近年來，醫療語音辨識需求快速增加，但現有 ASR 系統在臨床中遇到中英混合語境、專業術語多樣化與發音差異時，常出現辨識失準的問題，尤其是藥品名稱、醫療術語與非標準的護理師發音，皆導致辨識結果與實際語意偏離。為解決上述問題，本研究提出一個以主題式詞彙關聯為核心的語音辨識錯誤矯正框架。
此框架建置護理專用辭典與非標準的護理師英文發音規則，並建立英文 CMU ARPAbet 編碼與Mi2S (成功大學資訊工程系多與心智服務實驗室)開發的中臺客語CTLR音標對應表。研究同時蒐集護理人員真實錄音，並結合 Mi2S 與 Google TTS 生成語音辨識訓練語料，以強化模型對多語者的穩健性。系統整合六種偵測模組：音素混淆、音節錯誤、字詞相似度、N-Best 差異、雙詞詞性錯誤與主題式搭配詞，並以加權評分與閥值機制進行判斷與修正。
實驗結果顯示，發音規則的導入將藥名 WER 從 74.4% 降至 24.4%；語音辨識錯誤矯正框架則將 CER 從 16.7% 降至 4.9%，優於商用與開源系統(如 Whisper-Large-v3-turbo、Google Speech-To-Text等)。此結果證實本研究提出主題式詞彙關聯語言特徵語音辨識錯誤矯正系統能有效提升醫療語音辨識準確率，減少臨床文書修正負擔，並為未來醫療語音助理與智慧病歷系統的發展奠定基礎。

In recent years, the demand for medical speech recognition has grown rapidly. However, existing automatic speech recognition (ASR) systems often struggle in clinical settings involving Chinese–English code-switching, diverse terminology, and non-standard nurse pronunciations. These challenges frequently lead to recognition errors in drug names and medical terms. To address this issue, this study proposes a thematic lexical association framework for ASR error correction in the nursing domain.
The framework builds a nursing-specific lexicon, defines non-standard pronunciation rules, and establishes a phonetic mapping between CMU ARPAbet for English word and the CTLR system developed by the Mi2S Laboratory, NCKU. Real nurse recordings were collected and augmented with Mi2S and Google TTS synthetic data to enhance robustness for multiple speakers. Six detection modules—phonetic confusion, syllabic error, word similarity, N-best difference, POS bigram error, and thematic collocation—are integrated with a weighted scoring and threshold mechanism for correction.
Experiments show that applying pronunciation rules reduced the WER for drug names from 74.4% to 24.4%, while the full framework lowered the CER from 16.7% to 4.9%, outperforming commercial and open-source systems such as Whisper-Large-v3-turbo and Google Speech-to-Text. The results demonstrate that the proposed method effectively improves medical ASR accuracy and reduces clinical documentation effort.

Daniel, P., Arnab, G., Giles, B., & Lukas, B. (2011). The Kaldi Speech Recognition Toolkit. IEEE Signal Processing Society.
Baevski, A., Zhou, H., Mohamed, A., & Auli, M. (2020). wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations. In NIPS’20: Proceedings of the 34th International Conference on Neural Information Processing Systems. Association for Computing Machinery.
Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey, C., & Sutskever, I. (2023). Robust speech recognition via large-scale weak supervision. In ICML’23: Proceedings of the 40th International Conference on Machine Learning. Association for Computing Machinery.
Ma, R., Gales, M. J. F., Knill, K. M., & Qian, M. (2023). N-best T5: Robust ASR Error Correction using Multiple Input Hypotheses and Constrained Decoding Space. In INTERSPEECH.
Ma, R., Qian, M., Gales, M., & Knill, K. (2024). ASR Error Correction using Large Language Models. In IEEE.
López-Cózar, R., & Callejas, Z. (2008). ASR post-correction for spoken dialogue systems based on semantic, syntactic, lexical and contextual information. In Speech Communication. Association for Computing Machinery.
Serai, P., Wang, P., & Fosler-Lussier, E. (2019). Improving Speech Recognition Error Prediction for Modern and Off-the-shelf Speech Recognizers. In IEEE ICASSP.
Wang, H.-W., Yan, B.-C., Wang, Y.-C., & Chen, B. (2022). Effective ASR Error Correction Leveraging Phonetic, Semantic Information and N-best hypotheses. APSIPA.
Hou, J.-D. (2024). Mixed-language Speech Recognition Based on Dual-model Detection of Candidate Words and Thematic Collocation Words Error Correction System.
Guo, J., Wang, M., Qiao, X., Wei, D., Shang, H., Li, Z., Yu, Z., Li, Y., Su, C., Zhang, M., Tao, S., & Yang, H. (2023). UCorrect: An Unsupervised Framework for Automatic Speech Recognition Error Correction. ICASSP.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017). Attention Is All You Need.
Lin, H.-Y. (2023). Multi-feature based ASR Error Detection and Correction system for Domain-Specific Task.
Yeh, C.-F., & Lee, L.-S. (2015). An improved framework for recognizing highly imbalanced bilingual code-switched lectures with cross-language acoustic modeling and frame-level language identification. In IEEE/ACM Transactions on Audio, Speech and Language Processing. Association for Computing Machinery.
Wei, V. junqiu, Wang, W., Jiang, D., Song, Y., & Wang, L. (2024). ASR-EC Benchmark: Evaluating Large Language Models on Chinese ASR Error Correction.
Zhu, L., Liu, W., & Lin, L. (2021). Improving ASR Error Correction Using N-Best Hypotheses. IEEE.
Shih-Hsuan Chiu and Berlin Chen. 2020. Innovative Pretrained-based Reranking Language Models for N-best Speech Recognition Lists. In Proceedings of the 32nd Conference on Computational Linguistics and Speech Processing (ROCLING 2020), pages 148–162, Taipei, Taiwan. The Association for Computational Linguistics and Chinese Language Processing (ACLCLP).
Errattahi, R., Hannani, A. E., & Ouahmane, H. (2018). Automatic Speech Recognition Errors Detection and Correction: A Review. In Procedia Computer Science.