In this paper, the GMM-based text-independent speaker identification system for Mandarin speech is modified by adding an upper layer to form a two-state HMM system. The two-state HMM aims at modeling the initial-final phonetic structure of Mandarin syllables for assisting in speaker identification. The GPD/MCE training algorithm is also applied to further improve the system. The performance of the proposed system was examined by using a 300-speaker speech database. Error rate reductions of 25-50% were achieved for the proposed two-state HMM system over the conventional GMM system.