Automatic Classification of Functional Dysphonia Using Voc2Vec Speech Representations.

Area of Science:

• Speech processing
• Machine learning
• Medical diagnostics

Background:

• Functional dysphonia is a voice disorder affecting vocal fold function.
• Accurate classification of dysphonia subtypes is crucial for effective treatment.
• Existing methods often rely on subjective perceptual analysis or limited acoustic features.

Purpose of the Study:

• To develop an automated classification system for functional dysphonia using advanced speech representations.
• To leverage a pre-trained Voc2Vec model for extracting relevant speech features.
• To differentiate between healthy, hyperfunctional dysphonia, and hypofunctional dysphonia.

Main Methods:

• Utilized the VOICED database containing speech recordings from healthy individuals and dysphonia patients.
• Extracted speech features from transformer layers of a pre-trained Voc2Vec model.
• Employed a Support Vector Machine (SVM) classifier with radial basis function kernel and addressed class imbalance using SMOTE.

Main Results:

• The Voc2Vec-based system achieved a classification accuracy of 83.37%, significantly outperforming baseline acoustic features (eGeMAPS: 72.31%, IS10: 65.21%).
• Demonstrated significant improvements in class-wise precision, recall, and F1-scores.
• Statistical analysis confirmed the significance of performance improvements (P < 0.05).

Conclusions:

• Voc2Vec-based speech representations effectively capture phonatory characteristics for functional dysphonia classification.
• The developed framework offers a non-invasive and reliable method for voice disorder assessment.
• Potential applications include clinical decision support systems for voice disorders.