Machine learning based estimation of hoarseness severity using sustained vowelsa).

Area of Science:

• Speech science
• Acoustic analysis
• Machine learning in healthcare

Background:

• Auditory perceptual evaluation is the standard for voice quality but suffers from subjectivity and limited scales.
• Objective measures are needed to overcome inter-rater variability and improve hoarseness assessment.

Purpose of the Study:

• To develop a continuous, objective method for evaluating hoarseness severity.
• To combine machine learning with acoustic analysis of sustained phonation.
• To correlate objective measures with subjective hoarseness ratings.

Main Methods:

• Collected 635 acoustic recordings of sustained /a/ vowel from 595 subjects with corresponding subjective hoarseness ratings.
• Extracted 50 temporal, spectral, and cepstral features, selecting a subset using statistical analysis.
• Employed logistic regression (LR) for classifying hoarseness levels and generating probability scores.

Main Results:

• Achieved 0.867 accuracy and 0.805 correlation between model predictions and subjective ratings using five acoustic features and LR.
• Demonstrated high qualitative agreement between model predictions and subjective hoarseness changes pre- and post-treatment.
• Obtained a moderate quantitative correlation of 0.567 for pre- and post-treatment assessments.

Conclusions:

• The proposed quantitative approach using machine learning shows significant promise for objective hoarseness severity estimation.
• This method has the potential to enhance the reliability and objectivity of voice quality assessment.
• Further research can refine this technique for clinical applications in voice disorder diagnosis and treatment monitoring.