Optimization Of Blind Speech Separation Based On Independent Component Analysis Algorithm

Independent component analysis(ICA)algorithm aims to isolate independent components from different sources in the observations.ICA is based on the non-Gaussianity and statistical independence of source signals.When compared with the traditional signal processing technology,ICA is more powerful to capture the essential structure of data.ICA is mainly used to optimize the blind speech separation in this paper.The key of ICA model is to maximize the non-Gaussianity or independence between signals.From the perspective of information theory and statistical hypothesis test,several common measurements of independence or non-Gaussianity are maximum likelihood,minimum mutual information,negentropy and nonparametric likelihood ratio etc.Generally,the objective functions above can't be directly calculated by definition,but by substituting the probability density estimation into the objective function approximation.Therefore,according to different estimating methods of probability density function,the ICA algorithm can be divided into two different types,named as parametric and nonparametric method.After introducing the definition,objective functions and optimization algorithms,the FastICA method with approximate negentropy served as objective function is emphatically discussed by the paper.FastICA is a commonly-used fast fixed-point algorithm;what's more,the assumption of distribution of transformed signals makes it a parametric ICA algorithm.With the measuring functions,the paper puts forward a series of constraint conditions and uses the variational method to get the density function which satisfies the constraints and has maximum entropy.With the assumption that the density is not far from the standardized Gaussian density,the density and negentropy approximations are then arised.The influence of measuring function for FastICA algorithm with negentropy as objective function has been further researched by this article.The ICA algorithm based on nonparametric density estimation has also been discussed.After choosing the kernel density estimation algorithm,the paper presents two kinds of ICA algorithm which respectively select the minimum mutual information(MMI)and nonparametric likelihood ratio(NLR)as objective function.The MMI is obtained from the information theory,while the NLR is drived from the hypothesis test.Both of them are measuring the independence between the transformed signals,and there exists a certain correlation between them.When using the nonparametric ICA algorithm,the selection of kernel function and smoothing parameter is discussed from the perspective of minimizing mean integrated square error and decreasing computational complexity.With MMI as objective function,the paper gives three different NMMI-ICA methods based on Epanechnikov ? Gaussian and Laplace kernels,while With NLR as objective function,the paper also gives two different NLR-ICA methods based on Epanechnikov?Gaussian kernels.What's more,the various effects on the algorithm of these kernel functions are also discussed.Finally,the article has carried simulation experiments on multiple sets of speech blind separation and analysised the influence of the kernel functions and smoothing parameter for nonparametric ICA algorithms.In addition,the comparision of parametric and nonparametric ICA algorithm is also presented.