一种非限定语音说话人实时识别的方法及其实现  

Text—Independent Speaker Indentification and Its Real—time Implementation

作　　者：俞振利[1] 张礼和[1] 

机构地区：[1]杭州大学

出　　处：《信号处理》1990年第3期183-190,共8页Journal of Signal Processing

摘　　要：本文首次使用了最大熵谱法估计的LPC反射系数矢量的长期平均作为说话人的语音特征矢量,利用所定义的特征矢量的平均自差异函数,平均互差异函数及平均互——自差异比函数分析了特征矢量用于非限定语音的说话人识别的有效性和说话人的可区分性;从模式识别分类的Bayes判决准则出发,导出了便于计算和程序实现的简化判决公式——欧几里德空间的加权和距离公式,并利用平均差异函数选择加权系数;提出了用序贯判别法对集外说话人的拒识方法;研制了相应的以微机为核心的实时响应的实验系统,响应速度为3秒。用此系统对20个说话人进行了非限定语音的说话人识别试验,误音率为10.67％,误拒率为5.67％,正确识别率95.41%。In this paper, the long—term averaged vector of the LPC reflecting coefficient vector abstracted by the Maximum Entropy Spectral Estimation method is used as the speech feature vector of speakers. The efficiency of the feature vectors and the discrimination of speakers for the purpose of text—independent speaker identification is analyzed using the average within speaker variance funeton, the average between speaker variance fuction and the ratial of average between—within speaker variance ruction defined in this paper. A simplized decision equation——the weighted Euclead space distance sum of feature vectors, which is easy for computing and programming, is conduced from the Bayes decision principle. The reciprocal of the average within speaker variance fundtion isselected as the optional weighting function. A sequencial decison principle for refusing the out—set speakers is presented. A correspondent experimental system related with a microcomputer is implemented. The response time is 3 seconds. The performants of text—independent speaker identification test is carried using this system for the population of 20 speakers. The correct rate of the identification is 95. 41percent, the faulse accept rate is 10. 67percent and the faulse refuse rate is 5. 67 percent.

关 键 词：语音识别 最大熵谱法 实时识别 

分 类 号：TN912.34[电子电信—通信与信息系统]