Abstract: There is a basic choice in the form of covariance matrix to be used with Gaussian mixture model in text-independent speaker identification. In general, diagonal covariance matrix is chose, which implies strong assumption that elements of the feature vector are independent, because full covariance matrix suffers from too many parameters and large computational requirement. Unfortunately, in most application the assumption is not reasonable. In order to make feature vectors more suit to be modeled with diagonal covariance, features are usually de-correlated in feature space or model space. In this paper, an improved model-based PCA transformation algorithm is presented to de-correlate the elements of feature vectors. In this algorithm, principal component analysis is directly made for covariance of Gaussians. Also, the number of parameter is deduced through tying the PCA transformation between Gaussians. Experiments on the MSRA mandarin task show that the algorithm could achieve above 35% identification error reduction over the best diagonal covariance models.