Although a previous study indicated that a microphone-array system consisting of seven microphones and a neural network realizes sharp sensitivity, it is only effective for a single frequency. In this work, we propose a new system with a modified input structure. Unlike the previous system, which was trained with the spatial patterns for a single frequency, the proposed system is trained with temporal-spatial patterns of the sound pressure distributions for sinusoidal signals at multiple frequencies. Three frequencies (425, 850, and 1,700 Hz) are used for the training process of a neural network in the proposed system. A computational simulation shows that the proposed system can realize sharp sensitivity with a half width of 5° at 425–1,700 Hz including untrained frequencies. Moreover, in an examination using an amplitude-modulated (AM) or frequency-modulated (FM) wave as the input signal, the proposed system achieves a higher performance than those in the previous study.
2017 by The Acoustical Society of Japan