Non-linear beamformer with long short-term memory network

Acoustic beamforming with a microphone array enables spatial filtering in a wide frequency range. It is a challenging issue to sharpen the main-lobe in the lower frequency region with a small-scale microphone array, of which the number and spacing of microphones are small. A neural network-based non-linear beamformer achieves a breakthrough in sharpening the main-lobe. The non-linear beamforming works well for the narrowband signals but is weak in wideband beamforming. The non-linear beamforming with the long short-term memory is proposed to deal with wideband speech signals. The long short-term memory network is trained in the recurrent neural network architecture with the sequence of audio data such as speech signals. The performance of the proposed beamformer is confirmed using a small-scale 8-ch MEMS microphone array, where eight microphones are linearly arranged with the neighboring spacing of 10 mm, under a real environment. The beam-pattern of the proposed non-linear beamformer succeeds in sharpening the main-lobe although the linear delay-and-sum beamformer could not achieve frequency selectivity. The feasibility of the proposed beamformer is also confirmed in speech enhancement.