弱稀疏性下的欠定语音盲分离方法  

Underdetermined Blind Speech Separation Method Under Weak Sparseness

作　　者：王国鹏[1] 刘郁林[1] 罗颖光[1] 

机构地区：[1]重庆通信学院DSP研究室,重庆400035

出　　处：《计算机工程》2009年第18期246-248,共3页Computer Engineering

基　　金：国家自然科学基金资助项目(60672157;60672158);重庆市自然科学基金资助项目(CSTC2005BB4219)

摘　　要：针对语音信号的弱稀疏性,提出一种新的基于混合矩阵估计的欠定语音盲分离方法。该方法通过主成分分析检测只有一个源信号存在时的时频点并用于估计混合矩阵,从而克服语音信号稀疏性变弱时的影响,提高混合矩阵估计精度。结合子空间法重构源信号,进一步提高分离性能,并从几何角度证明子空间方法,仿真结果表明该方法的分离性能优于Cluster-UBSS,且鲁棒性更好。This paper proposes a new method based on mixing matrix estimation for underdetermined blind speech separation, aiming at speech signals under weak sparseness. The method detects and esploits time-frequency bins with only one source by Principal Component Analysis （PCA） to estimate mixing matrix, it overcomes the shortcoming of weak sparseness of speech signals and improves the estimation precision of mixing matrix. It combines a subspace method to reconstruct the sources to improve the separation performance further. The subspaee method is proved from geometric interpretation. Simulation results show the separability of the method is better than Cluster-BSS, and it robustness is better.

关 键 词：语音盲分离 混合矩阵估计 稀疏性 主成分分析 子空间方法 

分 类 号：TN912.3[电子电信—通信与信息系统]