Transformer-Based Punctuation Restoration for Automatic Speech Recognition Systems
Mehmet Efe Yuzuguler, C. Okan Sakar · 2025
Standard automatic speech recognition (ASR) systems are successful in converting speech to text but are inadequate in correctly placing punctuation marks. With advancements in natural language processing, some advanced ASR software can restore punctuation marks; however, they are costly. Additionally, due to their predictive features, they correct errors within the text, making them unsuitable for situations where errors need to be preserved to provide user feedback, such as foreign language speaking practice platforms. In this study, a BERT-based model is proposed to correct punctuation marks in the outputs of ASR systems. Using a dataset obtained from English movie subtitles, different labeling approaches, class imbalance problems, and contextual modeling strategies were tested. The results demonstrate the effectiveness of transformer-based models and provide a detailed analysis of the impact of contextual cues on model performance.