Code-Switched Automatic Speech Recognition on CS-FLEURS XTTS-TRAIN read (test)
7.9CERLoRA SFT
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| LoRA SFTLanguage Pair=deu-eng, Data Scale=100%, Method Configuration=LoRA SFT2026.07 | 7.9 | 0.3 | — | — | |
| RLVR (CER+SHR+refine)Language Pair=deu-eng, Data Scale=10%, Method Configuration=CER+SHR+refine2026.07 | 10.2 | 0 | — | — | |
| RLVR (CER+SHR+refine)Language Pair=deu-eng, Data Scale=20%, Method Configuration=CER+SHR+refine2026.07 | 10.7 | 0.2 | — | — | |
| RLVR (CER+refine)Language Pair=deu-eng, Data Scale=10%, Method Configuration=CER+refine2026.07 | 11.2 | 0.7 | — | — | |
| RLVR (CER only)Language Pair=deu-eng, Data Scale=10%, Method Configuration=CER only2026.07 | 11.5 | 0.2 | — | — | |
| LoRA SFTLanguage Pair=deu-eng, Data Scale=20%, Method Configuration=LoRA SFT2026.07 | 12.3 | 0.8 | — | — | |
| RLVR (CER+SHR+refine)Language Pair=ara-eng, Data Scale=10%, Method Configuration=CER+SHR+refine2026.07 | 26.9 | 8 | — | — | |
| Base Qwen2-AudioLanguage Pair=deu-eng, Method Configuration=Base2026.07 | 28.8 | — | — | — | |
| LoRA SFTLanguage Pair=ara-eng, Data Scale=100%, Method Configuration=LoRA SFT2026.07 | 33.8 | 17.7 | — | — | |
| RLVR (CER+SHR+refine)Language Pair=Average, Data Scale=10%, Method Configuration=CER+SHR+refine2026.07 | — | 2.5 | 13.8 | 15.5 |