The automatic assessment of the severity of dysphonia

Perceptual evaluation of the patient’s voice is the most commonly used method in everyday clinical practice. We propose an automatic approach for the prediction of severity of some types of organic and functional dysphonia. By means of an unsupervised learning method, we have demonstrated that acoustic parameters measured on different phonetic classes are suitable for modelling the four grade assessments of the specialists (RBH subjective scale from 0 to 3). In this study, the overall hoarseness H was examined. Four specialists were asked to determine the severity of dysphonia. A k-means cluster analysis was performed for the decision of each specialist separately; the average accuracy of the four-grade classification was 0.46. The four-grade classification has been surprisingly close to the subjective judgements. Moreover, automatic estimation of severity of dysphonia was also determined. Linear regression and RBF kernel regression models were compared. The average rating of the four specialists were used as target in the experiments. Low RMSE and high correlation measures were obtained between the automatically predicted severity and perceptual assessments. The best RMS value of H was 0.45 for the model with RBF kernel, however, a simpler linear model provided the highest correlation value of 0.85, using only eight acoustic parameters.