Combining Evidences from Variable Teager Energy Source and Mel Cepstral Features for Classification of Normal vs. Pathological Voices

In this paper, novel Variable length Teager Energy Operator (VTEO) based Mel frequency cepstral coefficients, namely, VTMFCC are proposed for automatic classification of normalvs. pathological voices. Experiments have been carried out using this proposed feature set, Mel Frequency Cepstral Coefficients (MFCC), and their score-level fusion. Classification was primarily performed using a discriminatively-trained 2
^nd
order polynomial classifier on a subset of the MEEI database for a feature dimension of 12. The equal error rate (EER) on fusion was reduced by 3.2% than MFCC alone which was used as the baseline. The classification accuracy was analyzed for different dimensions of feature vector. Furthermore, results obtained for the 2
^nd
order classifier were compared with the results obtained from the 3
^rd
order polynomial classifier for different feature dimensions. In addition, the effectiveness of dynamic features, in particular, delta, delta-delta, and shifted delta cepstral features have been investigated for this particular problem. It has been observed that the score-level fusion (with equal weights) of proposed feature set and state-of-the-art MFCC gave better classification performance than MFCC alone for various evaluation factors considered in this paper.