Improved Generative Adversarial Network Method for Flight Crew Dialog Speech Enhancement

Abstract

Traditional speech enhancement algorithms are only suitable for dealing with stationary noise, but the noise in the stage of flight is nonstationary noise, so the traditional method is not suitable for dealing with the noise in the stage of flight. This paper proposes a speech enhancement algorithm based on a generative adversarial network: Deep Convolutional–Wasserstein Generative Adversarial Network (DWGAN). Firstly, the model integrates the deep convolutional generative adversarial network and the Wasserstein distance based on the generative adversarial network. Secondly, it introduces a conditional model to improve the enhanced speech quality, and the spectral constraint layer is used to prevent the model from falling too fast and causing collapse. Finally, the L1 loss term is introduced into the loss function to reduce the number of training times and further improve the enhanced speech quality. The experimental results show that the intrusiveness of background noise and overall processed speech quality of DWGAN are improved by about 7.6 and 9.4%, respectively, compared with WGAN in the acoustic environment of simulated aircraft operation.