YOLO-Stutter: End-to-end Region-Wise Speech Dysfluency Detection.

Area of Science:

• Speech Processing
• Computational Linguistics
• Machine Learning

Background:

• Dysfluent speech detection is crucial for analyzing disordered speech and improving spoken language learning.
• Current rule-based systems are inefficient, lack robustness, and are sensitive to template design.
• There is a need for advanced, end-to-end methods for accurate dysfluency detection.

Purpose of the Study:

• To introduce YOLO-Stutter, the first end-to-end method for time-accurate dysfluency detection.
• To develop a robust model that handles imperfect speech-text alignment.
• To create novel dysfluency corpora for training and evaluation.

Main Methods:

• YOLO-Stutter processes imperfect speech-text alignment using a spatial feature aggregator and temporal dependency extractor.
• The model performs region-wise boundary and class predictions for dysfluency identification.
• Two new corpora, VCTK-Stutter and VCTK-TTS, were created to simulate natural dysfluencies.

Main Results:

• YOLO-Stutter achieves state-of-the-art performance in dysfluency detection.
• The method demonstrates high accuracy on both simulated and real-world (aphasia) speech data.
• The model requires a minimum number of trainable parameters, indicating efficiency.

Conclusions:

• YOLO-Stutter offers a significant advancement in end-to-end dysfluent speech detection.
• The proposed method is efficient, robust, and achieves superior performance.
• Open-sourced code and datasets facilitate further research and development in this area.