Experimental Analysis of EM and MU Algorithms for Optimizing Full-rank Spatial Covariance Model

Abstract

Full-rank spatial covariance analysis (FCA) is based on a flexible source model, and achieves high-quality results for blind source separation. An expectation-maximization (EM) algorithm as well as a multiplicative update (MU) algorithm are known to optimize the FCA model parameters. In this paper, we first investigate the behaviors of both algorithms. We observed that the MU algorithm minimizes the FCA objective function faster than the EM algorithm, but the separation performance at the converged point is better by the EM algorithm than the MU algorithm. We found that the MU algorithm tends to push the covariance matrices towards rank deficient. To mitigate this tendency, we propose a modified FCA model where the temporal parameters are shared within a time block. Experimental results show that the modified model provides better separation performance not only by the MU algorithm but also by the EM algorithm.