Automated Diagnosis of Voice Disorders using Acoustic Voice Metrics: A Machine Learning Approach

Objective: Currently, numerous acoustic voice metrics are being utilized to estimate an objective assessment of voice quality, including acoustic quantities such as Cepstral Peak Prominence (CPP), Pitch Strength, and Pitch Period Entropy. Additionally, indexes such as the Acoustic Voice Quality Index (AVQI) and the Acoustic Breathiness Index (ABI) combine multiple acoustic quantities into one number have been acknowledged for their efficacy in differentiating symptomatic and asymptomatic voices. Using machine learning models, this study leverages these metrics as compelling acoustic features for the classification of dysphonic and non-dysphonic voices through, thereby examining their reliability in automated voice disorder diagnosis.

Methods: This ongoing retrospective project encompasses a cohort of over 190 participants: 25% having healthy vocal attributes and 75% presenting with dysphonic voices characterized by various degrees of dysphonia severity and different etiologies. All samples were analyzed using a custom MATLAB script and the “Voice” function in Pratt. The analysis yielded values including fundamental frequency (f0), standard deviation of the fundamental frequency, jitter, shimmer, Harmonics-to-Noise Ratio (HNR), Cepstral Peak Prominence (CPP), smoothed CPP (CPPs), Alpha ratio, and Pitch Period Entropy (PPE), along with AVQI-3 and ABI. Acoustic metrics were extracted from both sustained vowel [a:] and running speech tasks. The computed acoustic measures were evaluated across different machine learning models, utilizing performance metrics such as sensitivity, specificity, accuracy, precision, and F1-Score.

Results and Conclusions: Machine learning models are anticipated to be successfully developed as classifiers. Comparative analysis identifies which acoustic measurements are more effective in distinguishing between dysphonic and non-dysphonic voices. The models are projected to demonstrate varying levels of accuracy, specificity, and sensitivity scores based on the employed acoustic measurements. By identifying the most pertinent acoustic measurement, machine learning models can be enhanced for more accurate and efficient diagnosis of voice disorders. The promising performance of these models, as anticipated, would underscore their potential in bolstering the clinical assessment process of voice disorders.