高噪声环境下的生成对抗网络人机语音增强技术

Man-Machine Speech Enhancement Technology for Generating Countermeasure Network in High Noise Environment

在复杂环境下,传统的语音增强技术存在泛化能力弱、性能表现不足等缺点。近年来,生成对抗网络技术在语音信号处理领域有着重大突破。通过改进传统的生成对抗网络模型,提出了基于深度完全卷积生成对抗网络的高噪声环境下人机语音增强方法。该方法将语音信号语谱图作为生成器输入,判别器根据纯净语音信号指导生成器生成高质量的语音信号,滤除噪声信号。实验表明,通过语谱图和客观质量评分评估,可以发现所提方法可以明显改善语音质量,减少语音失真,增强系统的鲁棒性。

The traditional speech enhancement technologies have the disadvantages of weak generalization ability and insuffi cient performance in complex environments.In recent years,generative adversarial networks(GAN)have a very promising future in the fi eld of speech enhancement.Therefore,this paper proposes a human-machine speech enhancement technology based on deep full convolutional generative adversarial networks(DFCNN-GAN)by significantly improving the traditional GAN model.Specifically,the proposed method uses the speech spectrum of the speech signals as an input of the generator,and the discriminator guides the generator to generate a highquality speech signal according to clean speech signal.Through evaluating the spectrogram and objective quality score,experiments show that the proposed method can improve the speech quality signifi cantly,reduce the speech distortion,and enhance the robustness of the system.