Time Domain Optimization Techniques for Blind Separation of Non-Stationary Convolutive Mixed Signals

This paper aims to solve the problem of Blind Signal Separation (BSS) in a convolutive environment based on output correlation matrix diagonalization. Firstly an extension of the closed form gradient and Newton methods used by Joho and Rahbar is developed which encapsulates the more difficult convolutive mixing case. This extension is completely in the time domain and thus avoids the inherent permutation problem associated with frequency domain approaches. We also compare the performance of three commonly used algorithms including Gradient, Newton and global optimization algorithms in terms of their convergence behavior and separation performance in the instantaneous case and then the convolutive case.