Multistage Automatic Speech Recognition- Named Entity Recognition Framework for Privacy Sensitive Information Recognition in Medical Speech Data

The semantic analysis of medical speech data requires two fundamental technologies, i.e., automatic speech recognition to transcribe speech signals into text and named entity recognition to identify sensitive health information and its temporal location within the transcribed text. This study addresses both challenges in the context of bilingual medical audio containing English and Mandarin Chinese. First, we propose a multistage framework that employs language-specific ASR models with tailored training strategies based on data availability and quality. Second, we present an embedding–augmentation–labeling architecture that employs a large language model- based data augmentation technique to improve NER performance, particularly for underrepresented entity categories. Evaluated on the AICUP 2025 Medical Speech Sensitive Personal Data Recognition Competition dataset, the proposed method obtained a mixed error rate of 0.1299 and a macro F1-score of 0.6051, ultimately ranking third among 271 participating teams.