说话人识别中改进特征提取算法的研究  

Study of improving feature extraction algorithm in speaker recognition

作　　者：宋乐[1] 白静[1] 

机构地区：[1]太原理工大学信息工程学院,山西太原030024

出　　处：《计算机工程与设计》2014年第5期1772-1775,1781,共5页Computer Engineering and Design

基　　金：国家自然科学基金项目(61072087);山西省科技攻关基金项目(20120313013-6)

摘　　要：为了提取到能够区分不同说话人个性特征的最优特征参数,采用在Mel频率倒谱系数(Mel-frequency cepstrum coefficients,MFCC)基础上进行改进的复合参数,即增加归一化短时能量参数和一阶差分所构成的特征矢量作为特征。针对高维特征参数,提出了一种基于相关距离Fisher准则的特征选取方法,利用该方法对提取出的参数进行加权降维。通过实验对比结果表明,该算法提高了识别率,具备可行性与优越性,是一种有效的特征提取算法。To extract speaker＇s personality characteristics that different speakers can be distinguished better. Firstly, a design method of Mel cepstrum composite coefficients based on the Mel Frequency cepstrum coefficient （MFCC） is constructed. The normalized short-time energy parameters and first-order difference are used to be the improved feature parameters. Then, in view of the high dimensional parameters, a algorithm for feature selection about Fisher criterion with correlation distance is introduced. The weighted algorithm designed to lower dimension for Mel cepstrum composite coefficients. Finally, an simulation example is presented to prove the recognition rate is improved as well as the feasibility and superiority of the proposed method, indicated that this study is an effective feature extraction algorithm.

关 键 词：说话人识别 特征提取 归一化短时能量 梅尔倒谱复合参数 相关距离Fisher准则 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]