Multitaper Based MFCC Feature Extraction for Robust Speaker Recognition System

In present scenario, speaker identification under different types of noise conditions is the most challenging task in the area of speech processing. Due to adverse effect of noise on the speech samples, there will be substantial degradation in the system accuracy. In order to overcome this problem, we proposed a novel spectrum estimation approach using multitaper windowing function. Using frequency domain averaging and multiple windowing functions, the multitaper forms a better spectrum estimation when compare to traditional hamming window function. The multitaper spectrum estimation is used in the computation of MFCC and the proposed novel approach is compared with the other feature extraction techniques. The simulation results were conducted using i-vector based model with mixture of PLDA (mPLDA) scorings. In this work the simulation results are conducted using TIMIT database and also different types of noises like car, babble and factory noises are used with different signal to noise ratio.