Multitaper MFCC Features for Acoustic Stress Recognition from Speech

Abstract

Ameliorating the performances of speech recognition system is a challenging problem interesting recent researchers. In this paper, we compare two extraction methods of Mel Frequency Cepstral Coefficients used to represent stressed speech utterances in order to obtain best performances. The first method known as traditional is based on single window (taper) generally the Hamming window and the second one is a novel technique developed with multitapers instead of a single taper. The extracted features are then classified using the multiclass Support Vector Machines. Experimental results on the SUSAS database have shown that the multitaper MFCC features outperform the conventional MFCCs.