Underdetermined Mixture Matrix Estimation Based on Neural Network and Genetic Algorithm

Abstract

This paper proposes an improved approach to estimate the underdetermined mixture matrix to improve the performance of underdetermined blind source separation (UBSS) for speech sources. This approach only use two observed signals and consider a tangent value instead of each vector of the mixture matrix for estimation. An improved clustering method based on competitive neural network and genetic algorithm is then designed to estimate these tangent values. In the proposed method, those tangent values are designed as clustering centers. The competitive neural network is used first to obtain the initial clustering centers, and genetic algorithm is applied to search for the global optimum around the initial clustering centers. Experimental results show that the tangent values of the observed vectors have better clustering characteristics, which could reduce the computational complexity for mixture matrix estimation. The improved clustering algorithm based on neural network and genetic algorithm can estimate a better mixture matrix with high precision than the general neural network clustering algorithm, and it can improve the performance of underdetermined blind signal separation.