基于UBM降阶算法的高效说话人识别系统  

Universal background model reduction based efficient speaker recognition

作　　者：单振宇[1] 杨莹春[1] 

机构地区：[1]浙江大学计算机科学与技术学院,浙江杭州310027

出　　处：《浙江大学学报（工学版）》2009年第6期978-982,共5页Journal of Zhejiang University：Engineering Science

基　　金：国家杰出青年基金资助项目(60525202);国家自然科学基金资助项目(60533040);教育部新世纪优秀人才计划资助项目(NCET-04-0545);国家"863"高技术研究发展计划资助项目(2006AA01Z136);长江学者和创新团队发展计划资助项目(IRT0652);浙江省自然科学基金资助项目(Y106705)

摘　　要：为了提高基于高斯混合模型-通用背景模型(GMM-UBM)说话人识别系统的运算速度,提出了通用背景模型(UBM)降阶算法,该方法采用极大似然估计法训练一个高阶UBM,再采用UBM降阶算法得到低阶UBM.采用最短距离高斯分量替换空映射集合的方法解决了空映射集问题.通过实验方法分析了3种初始化低阶UBM方法的识别结果,发现不同的初始化方法对结果影响很小.在NIST2001 SRE数据库上的实验显示,该算法使基于GMM-UBM说话人识别系统的运算速度提高了8倍,而等错误率仅上升了4.59%,表明了UBM降阶算法在小幅降低系统识别率的情况下,可大幅度提高GMM-UBM系统的运行效率.A universal background model （UBM） reduction method was proposed to speed up the Gaussian mixture model-universal background model （GMM-UBM） based speaker recognition system. A high-order UBM was trained by expectation maximization （EM） algorithm and then clustered into a new UBM with lower order. The Gaussian component with the shortest distance was adopted to replace the empty set to solve the empty mapping set problem. Three methods of initialization low-order UBM were experimentally analyzed to find out that different initialization methods converged to similar recognition results. The experiments on NIST2001 SRE Corpora showed that the equal error rate （EER） of the system only increased 4.59%, while the computation speed increased by 8 times. The UBM reduction method can considerably improve the efficiency of the GMM-UBM system while maintaining the performance.

关 键 词：说话人识别 高斯混合模型 通用背景模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]