Optimal microphone subset selection for beamforming

Abstract

Microphone arrays are widely utilized in various speech-related applications. However, using all available microphones enlarges the number of filter coefficients to be estimated, thereby increasing the computational burden without benefitting the overall performance. Consequently, selecting an optimal subset of microphones is crucial for enhancing beamformer performance. This problem is inherently combinatorial and conventionally solved through greedy-based methodologies. In this paper, we propose a novel microphone subset selection problem for beamforming and reformulate the combinatorial constraints into algebraic constraints, thereby transforming the problem into a novel mixed-integer linear programming (MILP) problem. The optimal subset is derived from a multi-objective optimization problem that maximizes beamforming performance while minimizing the number of selected microphones. The branch-and-bound method is employed to guarantee global optimality. Numerical experiments demonstrate the proposed method achieves similar beamforming performance to the greedy method and genetic algorithm (GA) while utilizing fewer microphones. This makes it particularly valuable in applications where hardware scale is strictly constrained.