基于多特征序贯判决的电话语音声纹鉴别方法研究  

Based on multi-features and multi-stages decision speaker identification in telecommunication network

作　　者：田岚[1] 白树忠[1] 郑丽娜[1] 

机构地区：[1]山东大学信息科学与工程学院,山东济南250061

出　　处：《山东大学学报（工学版）》2003年第6期648-651,共4页Journal of Shandong University（Engineering Science）

基　　金：山东省自然科学基金 (Y2 0 0 2G1 3)资助项目

摘　　要：提出一种基于倒谱特征矢量量化结合音调特征序贯判决的电话语音声纹识别方法 .文中分析了长时LPC倒谱特征和音调特征的特性 ;根据LPC倒谱各维分量对说话人识别贡献不同 ,改进了失真测度 ;用矢量量化方法建立每个说话人语音模型 ,构成一种分类准确、存储数据少、响应速度快的不依赖文本的声纹鉴别系统 ;构造一种简化的音调测度 ,将声道和声源特征结合 ,采用序贯判决策略 ,提高了系统的顽健性 .实验表明 ,在 60位说话人集合中 ,用大约 10秒的话音做识别 ,正确率达Using the cepstrum and pitch parameters, a VQ sp eaker iden tification method is realized and used in telecommunication network. In this pap er, we analyze the characteristics of long-time LPCCEP parameter and the pitch parameter for text-independent speaker identification. Considering the differen t contribution of each component of LPC cepstrum, the weight is added to modify the distortion measurement. Based on the improved vector quantification (VQ), th e speaker individual model is constructed and the system is training shorter, th e response real time and the data memory smaller. In order to increase the syste m robustness, a simple pitch measurement is introduced and applied in multi-sta ge decision. The test results show that using about 10 second voice, in 60 speak er set, the correct identification rate is above 95%.

关 键 词：电话语音 矢量量化 序贯 调测 声纹识别 倒谱 话音 顽健性 存储数据 语音模型 

分 类 号：TN912[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]