Robust Speaker Identification System Based on Variational Bayesian Inference Gaussian Mixture Model and Feature Normalization

Abstract

Voice is an important human feature for identifying an individual in normal human’s communication. Automatic speaker recognition (ASR) mechanisms can, therefore, be considered a client-friendly form of biometric type that is used in applications such as banking, forensics, teleconferencing, and so on. This paper presents a text-independent speaker identification system based on Variational Bayesian Gaussian Mixture Model (VBGMM). Four types of features which are: MFCCs, derivatives of MFCCs, Log Filter-bank Energies and Spectral Sub-band Centroids, in addition to feature normalization have been used in the proposed system. The performance evaluation of proposed system is compared with the traditional Gaussian Mixture Model (GMM). The two modeling techniques with different types of covariance (Diagonal and Tied) are examined using the TIMIT and Arabic corpus. The recognition rates for the two datasets indicate that VBGMM is superior to GMM particularly when using data normalization for the extracted features. The recognition rates that achieved in this experiment were 98.3% and 93.3% for the TIMIT and the Arabic corpus respectively.