25.97 | 31.1 | 28.97 | 32.3 | 31.84 | |
| MinMo2025.11 | 40.68 | 25.14 | 32.05 | 23.39 | 34.34 | 31.12 | |
| POTSA (w/o Bias Comp.)Bias Compensation=false, OT Alignment=true2025.11 | 40.44 | 24.83 | 30.91 | 28.06 | 31.31 | 31.11 | |
| POTSA (w/o OT Align.)Bias Compensation=true, OT Alignment=false2025.11 | 40.22 | 25.22 | 30.89 | 25.91 | 31.4 | 30.73 | |
| POTSA (En-Anchor Fixed)Anchor Strategy=Fixed2025.11 | 40.15 | 24.51 | 30.68 | 27.17 | 31.51 | 30.8 | |
| WhisperV3+Qwen2.5Training Approach=end-to-end2025.11 | 40.09 | 24.32 | 30.14 | 27.07 | 31.13 | 30.55 | |
| Qwen2.5-Omni2025.11 | 38.35 | 23.76 | 27.81 | 27.05 | 29.69 | 29.33 | |
| Qwen2-Audio2025.11 | 37.45 | 20.02 | 28.69 | 25.69 | 29.01 | 28.17 | |
| POTSA (En-Anchor Train.)Anchor Strategy=Trainable2025.11 | 23.69 | 25.22 | 19.5 | 27.09 | 31.21 | 25.34 |