A learnable, pinyin-aware post-ASR correction method with accent routing for domain-specific Chinese speech recognition

Background Automatic speech recognition (ASR) in Traditional Chinese Medicine (TCM) is hindered by dense homophones, rare terms, and accent-induced pronunciation variation. We propose Hierarchical Phonetic Fuzzy Search (HPFS), a post-ASR correction framework for domain-specific terminology. Methods HPFS measures phonetic similarity between ASR hypotheses and lexicon entries using a learnable, structure-aware pinyin edit distance optimized by metric learning. It combines hierarchical candidate retrieval with minimal-change rewriting. We evaluate a general unified model and a specialized accent-routed strategy on a unified test set derived from a 6,541-term TCM lexicon. Results The general HPFS strategy reduces baseline word error rate (WER) by 30–40% on average. The specialized workflow achieves 10.98% WER with hotwords. In the most difficult setting, accented speech without hotwords, it reaches 14.58% WER, a 48.4% relative reduction vs the general-weight configuration. Ablation studies show that hard clustering combined with multi-path search yields the optimal efficiency–recall trade-off. Discussion HPFS provides a deployment-ready, ASR-agnostic solution for accurate recognition of accented Mandarin terminology.