Multichannel semi-blind source separation via local Gaussian modeling for acoustic echo reduction

In this paper, we propose a semi-blind source separation method under the assumption that the original signal of a certain source is given in advance. Main purpose of the proposed method is acoustic echo reduction. The proposed method can be regarded as an optimal coupling of linear echo canceller, multichannel source separation, and nonlinear echo suppressor under a newly proposed semi-blind local Gaussian modeling. We analyze the behavior of the proposed method when phase component of the acoustic transfer function of a loudspeaker is fluctuated. It is shown that the proposed method is more robust against phase fluctuation than the conventional linear echo canceller. Furthermore, the proposed method is superior to full-rank local Gaussian modeling and a cascade method of the linear echo canceller and the full-rank local Gaussian modeling.