基于多特征有效组合的说话人识别  

Speaker Identification Based on Efficiently Combining Manifold Features

作　　者：谢迎春[1] 于湘珍[2] 刘建平[3] 张卫华[4] 

机构地区：[1]武警工程学院研究生队,陕西西安710086 [2]武警工程学院电子技术基础实验室,陕西西安710086 [3]武警工程学院通信工程系无线通信工程教研室,陕西西安710086 [4]武警工程学院通信工程系信息工程教研室,陕西西安710086

出　　处：《现代电子技术》2005年第9期68-70,73,共4页Modern Electronics Technique

摘　　要：通过分析当今说话人识别系统中常用的一些特征参数,以提高说话人识别的识别率为目的,在Matlab 6 .5软件环境下提出了将Mel频率倒谱(MFCC)、线性预测倒谱(L PCC)及他们的一阶差分和基音周期等多种特征有效结合进行说话人识别的方法。采用短时自相关法提取基音周期,在识别过程中采用改进的动态规整算法,将模板的匹配过程与检验量的计算分离开,每帧给出一个说话人辨认结果,最后综合各帧的辨认结果,得出最佳匹配结果。经过多次实验证明,采用以上方法使用多特征有效结合比单个使用各种特征效果要好,能在一定程度上提高系统区分说话人的能力。Through analyzing some features that be used usually in speaker identification system nowadays,in order to improve the rate of identification,this paper puts forward a method that combining efficiently more features such as MFCC and LPCC and their one ranks coefficients and keynote period and so on to do speaker verification under Matlab 6.5. We pick up keynote periods by self correlation method and use a new Dynamic Time Warping(DTW) method to do identification.This new DTW method is a way that dividing template matching and calculation of test measure and calculating identification results of all frames after every frame getting out a identification result.At last,we can make out the best matching result.Through series of experiments,it proves that the method of using manifold features is better than the method of using single feature and the ability of speaker identification can be improved by using this way.

关 键 词：说话人识别 动态规整 MFCC LPCC 基音周期 

分 类 号：TN912[电子电信—通信与信息系统]