Automatic assessment of voice quality using machine learning.

Area of Science:

• Speech and Language Pathology
• Computational Linguistics
• Biomedical Engineering

Background:

• Clinical voice assessment relies on subjective perceptual ratings of dysphonia severity.
• Existing methods lack objectivity, reproducibility, and efficiency.
• Automated acoustic analysis offers potential for standardized voice evaluation.

Purpose of the Study:

• To develop and validate machine learning models for automated prediction of dysphonia severity (Grade parameter of the GRBAS scale).
• To enhance objectivity, reproducibility, and efficiency in clinical voice assessment.
• To identify key acoustic features predictive of perceptual dysphonia severity.

Main Methods:

• Collected 524 sustained /a/ voice samples from three databases.
• Extracted 47 acoustic features (spectral, cepstral, perturbation, noise-based) using Parselmouth (Praat).
• Trained and evaluated five machine learning classifiers (DT, RF, XGBoost, LightGBM, CatBoost) using 5-fold cross-validation.

Main Results:

• Gradient boosting algorithms (LightGBM, CatBoost, XGBoost) outperformed traditional tree-based models.
• LightGBM achieved the highest quadratic weighted kappa (QWK) of 0.945.
• Cepstral measures (CPPS, CSID, AVQI) and HNR were the most influential predictors of Grade, while jitter and shimmer contributed minimally.

Conclusions:

• Gradient boosting methods, particularly LightGBM, demonstrate near-expert agreement with perceptual dysphonia ratings.
• These models offer objective, interpretable tools for clinical voice assessment.
• Automated prediction of dysphonia severity can improve clinical workflow and diagnostic consistency.