Analysis of vocal tract disorders using Mel-Frequency Cepstral Coefficients and Empirical Mode Decomposition based features

This work investigates the possibility of developing a non-invasive technique for the detection of vocal tract disorders from voice samples of patients. The existing techniques are invasive, expensive or both and hence the relevance of this study. Mel-Frequency Cepstral Coefficients (MFCC), dynamic measures derived from MFCC and statistical features extracted from Empirical Mode Decomposition (EMD) of voice samples provide distinct features capable of discriminating pathological and normal voice samples. A Support Vector Machine (SVM) classifier is used for classification. Experimental evaluations on a voice database created from videostroboscopy data yield accuracies more than 90%. It is observed that although MFCC is a good discriminating feature as far as speech/voice segments are considered, EMD, being a significant analysis technique for non-linear, non-stationary signals, also proves to give good discrimination possibilities for detecting vocal tract disorders.