基于新型混合模型的欠定盲分离方法

A Method for Under-determined Blind Source Separation Based on New Mixture Model

针对欠定盲分离问题,提出了一种新的源恢复方法.在时频域局部区域采用复高斯分布对源信号进行建模,将语音信号的稀疏性和局部平稳性结合在一起,提出了一种新的混合模型来描述观测信号在局部区域的概率分布.通过该模型,将每个时频点的源信号状态的判断问题转换成模型的参数估计和后验概率的计算问题,最后通过子混合矩阵的逆恢复出源信号.实验结果表明,该方法具有很快的收敛速度,并且比已有方法具有更好的分离性能.

To solve the problem of under-determined blind source separation, we propose a new source recovery method. By utilizing the complex valued Gaussian model to characterize the local distribution of source signals in each micro-region in the time-frequency domain and combining speech signalsr sparsity with their local stability, a new mixture model is derived to characterize the local distribution of observed signals. We convert the problem of judging the state of each source signal at each time-frequency point into a problem of model＇s parameters estimation and posterior probability computation. Finally, the source signals are recovered by sub-mixing matrixes inverse. Experiment results show that the proposed method converges very fast and has better separation performance compared with the existing methods.