基于升降编解码全卷积神经网络语音增强技术  

Speech enhancement technology based on lift codec Full Convolutional Neural Network

作　　者：孙立辉[1] 曹丽静 张竟雄 SUN Lihui;CAO Lijing;ZHANG Jingxiong(School of Information Technology,Hebei University of Economics and Business,Shijiazhuang 050061,China)

机构地区：[1]河北经贸大学信息技术学院,石家庄050061

出　　处：《智能计算机与应用》2021年第2期19-22,共4页Intelligent Computer and Applications

摘　　要：步兵战车强噪声背景下由于强背景噪声的存在,既影响了口令识别的正确率,又降低了指挥所后台监听的清晰度,为了提高语音质量,本文对口令数据进行增强处理。为此,本文提出了一种基于升降编解码全卷积神经网络(Increase Decrease Encoder Decode Convolution Neural Network,IDEDCNN)的语音增强算法,该算法将输入语音信号通过预处理,获取其傅里叶幅度谱特征,并将连续8帧的语音信号作为网络的输入,通过编码器来对相邻多帧语音信号建模以提取上下文信息,利用解码器挖掘当前待增强语音帧和上下文信息之间的联系,从而实现语音增强的目的。通过实验证明了该算法能够实现较好的语音增强效果。Due to the presence of strong background noise in the background of infantry fighting vehicles,the accuracy of password recognition is not only affected,but also the clarity of background monitoring of command post is reduced.In order to improve the voice quality,this paper carries out enhanced processing of password data.To this end,this paper puts forward a lift decoding the convolutional Neural Network(happens Decrease Encoder Decode Convolution Neural Network,IDEDCNN),which is the speech enhancement algorithm.In this algorithm,the input speech signal is preprocessed,the Fourier amplitude spectrum features are obtained,and eight adjacent frames of speech signal are taken as network input,model of adjacent frames of voice signal is modeled through the use of the encoder to extract context information.The decoder is used to mine the connection between the speech frame and the context information so as to realize the purpose of speech enhancement.Experimental results show that this algorithm can achieve better speech enhancement effect.

关 键 词：噪声估计 语音增强 全卷积神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]