基于准正交原理的多信源少观测源的盲语音信号分离

信号源个数多于观测信号个数情况下的盲源分离问题是盲信号分离中的一个难题,也是一个很实际的问题。论文在A.Hyvrinen提出的一种基于准正交原理的盲分离算法基础上,指出当混合矩阵的基矢量不满足准正交性时,可以对观测信号预白化,使混合矩阵的基矢量的准正交性得以很大提高。然后将此方法用于多信源少观测源情况下的混合语音信号分离。实验分为两个过程:(1)估计混合矩阵;(2)用最大后验概率的估计方法分离源语音信号。实验结果证明了该算法能够有效用于高维情况下多信源少观测源的盲语音信号分离。

Blind Speech Signals Separation of More Sources Than Sensors Based on the Quasi-orthogonality Theory

Blind source separation of more sources than sensors is a difficult and practical problem.This article introduces a method based on the quasi-orthogonality theory of A.Hyvrinen at first,and infers that,when the basis vectors of the mixing matrix are not quasi-orthogonal,can improve the quasi-orthogonality by prewhiting the observed signals.The second,it uses this method to blind speech signals separation of more sources than sensors.The experiment has two steps:(1)estimating the mixing matrix;(2)separating the sources by maximum a posteriori estimation.Results of the experiment show that the method is effective for blind speech signals separation.