On the limitations of narrow 2D fan filters speech processing

Two-dimensional (2D) discrete-domain FIR narrow fan filters may be used for the selective filtering of sampled broadband 2D plane waves (PWs) on the basis of their directions of arrival (DOAs). We investigate the performance limitations of such filters for the selective filtering of 2D space-time PWs that propagate broadband speech. Employing both speech and synthetic speech, we show that, for a given computational complexity, the signal distortion of the passband speech signal is strongly dependent on the low-frequency part of the speech spectrum. It is shown that spatial undersampling is a particularly effective way of reducing computational complexity, as well as the number of sensors, and that the required speech signal may be fully recovered from a column of the 2D output image.