Comparison of superimposition and sparse models in blind source separation by multichannel Wiener filter

Multichannel Wiener filter proposed by Duong et al: can conduct underdetermined blind source separation (BSS) with low distortion. This method assumes that the observed signal is the superimposition of the multichannel source images generated from multivariate normal distributions. The covariance matrix in each time-frequency slot is estimated by an EM algorithm which treats the source images as the hidden variables. Using the estimated parameters, the source images are separated as the maximum a posteriori estimate. It is worth nothing that this method does not assume the sparseness of sources, which is usually assumed in underdetermined BSS. In this paper we investigate the effectiveness of the three attributes of Duong's method, i.e., the source image model with multivariate normal distribution, the observation model without sparseness assumption, and the source separation by multichannel Wiener filter. We newly formulate three BSS methods with the similar source image model and the different observation model assuming sparseness, and we compare them with Duong's method and the conventional binary masking. Experimental results confirmed the effectiveness of all the three attributes of Duong's method.