Wavelet Transform Features to Hybrid Classifier for Detection of Neurological-Disordered Voices

Abstract

Neurological problems may lead to speech-related disorders because of improper movements of vocal fold or incomplete closure of glottis. This may alter the acoustic characteristics of voice signal, which may provide valuable clues for diagnosing certain neurological diseases. In this work, the wavelet transform and Mel-frequency cepstral coefficients, which are features of short-time analysis techniques, fused with the time-domain features, and given to a hybrid model designed using Gaussian mixture model (GMM) and support vector machines (SVM). The fusion of features and fusion of different classifiers are carried out to avoid the generation of large feature space, complexity, and delayed results. Linear predictive coded–Mel-frequency cepstral coefficients computed for selected 6-level discrete wavelet transform are given to GMM. The output of which is combined with SVM scores obtained with time-domain features as input and given to another SVM, which makes the decision to classify the data as normal or neurological-disordered voice. It is observed that this hybrid classifier model has shown an improvement with a classification accuracy of 94.3% compared with individual SVM classifier with time-domain features with 81.43% and the GMM-SVM classifier with 85.71%.