STUTTER-SOLVER: END-TO-END MULTI-LINGUAL DYSFLUENCY DETECTION.

Area of Science:

• Speech processing
• Computational linguistics
• Artificial intelligence

Background:

• Current dysfluency modeling relies on non-generalizable template matching.
• Existing methods lack scalability with growing datasets and cross-lingual capabilities.

Purpose of the Study:

• Introduce Stutter-Solver, an end-to-end framework for accurate dysfluency detection.
• Enable accurate type and time transcription of dysfluencies.
• Develop a scalable and multilingual dysfluency detection system.

Main Methods:

• Inspired by YOLO object detection, Stutter-Solver employs an end-to-end deep learning approach.
• Introduced three novel corpora (VCTK-Pro, VCTK-Art, AISHELL3-Pro) for dysfluency simulation.
• Utilized articulatory-encodec and TTS-based methods for simulating various dysfluency types.

Main Results:

• Stutter-Solver achieves state-of-the-art performance across all evaluated dysfluency corpora.
• Demonstrated accurate detection of dysfluency type and timing.
• Showcased effectiveness in handling co-dysfluencies and multilingual detection.

Conclusions:

• Stutter-Solver offers a scalable, accurate, and multilingual solution for dysfluency modeling.
• The novel corpora contribute to advancing dysfluency research.
• Open-sourced code and datasets facilitate further research and development.