Multi-Channel Speech Enhancement Using Beamforming and Nullforming for Severely Adverse Drone Environment

In this paper, we present an end-to-end neural beamforming method for multi-channel speech enhancement in drone environments with severe noise levels. In flying drone environment, recording with microphones attached to the drone creates a situation where the proximity and intensity of propeller and motor noise result in a low signal-to-noise ratio (SNR) compared to the target speech. EaBNet is a Deep Neural Network (DNN) based beamforming that utilizes embedding and beamforming modules to address the inherent uninterpretability in previous end-to-end beamforming models. EaBNet utilizes spatial information for speech enhancement and seeks additional improvement through PostNet. Building upon EaBNet, we incorporate a module inspired by the structure of the Generalized Sidelobe Canceller (GSC) algorithm to estimate the spatial information of ego-noise through nullforming. Additionally, we suggest estimating the spectral features of nullforming estimation outputs in addition to the beamforming output and for the input of the PostNet to achieve higher performance. As a result, it was confirmed that performance improved through these two methods.