Nonlinear Spatial Filtering in Multichannel Speech Enhancement

The majority of multichannel speech enhancement algorithms are two-step procedures that first apply a linear spatial filter, so-called beamformer, and combine it with single-channel approach for postprocessing. However, the serial concatenation filter postfilter is not generally optimal in minimum mean square error (MMSE) sense noise distributions other than Gaussian distribution. Rather, MMSE joint spectral nonlinear function. While estimating parameters such traditional methods challenging, modern neural networks may provide an efficient way to learn function directly from data. To see if further research this direction worthwhile, work we examine potential performance benefit replacing common procedure filter. We analyze three different forms non-Gaussianity: First, evaluate on super-Gaussian high kurtosis. Second, inhomogeneous fields created by five interfering sources using two microphones, third, real-world recordings CHiME3 database. In all scenarios, considerable improvements be obtained. Most prominently, our analyses show uses available information more effectively as capable suppressing $D-1$ directional $D$-dimensional microphone array without adaptation.