Instruction-Tuned LLMs for Multilingual Medical ASR and Privacy Entity Extraction

In this study, we propose a comprehensive framework for sensitive information recognition in bilingual doctor-patient speech, developed for the AI CUP 2025 Spring Challenge. The task consists of two subtasks: automatic speech recognition (ASR) and named entity recognition (NER) for privacy-related spans. For ASR, we fine-tuned the Whisper-large-v3-turbo model separately on Chinese and English audio, achieving improved Mixed Error Rates (MER) through language-specific modeling. For NER, we reformatted the data into Alpaca-style instruction-output pairs and employed QLoRA fine-tuning on two complementary LLMs, Gemma-27B and Qwen2.5-32B-Instruct. We further applied data augmentation using semantic rewriting with a distilled Qwen-7B model and per-formed category-wise training to enhance low-frequency entity detection. To align predicted spans with timestamps, we devised a sliding window approach combined with Gemini-based LLM inference for span recovery. Finally, we integrated predictions from Gemma and Qwen using a non-overlapping span merging strategy. Our system achieved second place in the ASR subtask and ranked sixth overall, demonstrating the effectiveness of combining LLMs, instruction tuning, and hybrid post-processing for medical speech privacy protection (Our code and data for this paper are made available at: https://github.com/pango0/aicup ).