Nonstationary source separation
Seungjin Choi, Oyoung Lee · 2003
Source separation is a statistical method, the goal of which is to recover mutually independent sources from their linear instantaneous mixtures without resorting to any prior knowledge. Most existing methods have been focused on stationary sources, so higher-order statistics was necessary for separation, unless sources are temporally correlated. For nonstationary sources, however, it was shown by Matsuoka, Ohya and Kawamoto (1995) that source separation could be achieved by only decorrelation (second-order statistics). In present paper, we adopt the natural gradient method of Amari (see Neural Computation, vol.10, p.251-76, 1998) and derive an efficient source separation algorithm by minimizing the cost function proposed by Matsuoka et al. The useful behavior of the proposed algorithm is demonstrated through computer simulations.