Abstract: This paper improves the traditional method of Gaussian Mixture Model（GMM） parameters initialization at the time of GMM training.A new approach which combines division and K-means clustering is presented.The experiment shows that the proposed method can achieve the average recognition rate increase by 15.47% and 7.5% compared with the randomization and K-means clustering.At the same time，the impact of the order of GMM，covariance threshold and pre-emphasis coefficient on system recognition rate are studied.Meanwhile，the experiment results are analyzed in detail.In order to make the speaker recognition system get a higher recognition rate，their optimal values are chosen from the experiment data.The experiment shows that the system can achieve the recognition rate with above 90% under the provided experimental condition.

Key words: speaker recognition, Gaussian Mixture Modal（GMM）, Mel Frequency Cepstrum Coefficient（MFCC）, combination division and K-means clustering

摘要： 在高斯混合模型（Gaussian Mixture Model，GMM）训练时，对传统的模型参数初始化方法（随机法、K均值聚类法）进行改进，提出分裂法与K均值聚类相结合的新方法。实验表明，采用改进的方法与传统方法相比，系统平均识别率有15.47%和7.5%的提高。研究了GMM的阶数、协方差阈值、预加重系数对系统识别率的影响。对实验结果进行详细分析，并根据实验数据，取它们各自表现最好的值，从而使构建的说话人识别系统获得一个较高的识别率。实验表明，在规定的实验条件下，系统可达到90%以上的识别率。

关键词: 说话人识别, 高斯混合模型, 美尔频率倒谱系数（MFCC）, 分裂法与K均值聚类结合法