Wavelet-based methodology for non-invasive detection and multiclass classification of voice disorders: a comprehensive study across multilingual datasets

Impaired voice function affects 1.2% of the global population and is often diagnosed through invasive procedures. Past efforts in automated voice disorder detection mainly tackled the binary 'healthy vs. unhealthy' classification. In this study, we suggest a non-invasive alternative based on speech analysis, diverging from the conventional invasive surgical methods. Both binary and multiclass classification is carried out in the present work by decomposing the speech signal extracted from German, Spanish, English, and Arabic datasets using discrete wavelet transform (DWT). The impact of varying decomposition levels on detection and classification accuracy is evident, with the fifth level of decomposition demonstrating the highest recognition rate of 90% to 99% for tasks involving voice disorder identification and multiclass classification. Results indicate that energy and statistical features derived from DWT offer richer information on pathological voices. Consequently, the proposed system could serve as a valuable adjunct for clinical diagnosis of laryngeal pathologies.