一种GMM对数似然得分变换方法  

GMM Logarithm Likelihood Score Transformation

作　　者：冷自强[1] 王金明[2] 

机构地区：[1]解放军理工大学通信工程学院研究生1队,江苏南京210007 [2]解放军理工大学通信工程学院电子信息工程系,江苏南京210007

出　　处：《电子质量》2009年第1期8-9,12,共3页Electronics Quality

摘　　要：基于高斯混合模型(GMM)的说话人识别方法通常采用对数似然得分作为测试时判定目标说话人的依据。文章在分析对数似然得分特点的基础上,提出了一种改进方法,提高了测试语音帧对于目标模型和非目标模型得分的相对差值。基于TIMIT数据库的实验证明了采用变换后似然得分的说话人识别系统比采用对数似然得分的系统具有更好的识别性能和抗噪声性能。logarithm likelihood score is usually used as the criterion of judging the target speaker when testing by the speaker recognition system based on Gaussian Mixture Models （GMM）. Based on the analysis of characteristic of logarithm likelihood score, a kind of transform is presented, which increases the relative difference between the scores of target model and the other models. Experiment based on TIMIT show that speaker recognition systems use the transformed likelihood score have better performance than the systems use the logarithm likelihood score.

关 键 词：文本无关说话人识别 高斯混合模型 对数似然得分 

分 类 号：TN912.34[电子电信—通信与信息系统]