A microphone clustering approach for improved Generalized Inverse Beamforming formulation

The combined use of Generalized Inverse Beamforming (GIBF) and Principal Component Analysis (PCA), has demonstrated to be effective in resolving complex and partially correlated sound source distributions in noise source identification problems. However the approach still suffers of some numerical instabilities which can give rise to a wrong identification (such as: not optimal localization and not correct quantification of the actual sources or presence of ghost images in the beamforming map) when the sources are either highly correlated or characterized by a similar strength. In this paper the authors propose a microphone clustering approach aiming at reducing these phenomena on the basis that different microphones distributions in the array produce different numerical instabilities patterns while real sources are located on the same positions. In other words the effect of the presence of actual sources is not dependent on the chosen cluster of microphones, while this is the case for the ghost images or other artifacts. An inverse beamforming formulation proves to be particularly suitable for adopting such an approach because the results are less sensitive to the geometrical distribution of the microphones in the array with respect to more conventional beamforming techniques. This gives more freedom in selecting criteria for subdividing the array in clusters of microphones.