Abstract: This paper presents an improved speech blind separation algorithm based on natural gradient algorithm.The first step performs pre-whitening.An algorithm based on Singular Value Decomposition（SVD） is proposed for pre-whitening.Then recovers unknown independent sources from pre-whitening observed mixtures.It has examined the algorithm by computer.Simulation results demonstrate the algorithm can effectively separate random mixing natural speech signals.

摘要： 提出了一种基于自然梯度的语音盲分离改进算法。该算法首先使用奇异值分解（SVD）的方法对观测信号进行预白化处理,而后使用自然梯度算法对预白化处理后的观测信号进行分离。通过计算机模拟试验,结果显示该算法能够有效地分离随机混合的自然语音信号。