Text-independent speaker verification based on relation of MFCC components

Abstract

GMM is prevalent for speaker verification. It performs very well but needs a background model to give a reference value, which greatly influences the error rate. In order to get a better generalization result, a large database with lots of people is needed to train the background model. In this paper, a new method without background model is proposed, which is called the correlation and kernel function method (CK method). In the CK method, the correlation and uncorrelation of MFCC are used to identify individuals, and a kernel function is used to work out the likelihood of two models. It works more than 30 times as fast as GMM method does, but requires fewer data to train and less space to store the model. But its performance is nearly identical to that of GMM. So it is suitable for real-time computation.