Automatic Separation of Various Disease Types by Correlation Structure of Time Shifted Speech Features

Special disease types may affect the complex mechanisms of speech production in different ways, causing various speech disorders. This is the reason why extraction of biomarkers from speech could be reliable indicators of those diseases. The present paper aims to separate healthy speech samples and different groups of disordered speech of patients with various disease types, namely depression, Parkinson, morphological alteration of vocal organs, functional dysphonia and recurrent paresis. The correlation matrices of the time shifted values of formant frequencies (F1, F2, F3), mel-filter band energy values, mel-frequency cepstral coefficients (MFCCs), fundamental frequency (F0) and intensity were used as input for the classification of the diseases. Support vector machines and k-nearest neighbor methods were utilized to compare performances. In six-class classification experiment, the best overall accuracy was 54.75%, and the accuracy was 77.59% using re-categorization of disorders into four classes. Based on the achieved results, a speech-based diagnostic tool can be created that helps clinical staff by giving them a novel marker for diagnosis.