欠定情形下语音信号盲分离的新方法

New method of underdetermined blind voice source separation

提出了一种新的两步法来实现欠定情形下语音信号的盲分离。第一步,采用一种重构观测信号采样点搜索法来估计混合矩阵;第二步,提出了一种伪提取矢量的概念,通过伪提取矢量来提取取值占优的源信号的采样值来恢复源信号。在源信号的恢复过程中,还使用了经典的基于线性规划的欠定盲源分离方法。结果表明:该方法由于在信号的各采样点处无须优化,在源信号的分离过程中,分离速度要比基于线性规划的方法快数倍,且分离精度不低于基于线性规划的方法。仿真结果表明了该算法的良好性能。

This paper put forward a new kind of two-step approach to separate underdetermined blind voice signals.The first step：the mixing matrix was estimated by searching observation signals newly constructed.The second step：a new concept of pseudo extraction vectors was put forward and the pseudo extraction vectors were used to recover the signal dominating for each sampling.In the process of separating the source signals,the method based on linear programming was also used.Some conclusions could be seen： the method based on pseudo extraction vectors free of optimizing process increases the velocity of separating source signals.The velocity of separation of the method based on pseudo extraction vectors was several-fold of the method based on linear programming.The precision of separation of the method based on pseudo extraction vectors was also not lower than the method based on linear programming.The simulating results illustrate the better performance of the method.