Performance of Articulation Kinetic Distributions Vs MFCCs in Parkinson’s Detection from Vowel Utterances

Abstract

Speech is a vehicular tool to detect neurological degeneration using certain accepted biomarkers derived from sustained vowels, diadochokinetic exercises, or running speech. Classically, mel-frequency cepstral coefficients (MFCCs) have been used in the organic and neurologic characterization of pathologic phonation using sustained vowels. In the present paper, a comparative study has been carried on comparing Parkinson’s disease detection results using MFCCs and vowel articulation kinematic distributions derived from the first two formants. Binary classification results using support vector machines avail the superior performance of articulation kinematic distributions with respect to MFCCs regarding sensitivity, specificity, and accuracy. The fusion of both types of features could lead to improve general performance in PD detection and monitoring from speech.