非稀疏欠定盲分离及其在语音分离中的应用

Underdetermined Blind Separation of Non-sparse Signals and Its Application in Speech Separation

本文提出一种基于马尔科夫链蒙特卡洛方法(MCMC)的贝叶斯非稀疏盲源分离算法。用广义高斯分布(GGD)来拟合源信号的分布,通过MCMC抽样得到GGD参数和隐变量的估计,并由此得到源信号的最小均方误差估计(MMSE),解决了GGD参数估计容易陷入局部极值点、鲁棒性差的问题。根据语音信号的局部平稳性,提出基于非稀疏度评判准则的盲分离算法,用MCMC方法分离非稀疏区的语音信号,进一步提高了语音信号分离精度。仿真实验证明,本文方法改善了非稀疏信号和语音信号的分离效果,而且具有更好的鲁棒性。

A method applicable to underdetermined blind separation of non-sparse signals is proposed, which is based on the Markov chain Monte Carlo（MCMC）Bayesian framework. The generalized Gaussian distribution （GGD） is used to model the source signals distribution, the model parameters and hidden variables are estimated by MCMC sampling to obtain the least mean square error estimation （MMSE） of source signals, by which the problem is solved that the GGD parameter estimation easily falls into a local extremum point and has poor robustness. Making use of the local stationarity of speech, a blind separation method based on non-sparse judgment criterions is proposed to enhance speech separation accuracy, which separates the speech in the nonsparse zone by MCMC sampling. Computer simulation shows that the proposed method can improve separation performance of non-sparse and speech signals, it also has better robustness.