说话人识别模型失配下的似然得分补偿变换  

Speaker recognition likelihood score compensation transformation under model mismatch

作　　者：包永强[1] 赵力[2] 邹采荣[2] 

机构地区：[1]南京工程学院通信系,江苏南京211167 [2]东南大学无线电工程系,江苏南京210096

出　　处：《电路与系统学报》2006年第4期51-55,共5页Journal of Circuits and Systems

基　　金：国家自然基金(NO:60472058);教育部博士点基金(NO:20050286001);教育部"新世纪优秀人才支持计划"

摘　　要：基于与文本无关说话人识别最常用的模型——高斯混合模型(GMM)的输出帧似然概率的统计特性,提出了一种非线性变换方法——似然得分补偿法。理论分析和实验结果表明:与常用的最大似然(ML)变换相比,该方法可降低误识率达20%。结果还表明:似然得分补偿法基本克服了在与文本无关说话人识别系统中,当说话人的个性特征不断变化、环境对系统识别构成影响从而导致识别模型失配情况下,需要对模型的得分进行补偿的局限。Based on the statistic characteristic of frame likelihood probability output by Gaussian mixture model （GMM） which is the best text-independent speaker recognition model, likelihood score compensation transformation as a non-linear transform method is presented. The theoretical analysis and experiment shows that it could reduce the error recognition ratio 20% as compared with Max/mum-Likelihood （ML） transformation. The result shows that likelihood score compensation transformation should be adopted for canceling the influence of variations in speech characteristics, noise and model mismatch. The result also shows that process on frame likelihood probability output by GMM is effectual way of decreasing the influence of noise and improving the recognition ratio.

关 键 词：与文本无关说话人识别 混合高斯模型 似然得分补偿变换 

分 类 号：TN912.34[电子电信—通信与信息系统]