Feature extraction in speaker verification under noisy conditions

Abstract

This thesis describes the development of a robust automatic speaker verification system (ASV) with specific interest in the extraction of dominant acoustic features. Our primary investigation involves the development of robust feature extraction techniques to improve the performance of the system under noisy conditions. By far, the most widely used feature in this area is the Mel Frequency Cepstral Coefficients (MFCC). The techniques developed here are processing strategies, which improves the MFCC feature set. We have introduced four techniques to improve the robustness of the system against noise, particularly additive white Gaussian noise (AWGN). The first three are integrated processing strategies and the last one a pre-processing technique. These features are subsequently used to train a speaker model which eventually is used to represent a particular speaker. The model that we have selected is the Gaussian Mixture Model (GMM). This model is used as opposed to the Hidden Markov Model (HMM) because of its simplicity and fast processing time.