基于环状生成对抗网络的深度语音去噪方法

Deep audio denoising by CycleGAN Network

针对基于深度学习的语音信号去噪方法存在难于收敛、性能不足的问题,本文提出了基于环状生成对抗网络的深度语音信号去噪方法,设计了新型的环状生成对抗语义去噪网络。通过40余种不同噪声语音集的试验,结果表明所提方法在5种衡量标准下都明显改善了去噪性能。

Traditional deep learning based audio denoising methods are difficult to convergence and their performances are insufficient to practical applications.This paper proposes a new audio denoising algorithms by CycleGAN,and design a new audio denoising network.By verifying the proposed method on 40 different types of audio noises,the experimental results demonstrate that the proposed method outperforms the existing methods obviously on five evaluation metrics.