Novel windowing technique of MFCC for speaker identification with Modified Polynomial Classifiers

Speech is one of the most popular parameter used to identify a speaker by her spoken phrase. Feature extraction from speech is a necessary first step in a speaker identification process. Traditionally computation of the Mel Frequency Cepstral Coefficient (MFCC) features use hamming window, as a preprocessing step to reduce spectral leakages. However, hamming window results in reasonable side lobes along with the desired main lobe. This paper deals with a modified algorithm to compute MFCC feature by integrating phase as well as slope information in computing power spectrum. A modified training algorithm is used to train the polynomial classifier which is used for speaker identification. Experimental results using Matlab show that the novel windowing technique with Modified Polynomial Classifier shows consistently better performance over hamming window. There is an improvement in the accuracy of identification especially for large database with low memory usage. It also reduces computational complexity.