摘要
Wasserstein距离生成对抗网络(Wasserstein Generative Adversal Network,WGAN)模型^([1])在语音增强中运用广泛,但存在梯度易爆炸、性能不稳定等问题。引入梯度惩罚(Gradient Penalty,GP)和弹性网络条件约束,并将生成器和判别器优化成深度全卷积神经网络(Deep Fully Convolutional Neural Networks,DFCNN)结构,提出一种基于DFCNN的弹性网络条件梯度惩罚(Wasserstein Conditional Generative Adversal Network Gradient Penalty,WCGAN-GP)模型。改进后的模型可以达到真实Lipschitz限制条件,提高了可控性、稳定性和特征提取能力,能更快优化训练。实验将改进后的模型与WGAN对不同噪声条件下的语音进行增强,结果证实了改进后的模型在语音增强方面的优越性。
Wasserstein generative adversal network(WGAN)model has been widely used in speech enhancement,but WGAN has problems such as gradient explosion and unstable performance.This paper introduced gradient penalty(GP)and elastic network condition constraints,and optimized the generator and discriminator into deep fully convolutional neural networks(DFCNN)structure,and proposed a kind of Wasserstein conditional gradient penalty generative adversal Elastic network(WCGAN-GP)model based on DFCNN.The improved model could reach the real Lipschitz constraints,improve the controllability,stability and feature extraction capabilities,and optimize training faster.The experiment enhanced the speech under different noise conditions with the improved model and WGAN.The results verify the superiority of the improved model in speech enhancement.
作者
许雯婷
龚晓峰
Xu Wenting;Gong Xiaofeng(College of Electrical Engineering,Sichuan University,Chengdu 610065,Sichuan,China)
出处
《计算机应用与软件》
北大核心
2024年第2期130-137,共8页
Computer Applications and Software
基金
四川省重点研发计划项目(2020YFG0051)
国家自然科学基金项目(61876114)
校企合作项目(19H1121,19H0355)。