说话人自动识别技术研究  

On Automatic Speaker Recognition

作　　者：曹敏[1] 王浩川[1] 

机构地区：[1]中州大学信息工程学院,郑州450044

出　　处：《中州大学学报》2007年第2期122-124,共3页Journal of Zhongzhou University

摘　　要：主要对文本无关的说话人识别技术进行一些探讨。与语音识别不同,说话人识别技术必须提取说话人依赖特点,而语音特征量的选取是利用说话人声音的频谱通过分离傅立叶变换(DCT)获得的。在训练阶段,每一个说话者通过矢量量化产生一个码书(语音数据库)。在认识阶段期间,通过对欧几里德距离代表VQ的计算来减少失真。在一定范围的说话人的语音库中,测试结果表明有很高的识别率,可以达到96%。This paper seeks to develop an Automatic Speaker Recognition（ASR） system, while this thesis concerns text -independent speaker recognition technology only. Contrary to the speech recognition, the speaker recognition is required to extract the speaker - dependent feature, thus the feature selected in this project is cepstrum, which is often applied to indicate any representation of a spectrum derived through a Discrete Fourier Transform （DCT）. During the training phase, codebooks based on extracted features are generated via Vector Quantization approach. During the recognition phase, Euclidean Distance representing VQ distortion in this project is calculated. The testing results indicate the system performs rather well ,which can achieve over 90% in general.

关 键 词：自动说话人识别技术(ASR) mel频标倒频系数(MFCC) 矢量量化(VQ) 欧氏距离测度 

分 类 号：TN912.1[电子电信—通信与信息系统]