The topic of this paper is speech modeling using the Variational Bayesian Mixture of Gaussians algorithm proposed by Hagai Attias (2000). Several mixtures of Gaussians were trained for representing cepstrum vectors computed from the TIMIT database. The VB-MOG algorithm was compared to the standard EM algorithm. VB-MOG was clearly better, its convergence was faster, there was no tendency to overfitting, and finally, it gave consistently better likelihoods for unseen test data using any given number of the mixture components.