说话人模型聚类算法研究与分析  

The research and analysis of speaker clustering algorithm

作　　者：谭萍[1] 邢玉娟[1] 高翔[1] 

机构地区：[1]兰州文理学院数字媒体学院,甘肃兰州730000

出　　处：《中国建材科技》2015年第5期87-88,91,共3页China Building Materials Science & Technology

基　　金：甘肃省教育厅基金项目(2014A-125)

摘　　要：针对说话人确认识别率不高及实时性差的问题,在深入研究传统高斯混合模型以及K均值聚类算法的基础上提出两种基于说话人高斯混合模型的说话人聚类算法:KL散度聚类算法和巴氏距离聚类算法。根据不同的聚类算法,得到各个类的聚类中心模型,将其作为SVM的输入得出最终识别结果。仿真实验将两种聚类算法进行详细的分析比较,实验结果显示巴氏距离聚类算法具有较好的识别性能和抗噪性。In order to improve speaker recognition rate and real-time,two speaker clustering algorithms based on GMM speaker model are proposed in this paper.They are KL divergence clustering algorithm and Bhattacharyya distance clustering algorithm depending on K-means clustering algorithm.According to different clustering algorithm, the cluster center models for each class are obtained as SVM input to get the final recognition result. We analyze these two kinds of clustering algorithm for detailed in simulation experiment. And the results show that Bhattacharyya distance clustering algorithm has better performance and noise immunity compared to KL divergence clustering algorithm.

关 键 词：说话人确认 高斯混合模型 KL散度 巴氏距离 支持向量机 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]