一种用于超高斯和亚高斯混合信号盲分离的新算法

A New Blind Separation Algorithm for Hybrid Mixture of Sub and Super Gaussian Signals

揭示了ＩｎｆｏＭａｘ盲源分离算法也是以Ｋｕｌｂａｃｋ－Ｌｅｉｂｌｅｒ散度为代价函数的，它之所以能有效地用于语音盲分离，是因为所选取的非线性函数的导数能够近似为源信号的概率密度函数（ＰＤＦ）．由此又提出一种广义非线性ＩｎｆｏＭａｘ算法，该算法在估计分离矩阵的同时也对非线性函数进行迭代估计．实验结果表明这一算法能有效地分离任何超高斯和亚高斯信号的混合信号，包括语音。

In this paper, we show that the InfoMax blind separation algorithm is also based on the contrast function of Kullback Leibler divergence. Its high separating performance for speech sources is closely related to the fact that the selected nonlinear functions approximate the probability density functions (PDFs) of source signals. With this understanding, we propose a new nonlinear InfoMax algorithm in which the nonlinear functions are iteratively updated simultaneously with the estimation of unmixing matrix. Simulation results show that the algorithm can extract independent sources from the hybrid mixture of any super Gaussian and sub Gaussian signals.