In this paper we propose a new method to select the optimal model order for the initialization of Gaussian Mixture speaker Models (GMM) based on regression tree in text-independent speaker identification system. The objective is to choose the optimal number of components which is necessary to adequately model a speaker for a good speaker identification performance according to the Bayesian Information Criterion (BIC) and agglomerative clustering. One obvious advantage of such method is that it provides a flexible framework to select an optimal speaker model order based on the training data for each client speaker. The experimental results on the YOHO corpus show that adaptive model mixture components achieves better performance, especially considering the fact that different speakers have different amounts of available enrollment data.