基于典范相关的频域盲解卷积排序方法

An Approach for Solving the Permutation Problem of Convolutive Blind Source Separation Based on Canonical Correlation Analysis

针对频域盲反卷积中的各频率点盲分离次序不确定性问题,提出了一种基于典范相关的排序方法。根据同一个源信号相邻频率点幅度相关性较高的原理,将每个频率点与分离出的邻近频率点进行幅度典范相关,以相邻频率点幅度向量及其时延之间最大的典范相关系数来衡量相邻频点幅度相关程度的大小,再以此为依据进行重新排序。仿真实例表明:该方法与采用简单相关系数和基于信号统计模型的方法相比,能更好地利用相邻频率点幅度之间的相关信息,因而能为各频率点独立分量的排序提供更准确的依据。

By utilizing the characteristic that amplitude correlation between neighbor bins of the same signal was better than that of different signals,an approach for solving the permutation problem of convolutive blind source separation based on canonical correlation analysis was presented in this paper.Through canonical correlation analysis,the amplitude and their time-delay correlation between neighbor bins were studied.The maximal canonical correlation coefficient was looked as the degree of correlation between neighbor bins and the foundation of permutation.Compared with the simple correlation,the canonical correlation analysis could utilize the amplitude information of neighbor bins more efficiently and provided more exact warrant for permutation.Experimental results showed that the proposed algorithm could get a relatively high permutation quality and separate the mixed speech sources.