A Study on a Two-Stage UAV Noise Removal Method Based on Deep Residual Neural Networks

Abstract

To improve speech intelligibility in UAV noise environments, this paper proposes a two-stage UAV noise removal method based on deep residual neural networks (DRNN). In the training stage, a DRNN with four residual blocks is employed to estimate the spectral gain function. In the enhancement stage, the estimated spectral gain function is applied to the noisy speech to obtain the enhanced speech signal. Comparative experimental results demonstrate that, on both the publicly available TIMIT speech dataset and a self-constructed UAV noise dataset, the proposed method consistently achieves higher average PESQ scores across all tested signal-to-noise ratio (SNR) conditions. This indicates that the enhanced speech produced by the proposed method is perceptually closer to the ideal quality as perceived by the human auditory system. Compared to traditional speech enhancement methods based on mask estimation and spectral mapping, the proposed approach more effectively removes UAV noise and improves speech quality.