Pathological Voice Classification Based on a Single Vowel´s Acoustic Features

This research focuses on the classification of pathological voice from healthy voice based upon 30 acoustic features derived from a single sound of vowel /a/. The method includes two steps. The first is the feature space transformation and data dimension reduction based on PCA. The second step is the classification of transformed features using support vector machine (SVM). The method was validated with a sound database provided by the Massachusetts eye and ear infirmary (MEEI). 216 data files, collected from an identical phoneme vowel /a/ from subjects of healthy and pathological cases, were used for examination. The original 30 acoustic features and the transformed features derived with PCA were modeled by the SVM classifier using the radial basis function (RBF) as a kernel function. The deviation residual (DR) is employed as the index for performance evaluation. In the 5 fold cross-validation, the results show that the pathological cases suffered from various diseases were detected with classification rates of up to 98.1%, while the sensitivity and specificity were 92.5% and 99.4% respectively. This preliminary result suggests that the highly promising feasibility of the detection of mental and physical status through analyzing a single tone of voice.