结合优化U⁃Net和残差神经网络的单通道语音增强算法

Single channel speech enhancement algorithm combining optimized U⁃Net and residual network

语音增强的目的是从带噪语音中恢复出干净的语音信号,为了解决现有深度神经网络中语音增强算法不稳定,语音增强效果不理想的问题,提出一种改进的U⁃Net网络与残差神经网络相结合的语音增强算法。首先,该方法构建了一个基于U⁃Net网络的端到端的语音增强模型;然后在该模型的编解码块中引入残差单元,将残差神经网络结构的跨层连接和拟合残差项应用到模型训练中,该方法更有利于恢复目标语音的细节特征信息,增强了模型训练的稳定性,提高了模型的特征提取能力和训练效率,改进后的Residual⁃U⁃Net网络模型能够实现更优的语音增强效果。仿真实验结果表明:与现有的其他几种语音增强方法相比,文中所提出的Residual⁃U⁃Net算法更有效地实现了语音增强,此外,该算法具有良好的去噪效果,进一步提高了语音信号的质量及其可懂度。

The purpose of speech enhancement is to recover the clean speech signal from the speech with noise.In order to solve the problems that the speech enhancement algorithms in the existing deep neural network(DNN)are unstable and their speech enhancement effects are not ideal,an improved speech enhancement algorithm based on the combination of U⁃Net network and residual network(ResNet)is proposed.An end⁃to⁃end speech enhancement model based on U⁃Net network is constructed,and then the residual unit is introduced into the codec block of the model,and the cross layer connection of ResNet structure and fitting residual term are applied to model training.This method is more conducive to recovering the detailed feature information of target speech,enhancing the stability of model training,and improving the feature extraction ability and training efficiency of the model.The improved Residual⁃U⁃Net network model can achieve better speech enhancement effect.The results of the simulation experiment show that,in comparison with the other existing speech enhancement algorithms,the proposed Residual⁃U⁃Net algorithm can achieve speech enhancement more effectively,has better denoising effect,and further improves the quality and intelligibility of speech signal.