Optimization for randomly described arrays based on geometry descriptors

The large degrees of freedom of irregular microphone placements render them difficult to analyze and optimize. This paper examines beamformer performance based on statistical descriptions of irregular geometries. Four geometry descriptors are developed to capture the properties of microphone distributions showing critical impact on array performance. Based on the relations of descriptors on performance matrices, a heuristic searching and a direct cluster design method are applied to obtain optimal arrays for speech applications. Results show significant SNR enhancements for optimized arrays over comparable regular arrays. Objective functions of heuristic searching result in rapid convergence to Monte Carlo optimizations, suggesting a strong correlation between performance and proposed descriptors. By clustering mics in the hyperbola areas to generate rich entropy, arrays with superior SNR performance are directly build according to the prior knowledge of acoustic scene. The feasibility of these optimization methods has also been demonstrated in real design cases.