Improved Generative Adversarial Network Method for Flight Crew Dialog Speech Enhancement

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