Neural Optimisation of Fixed Beamformers With Flexible Geometric Constraints

Longfei Felix Yan, Weilong Huang, Thushara Abhayapala, Jinwei Feng, W. Bastiaan Kleijn

Abstract: This paper presents a novel approach to optimising fixed broadband beamformers using neural networks. The proposed neural network model for fixed beamformers allows for the optimisation of spatial filters while incorporating flexible geometric constraints. We propose a framework for the unified signal model applicable to all geometric settings and employ two heterogeneous neural networks to simultaneously optimise both the geometry and spatial filter of fixed beamforming. Furthermore, we introduce a technique called constrained naked neurons for the optimisation of spatial filters. Experimental results show that our approaches outperform conventional approaches in terms of Directivity Factor (DF) and White Noise Gain (WNG). Our study reveals the competitive performance of a circular microphone array that matches the capabilities of a concentric circular microphone array with the same number of microphones. We also validate the effectiveness of our model in a circular discal setting, where microphones can be placed arbitrarily. Given the same parameter settings, a circular discal array can be significantly better than a linear array.