انجمن کامپیوتر ایران

Normalization methods are used to improve accuracy in speaker verification systems. In this paper, some of the most usual normalization methods including maximum, TNorm, and ZNorm normalization methods were investigated on GMM speaker modeling approach. Maximum normalization method was shown to have the best accuracy among the studied methods. Also, hybrids of normalization methods led to better speaker verification accuracy. In this paper, serial hybrid of normalization methods was also considered. It was shown that applying Maximum normalization method for two successive times leads to the best result. Serial hybrid approach increased recognition time. So we proposed twostage normalization approach that has special consideration on accuracy and recognition time of speaker verification system.