Delta-MFCC Features and Information Theoretic Expectation Maximization based Text-independent Speaker Verification System
Sheeraz Memon, Imran Ali Jokhio, SanaHoor Arisar, Margaret Lech, Namunu Chinthaka Maddage · IETE Journal of Research · 2012
Gaussian Mixture Model (GMM)-based speaker models yield very good performance in text.independent speaker verification systems. GMMs use expectation maximization (EM) as an optimization procedure to train speaker models. This paper proposes the information theoretic expectation maximization (ITEM) with improved convergence rates to train the speaker models. The approach amounts to information theoretic (IT) since it uses the parzen density estimation and Kullback.Leibler (KL) divergence measure. EM encounters the problem of convergence thus to enhance the convergence rates of EM, an IT procedure is incorporated. The proposed ITEM algorithm adapts means, covariances, and weights, like the conventional EM algorithm; however, this process is not conducted directly on feature vectors but on a smaller set of centroids derived by the IT procedure, which simultaneously minimizes the divergence between the Parzen estimates of the feature vector's distribution within a given class and the centroids distribution within the same class. The ITEM algorithm was applied to the speaker verification problem using NIST 2001, 2002, 2004, and 2006 speaker recognition evaluation corpora and MFCC with delta, energy, and zero-crossing features. The results showed an improvement of the equal error rate over the classical EM approach. The ITEM method also showed higher convergence rates compare with the EM method.