Broadbanding of a NN-based microphone-array system by decomposing into frequency components

Our previous microphone-array system with a neural network (NN) structure has yielded a sharp directivity by training the NN using temporal-spatial patterns in sound pressure for sinusoidal signals at multiple frequencies. Although this system achieved a sharp directivity for trained and untrained frequencies, the directivity is effective only for sinusoidal signals. In this study, we aim broadbanding of the system by decomposing a complex input signal into frequency components. The decomposed signals are fed into the NN units prepared for every frequency component in parallel. Finally, the outputs from all units are summed up to form a desired signal. As a result, a sharp directivity is obtained for a broadband signal.