Acoustic beamforming has been one of the essential issues in multi-channel signal processing. The traditional beamformers were analytically designed for each application, and recent nonlinear beamformers employ machine learning to optimize their beampatterns. The concept of the neural network-based beamformer was established in the early 1990s, where the nonlinear beamformer trained with a single sinusoidal signal succeeds in sharpening its beampattern at a specific frequency. In practical applications such as speech interfaces, we must deal with broadband signals, including speech signals. The authors investigated the broadband optimization of the neural network-based beamformer in the speech band from 300 Hz to 3.4 kHz, but a sharp mainlobe was not achieved in the lower frequency range. Next, the recurrent neural network with long short-term memory succeeded in sharpening the beampattern, but it caused severe frequency-dependent nonlinear distortion on beamformer outputs. This paper proposes an alternative broadband optimization of nonlinear beamformers by subband decomposition. A filterbank decomposed a broadband signal into narrowband signals in each subband. It is confirmed that beamformer optimization can be performed in each subband with suitable parameter settings.
Mizumachi et al. (Wed,) studied this question.