Residual Neural Network precisely quantifies dysarthria severity-level based on short-duration speech segments.

Area of Science:

• Speech processing
• Machine learning for healthcare
• Neurology

Background:

• Dysarthria severity classification is crucial for patient treatment monitoring and improving speech technologies.
• Current methods using Convolutional Neural Networks (CNNs) struggle with short speech segments.
• Enhanced detection of dysarthria severity from brief speech samples can improve system performance.

Purpose of the Study:

• To propose a novel Residual Network (ResNet)-based technique for dysarthria severity classification using short speech segments.
• To evaluate the efficacy of the proposed ResNet model against baseline CNNs and other methods.
• To enhance the performance and applicability of speech-based systems for impaired voices.

Main Methods:

• Developed a Residual Network (ResNet) model to process short-duration speech segments.
• Conducted experiments using the standard Universal Access corpus.
• Performed comparative analyses against baseline Convolutional Neural Networks (CNNs), Gaussian Mixture Models, and Light CNNs.

Main Results:

• The proposed ResNet approach achieved 98.90% classification accuracy and 98.00% F1-score.
• Demonstrated an average improvement of 21.35% in classification accuracy and 22.48% in F1-score compared to baseline CNNs.
• Experimental results confirm the efficacy and practical applicability of the ResNet model.

Conclusions:

• The novel ResNet-based technique effectively classifies dysarthria severity from short speech segments.
• This approach significantly outperforms existing methods, offering practical benefits for clinical assessment and speech technology.
• The study validates the potential of ResNet for real-world applications in speech-based health monitoring.