摘要
从基于GMM的与文本无关说话人识别系统的帧似然概率的统计特性出发,提出了一种对目标和非目标模型帧似然概率进行补偿变换的方法。理论推导和实验结果表明,与GMM常用的最大似然(ML)变换相比,该变换能使系统降低误识率达8.6%,因此,证明了该变换能够改善基于GMM的与文本无关说话人识别系统的识别率。
This paper presents a compensation transformation method for the frame likelihood probability of objected and non-objected models. It is according to the statistical characteristic of the frame likelihood probability in the text-independent speaker recognition system based on GMM. Theoretical analysis and experimental result indicates that the transformation can reduce the miss recognition rate up to 8.6%, compared to Maximum Likelihood (ML) transformation which is commonly used in GMM.
出处
《电声技术》
北大核心
2004年第9期40-42,共3页
Audio Engineering
基金
教育部<面向21世纪教育振兴行动计划>
教育部科学技术重点项目.