Speaker Verification Using Score Level Fusion of MFCC and GFCC

Abstract

The feature analysis component of an Automated Speaker Verification (ASV) system plays a crucial role in the overall performance of the system. There are many feature extraction techniques available, but ultimately we want to maximize the performance of these systems. Current state-of-the-art ASV system performs quite well in a controlled environment where the speech signal is noise free. The objective of this thesis investigates the results that can be obtained when you combine Mel-Frequency Cepstral Coefficients (MFCC) and Gammatone Frequency Cepstral Coefficients (GFCC) at score level. The MFCC and GFCC feature components combined are suggested to improve the reliability of a speaker verification system. The MFCC are typically the “de facto” standard for speaker recognition systems because of their high accuracy and low complexity; however they are not very robust at the presence of additive noise. The GFCC features in recent studies have shown very good robustness against noise and acoustic change. The main idea is to integrate MFCC & GFCC features to improve the overall Speaker Verification system performance in low signal to noise ratio (SNR) conditions. Also, we employed a simpler Gaussian membership function (GMF) based matching process. Last, we use Gaussian Mixture Model (GMM) and K-nearest neighbour (KNN) to measure the similarity in the verification stage. The experiments are conducted on the text independent VoxForge database, where the test utterances are mixed with noises at various SNR levels (-5, 5, 10 and 20 dB). Experimental results verify the validity of our proposed approaches in personal authentication.