Detection and Correction of Homophone Errors in Thai Language using Machine Learning Techniques in Courtroom Transcriptions

Homophone errors in Thai courtroom transcriptions pose significant challenges in ensuring accurate and reliable legal documentation. Due to the complexity of the Thai language, which includes numerous homophones and tonal variations, speech-to-text (STT) systems often struggle to distinguish between phonetically similar words in legal discourse. This study aims to enhance Thai courtroom transcription accuracy by detecting and correcting homophone errors using machine learning (ML) techniques. The research integrates Google’s Speech-to-Text API with an ML-based correction model that utilizes deep learning architectures, including Thai-BERT and Transformer-based models. The proposed approach employs contextual analysis, a domain-specific Thai Legal Term Dictionary, and post-processing algorithms to refine transcriptions. Experimental results demonstrate that the ML-enhanced system improves homophone detection and correction, increasing overall transcription accuracy from 91.10% to 93.73%, with a homophone correction rate of 71.5%. The evaluation further confirms the model’s effectiveness, achieving a precision of 89.2%, recall of 78.3%, and an F1-score of 83.3%. These findings highlight the potential of integrating machine learning into courtroom transcription workflows, offering a scalable and automated solution for improving judicial documentation. Future work will focus on expanding the dataset with a specialized Thai Legal Speech Corpus, optimizing the model’s adaptability to diverse courtroom environments, and integrating real-time transcription correction systems for enhanced legal applications.