Blind Separation of Convolutive Mixed Source Signals by Using Robust Nonnegative Matrix Factorization

Abstract

Most of existing convolutive nonnegative matrix factorization algorithms are sensitive to noise and outliers. In this paper, a robust convolutive nonnegative matrix factorization algorithm for convolutive BSS is proposed. The algorithm uses the projected gradient descent method to minimize the robust statistic energy function and yields two equations updated alternatively. Unlike other nonnegative matrix factorization algorithms, the robust convolutive nonnegative matrix factorization algorithm is resistant to noise and outliers. Experimental results on convolutive blind source separation are presented to illustrate the much improved performance of the algorithm.