Abstract: When training data is limited, the performances of speaker identification (SI) system deteriorate owing to the weak generalization of GMM trained by EM algorithm. This paper presents a solution of this problem by using a evolutionary Gaussian mixture model (EGMM) as a modeling method of SI system. Based on evolutionary algorithm, EGMM optimizes both the structure and the parameters of a GMM to better its generalization ability. Also, other speakers' discriminative information is integrated into objective function to increase the accuracy of classification. According to the characteristics of GMM, we design two special evolutionary operators and present a new hybrid evolutionary algorithm using genetic algorithms and evolutionary programming. The preliminary experimental results show that the speaker models based on EGMM have more generalization ability. Compared with traditional GMM, the correct recognition rate of SI system based on EGMM increases by approximate 3%. Furthermore, the GMMS in new system have smaller average size.

Key words: evolutionary algorithm, speaker identification, model selection