TranStutter: A Convolution-Free Transformer-Based Deep Learning Method to Classify Stuttered Speech Using 2D...

Area of Science:

• Speech Pathology
• Neurodevelopmental Disorders
• Artificial Intelligence in Healthcare

Background:

• Stuttering is a common neurodevelopmental disorder impacting speech fluency.
• Accurate classification of stuttering types is crucial for effective diagnosis and intervention.
• Existing methods may not fully capture the complex temporal patterns of speech disfluencies.

Purpose of the Study:

• To introduce TranStutter, a novel Convolution-free Transformer-based deep learning model for speech disfluency classification.
• To evaluate TranStutter's accuracy in identifying various stuttering types.
• To demonstrate the model's potential to enhance stuttering diagnosis and treatment.

Main Methods:

• Developed TranStutter, a deep learning model utilizing Multi-Head Self-Attention and Positional Encoding.
• Tested TranStutter on two benchmark datasets: Stuttering Events in Podcasts (SEP-28k) and FluencyBank Interview Subset.
• Evaluated model performance based on classification accuracy for different stuttering subtypes (Block, Prolongation, Repetition, Interjection).

Main Results:

• TranStutter achieved 88.1% accuracy on the SEP-28k dataset.
• The model demonstrated 80.6% accuracy on the FluencyBank dataset.
• The results indicate superior performance in capturing intricate temporal speech patterns.

Conclusions:

• TranStutter shows significant potential for revolutionizing stuttering diagnosis and treatment.
• The model's innovative architecture offers precise identification of nuanced disfluencies.
• This advancement contributes to speech pathology and neurodevelopmental research for targeted interventions.