利用密集卷积神经网络的语音变换欺骗检测  

Detection of voice transformation spoofing using the dense convolutional neural network

作　　者：王泳 苏卓艺 朱铮宇 WANG Yong;SU Zhuoyi;ZHU Zhengyu(School of Cyberspace Security,Guangdong Polytechnic Normal University,Guangzhou 510665,China;Audio,Speech and Vision Processing Laboratory,South China University of Technology,Guangzhou 510641,China)

机构地区：[1]广东技术师范大学网络空间安全学院,广东广州510665 [2]华南理工大学音频、语音与视觉处理实验室,广东广州510641

出　　处：《西安电子科技大学学报》2021年第4期168-175,共8页Journal of Xidian University

基　　金：国家自然科学基金(61672173);广东省普通高校青年创新人才类项目(2018KQNCX140)。

摘　　要：语音变换欺骗是指利用语音处理算法改变原说话人的语音特征,从而导致说话人识别系统产生极高的错误拒绝率,达到隐藏说话人身份的目的。其实现成本低廉,并且已集成在众多的音频处理工具中,对社会安全带来严重威胁。然而,目前对于变换欺骗的检测研究仍然不足。为此,提出了一种基于密集卷积神经网络的语音变换欺骗检测方法,以区分欺骗语音和真实语音。该网络总共包含135层的网络层,通过最大化短路径地连接强化数据传输,可同时利用深层和浅层的边缘特征进行分类,抑制退化现象,从而进一步提高检测的准确率。实验结果表明,该算法对不同欺骗因子下的欺骗语音的检测准确率超过了98%。Voice transformation(VT)spoofing refers to the operations for hiding the speaker’s identity which change a speaker’s acoustic features by speech processing algorithms and result in extremely high false reject rates for automatic speaker recognition(ASR)systems.VT spoofing is implemented with a low cost and has been integrated in many audio editing tools,thus presenting serious threats to social security.However,the research on VT spoofing detection is still insufficient.Hence,in this paper we propose a dense convolutional neural network(DenseNet)based VT detection method for distinguishing spoofed voices and genuine ones.The proposed network consists of 135 layers in total.By maximizing the skip-layers,the data transmission can be enhanced,and both the deep and shallow edge features can be used for classification,so as to alleviate the degradation phenomenon and further to improve detection accuracy.Experimental results show that the detection accuracy with various spoofing factors is over 98%.

关 键 词：语音变换欺骗 安全 检测 神经网络 

分 类 号：TP39[自动化与计算机技术—计算机应用技术]