Advancing Stuttering Detection with Time Delay Neural Networks and Attention Mechanism

Stuttering is a speech disorder characterized by interruptions in speech fluency, posing significant challenges to effective communication. Although recent advancements in machine learning have contributed to stuttering detection, ensuring consistent accuracy across diverse speech patterns remains challenging. This paper introduces a novel deep learning framework that leverages Time Delay Neural Networks (TDNN) and attention mechanisms to capture both temporal and contextual speech patterns, specifically aimed at detecting stuttering. The proposed model is trained and evaluated using the SEP-28K dataset, which provides a comprehensive set of labeled stuttering speech samples. By integrating TDNN and attention mechanisms, the model focuses on critical disfluency cues, thereby enhancing detection reliability. Experimental results show that the model achieves an accuracy of 73.56%, a precision of 77.04%, a recall of 90.53%, and an F1-score of 83.15%, indicating robust performance in identifying stuttering events. These results lay the groundwork for improving automated stuttering detection systems with a focus on real-time applications and scalability across various speech datasets.