Abstract
In recent years, speaker recognition has been an important task as the security issues become more and more important in many kinds of applications, such as information retrieval, banking transaction...etc. Many MFCC-based speaker recognizers work well when training and testing speech data are under matched conditions. If testing data in real situation where noise is hardly evitable, the performance degrades substantially.In this thesis, we try to explore other acoustic features which can be less sensitive to the noise and then can be more robust in speaker recognition. Experiments on MAT-160 and 1999 NIST SRE database with different level noise demonstrate the improvement as comparing with the traditional MFCC features.