A novel Swin transformer based framework for speech recognition for dysarthria

Area of Science:

• Speech-Language Pathology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Dysarthria, a speech disorder common in neurological conditions like Parkinson's disease and stroke, requires timely detection.
• Existing machine learning methods for dysarthria detection suffer from high false positive rates due to speech variability and background noise.

Purpose of the Study:

• To develop an improved Swin transformer-based model, DSR-Swinoid, for more accurate dysarthria speech detection.
• To enhance the Swin transformer's local feature extraction capabilities for better identification of speech impairments.

Main Methods:

• Speech signals were converted into mel-spectrograms to capture voice patterns.
• A novel DSR-Swinoid model was developed, integrating four modules (NLF, convolutional patch concatenation, MP, MVB) to enhance local feature learning in the Swin transformer.

Main Results:

• The DSR-Swinoid model achieved a high accuracy of 98.66% in detecting dysarthria.
• The proposed model demonstrated superior performance compared to existing methods.

Conclusions:

• The DSR-Swinoid model offers a significant advancement in the accurate detection of dysarthria.
• The enhanced Swin transformer architecture effectively addresses limitations of previous methods, paving the way for improved clinical diagnostics.