A Machine-Learning Algorithm for the Automated Perceptual Evaluation of Dysphonia Severity.

Area of Science:

• Speech science
• Computational linguistics
• Machine learning

Background:

• Auditory-perceptual assessments are the standard for evaluating voice quality.
• Objective and reliable methods for dysphonia severity assessment are needed.

Purpose of the Study:

• To develop a machine learning (ML) model for measuring perceptual dysphonia severity.
• To ensure the model's assessments align with expert human raters.

Main Methods:

• Utilized the Perceptual Voice Qualities Database, including sustained vowels and Consensus Auditory-Perceptual Evaluation of Voice sentences.
• Extracted acoustic and prosodic features using the OpenSMILE toolkit.
• Employed a support vector machine (SVM) model, combining features from vowels, sentences, and whole audio samples for prediction.

Main Results:

• The ML algorithm achieved a high correlation (r=0.847) with expert rater estimates.
• The root mean square error (RMSE) was 13.36.
• Combining features from different audio components (vowels, sentences, whole audio) yielded superior dysphonia severity estimation compared to individual components.

Conclusions:

• A novel ML algorithm can perceptually estimate dysphonia severity on a 100-point scale with high correlation to expert raters.
• This suggests ML algorithms can provide an objective method for voice sample evaluation.
• The developed algorithm shows promise for consistent and objective voice quality assessment.