Abstract: In order to overcome the influence of background noises and improve the robustness of speaker identification systems,two methods were proposed:One is to incorporate subband amplitude information with subband Mel-spectrum centroid (SMSC) because spectral peak positions remain practically unaffected in presence of additive noise.The other is to use a class transition probability matrix to model the high-level information hidden in Gaussian mixture correlation (GMC).Experiments showed that SMSC and GMC could improve the robustness of a speaker identification system in stationary noises,respectively.The average error rate of GMM-UBM system using SMSC and GMC can be reduced by 11.7% compared to conventional GMM-UBM system using MFCC.