Contrastive Training with LLM-generated Near-Misses for Robust Code-Switching Speech Recognition

In three linesContrastive training method to improve speech recognition for code-switching (language alternation). Authors identify critical regions, generate acoustically plausible near-misses via LLM, then fine-tune Whisper-small with LoRA. Results: >2% error reduction on CS-FLEURS and ViMedCSS datasets.

Summary generated by Claude — human-verified