Automatic Classification of Functional Dysphonia Using Voc2Vec Speech Representations

This study aims to develop a machine learning-based automatic classification framework for functional dysphonia using speech representations derived from a pre-trained Voc2Vec model. The proposed system targets three classes: healthy, hyperfunctional dysphonia, and hypofunctional dysphonia.

Experiments were conducted using the VOICED database, which includes speech recordings from healthy individuals and patients with hyperfunctional and hypofunctional dysphonia. Speech signals were preprocessed and segmented, and features were extracted from different transformer layers of a pre-trained Voc2Vec model. A support vector machine (SVM) with a radial basis function kernel was used for classification. The performance of the proposed approach was compared with standard acoustic feature sets, namely IS10 and eGeMAPS. A fivefold speaker-independent cross-validation scheme was employed, and class imbalance was addressed using SMOTE.

The proposed Voc2Vec-based system achieved a maximum classification accuracy of 83.37%, outperforming the baseline feature sets (72.31% for eGeMAPS and 65.21% for IS10). Improved class-wise precision, recall, and F1-scores were observed across all classes. Statistical analysis using the Wilcoxon signed-rank test confirmed that the performance improvements were significant (P < 0.05).

The results demonstrate that Voc2Vec-based representations effectively capture relevant phonatory characteristics for functional dysphonia classification. The proposed approach provides a non-invasive and reliable framework for distinguishing between healthy, hyperfunctional, and hypofunctional voices, with potential applications in clinical decision support systems.