基于聚类分析的实时语音信源数估计方法  

Real-Time Estimation Method for Number of Speech Source Based on Clustering Analysis

作　　者：杨杰[1] 张家树[1] 

机构地区：[1]西南交通大学信号与信息处理四川省重点实验室,成都610031

出　　处：《数据采集与处理》2009年第2期144-148,共5页Journal of Data Acquisition and Processing

基　　金：国家自然科学基金(60572027)资助项目;教育部新世纪优秀人才计划(NCET-05-0794)资助项目

摘　　要：在信源个数未知的语音信号盲分离中,首先要解决的是信源个数的估计问题。在深入分析基于判定主特征值个数的信源数估计方法出错原因和条件的基础上,提出利用对少量观测样本自相关矩阵的特征值进行聚类分析的方法来实时估计语音信源数。同时,依据期望聚类过程,定义了一种新的类间距离,并给出了特征值聚类过程的详细计算步骤。仿真实验表明该方法的估计准确率明显优于基于主特征值的估计方法。The accurate estimation of number of sources plays a key role in the blind speech source separation. The dominant eigenvalue-based method often results in the wrong estimation of the number of speech sources due to nonstationary properties of the speech. Therefore, an eigenvalue clustering-based estimation method for the number of speech sources is proposed based on a new distance among different classes defined by the expected process of clustering. Simulations show that the method can more accurately estimate the number of speech sources than the dominant eigenvalue-based method.

关 键 词：语音信源 盲源分离 信源数估计 聚类分析 

分 类 号：TP911.23[自动化与计算机技术]