Variable step-size speech blind separation employing Laplacian normal mixture distribution model

Abstract

For the blind source separation algorithm, choices of nonlinear function and step-size are very important. In this paper, we propose a novel natural gradient speech blind separation algorithm. The main ingredients are the use of a variable step-size technology and a nonlinear function based on Laplacian normal mixture distribution. Simulation results indicate the proposed method ensures steady state error of algorithm and accelerates convergence speed of algorithm.