GMM Classifier for Identification of Neurological Disordered Voices Using MFCC Features

Abstract

Automatic detection of neurological disordered subjects voice mostly relies on parameters extracted from time-domain processing. The calculation of these parameters often requires prior pitch period estimation; which in turn depends heavily on the robustness of pitch detection algorithm. In the present work cepstraldomain processing technique which does not require pitch estimation has been adopted to extract the features of voice signal. The Mel frequency cepstral coefficients (MFCCs) are computed using two methods; the fast Fourier transform (FFT) and the linear predictive coding (LPC) method. The cepstral parameters estimated from these methods are used as features to classify normal subject voice from neurologically disordered subject’s voice using Gaussian mixture model (GMM). The results of the two methods are compared, and it is found that the accuracy of LPC-MFCC based GMM classifier is 89.55% compared to FFT-MFCC based GMM classifier which is giving an accuracy of classification of 88.5%.