Application of neural network to source PSD estimation for wiener filter based array sound source enhancement

Abstract

The Wiener filter has been used as a post-filter applied to the output of beamforming, which boosts the overall performance of sound source enhancement. Since the power spectral density (PSD) of each sound source needs to be estimated to derive the Wiener filter, a previous study attempted to estimate source PSDs from the output signals of multiple beamformings using linear approximation realized by the least squares method. In this study, we propose an alternative approach to this estimation process that uses a neural network to implement the approximation by using a non-linear function. Experimental results reveal that the proposed method estimated the Wiener filter more accurately, resulting in higher source enhancement performance while reducing the distortion in the desired sound signal.