A robust clustering approach to fuzzy Gaussian mixture models for speaker identification

Abstract

The Gaussian mixture model (GMM) is a currently used method for speaker recognition. The fuzzy GMM (FGMM) proposed in previous work (D. Tran et al., 1998) is a fuzzy clustering based modification of the GMM. Although both the FGMM and the GMM are capable of achieving high identification accuracy, they have a common disadvantage in the problem of sensitivity to outliers. The paper presents an improvement for the FGMM to handle this problem. Experimental results on 16 speakers using the TI46 database are also reported.