Blind Speech Separation Employing Laplacian Normal Mixture Distribution Model

Careful choice of nonlinear function is necessary to obtain good performance from algorithms for blind source separation. In this paper, we propose a fast approach to perform blind speech separation based on natural gradient. The main ingredient is the use of a novel nonlinear function, which is accordant to the true PDF of speech signals. By appropriately choosing the shape parameter, we approximate a Laplacian normal mixture distribution to the source's PDF in time domain, then a new form of nonlinear function more suitable for speech separation is derived using the given distribution model. Simulation results indicate the good convergence and steady-state performance of our proposed method.