Speaker recognition is the task of validating individual s identity using invariant features extracted from their voices print. Speaker recognition technology common applications include authentication, surveillance and forensic applications. This Paper investigates the performance of three automatic model selections based on Gaussian Mixture Model (GMM). These approaches are Bayesian information criterion (BIC), Bayesian Ying–Yang harmony empirical learning criterion (BYY-HEC) and Bayesian Ying–Yang harmony data smoothing learning criterion (BYY-HDS). Experimental evaluation of these methods is presented.