基于LPC分析的语音特征参数研究及其在说话人识别中的应用  

A Study of Feature Parameters Based on LPC Analysis with Applications to Speaker Identification

作　　者：张玲华[1] 郑宝玉[1] 杨震[1] 

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003

出　　处：《南京邮电学院学报（自然科学版）》2005年第6期1-6,共6页Journal of Nanjing University of Posts and Telecommunications

基　　金：江苏省"青蓝工程"跨世纪学术带头人专项基金(QL003YZ);南京邮电大学科研发展基金(2001院17)资助项目

摘　　要：对LPC(线性预测系数)参数及其派生参数进行了研究,重点讨论了各参数的计算方法,在此基础上提出了一种由LPC参数和语音帧能量构成的组合参数。利用GMM对20个说话人进行了闭集文本无关说话人识别实验。结果表明,与LPC参数的派生参数相比,该组合参数可以以较少的运算量取得与LPC派生参数相当的识别效果;与直接使用LPC参数相比,该组合参数能够在运算量增加不明显的情况下改进系统的性能,特别是在测试音长度较短的情况下,对性能的改进尤为明显。In this paper, LPC predictor coefficients and LPC-derived coefficients are studied and compared from the point of view of computation method. A new set of features composed of LPC coefficients and speech frame energy is introduced. Closed-set text-independent speaker identification experiments with 20 speakers are conducted using a GMM classifier. The experimental results show that, compared to LPC-derived coefficients, the proposed feature parameters can provide comparable accuracy with lower computational complexity and compared to LPC coefficients, the proposed feature parameters can yield significantly improved performance with slightly higher computational complexity, especially for short test utterance.

关 键 词：说话人识别 特征参数 LPC分析 运算复杂度 

分 类 号：TN912.3[电子电信—通信与信息系统]