基于互信息的卷积混合语音时域盲分离算法  

Time Domain Blind Separation for Convolutive Mixtures of Speech Based on Mutual Information

作　　者：吕尧[1] 李双田[1] 

机构地区：[1]中国科学院声学研究所

出　　处：《信号处理》2008年第3期395-400,共6页Journal of Signal Processing

摘　　要：本文以互信息最小化作为分离准则,提出了一种适用于非平稳语音卷积混合信号的时域盲分离算法。其目标函数同时考虑了语音信号的短时平稳性和长时非平稳性,在短时间段中计算平均互信息,在长时间段中引入权值因子,对短时间段中计算得到的平均互信息进行加权;其分离矩阵的更新采用快速收敛的自然梯度算法;其"去白化"的后处理步骤提高了分离语音的自然度。仿真实验和分析表明了算法的有效性。In this paper,we present a time domain blind separation algorithm for convolutive mixtures of speech based on mutual information,in which a novel cost function is constructed where both short time stationarity and long time nonstationarity of speech signals are taken into account, the mutual information during short time segments are first averaged and then summed up with weight along the successive segments, natural gradient adaption is used for updating the separation system to improve the convergency speed, and a post-filtering method named as ＂Dewhitening＂ is presented to preserve the naturalness of the separated speech. Finally, an experiment with analysis is given to evaluate the algorithm.

关 键 词：盲源分离 互信息 非平稳 卷积混合 

分 类 号：TN912.3[电子电信—通信与信息系统] TN407[电子电信—信息与通信工程]