Code-switching Automatic Speech Recognition on CS-FLEURS human-read Mandarin-English (test)
14.06WERWCE + CL (tri-level)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| WCE + CL (tri-level)Backbone=Whisper-small, Adaptation=LoRA, Filtering Strategy=tri-level (Acoustic + Text + Phoneme), Loss=WCE + Contrastive Alignment2026.06 | 14.06 | 15.1 | |
| WCE + CL (N-best NM)Backbone=Whisper-small, Adaptation=LoRA, Filtering Strategy=N-best NM, Loss=WCE + Contrastive Alignment2026.06 | 14.93 | 15.72 | |
| CE + CL (N-best NM)Backbone=Whisper-small, Adaptation=LoRA, Filtering Strategy=N-best NM, Loss=CE + Contrastive Alignment2026.06 | 15.64 | 16.21 | |
| MWERBackbone=Whisper-small, Adaptation=LoRA, Loss=Sequence-level baseline2026.06 | 15.75 | 16.41 | |
| WCEBackbone=Whisper-small, Adaptation=LoRA, Loss=Weighted cross-entropy2026.06 | 16.42 | 16.68 | |
| CEBackbone=Whisper-small, Adaptation=LoRA, Loss=Standard cross-entropy2026.06 | 16.67 | 17.25 |