面向复杂声学环境的伪装语音检测  

Speech Anti-Spoofing for Complex Acoustic Environments

作　　者：周晔 章坚武[1] 程继承 ZHOU Ye;ZHANG Jianwu;CHENG Jicheng(College of Telecommuncation Engineering,Hangzhou Dianzi University,Hangzhou Zhejiang 310018,China;Zhejiang Uniview Technologies Co.,Ltd.,Hangzhou Zhejiang 310051,China)

机构地区：[1]杭州电子科技大学通信工程学院,浙江杭州310018 [2]浙江宇视科技有限公司,浙江杭州310051

出　　处：《传感技术学报》2022年第10期1355-1362,共8页Chinese Journal of Sensors and Actuators

基　　金：国家自然科学基金项目(U1866209,61772162)。

摘　　要：随着技术更迭,最新语音合成和语音转换模型生成的伪装语音在感知上与真正语音无法区分,这严重威胁着公众的个人信息安全。现有的检测方法虽有可观进展,但仍存在检测声学环境单一、对未知欺骗攻击泛化能力差等问题。针对以上问题,提出了一种基于深度残差收缩网络(Deep Residual Shrinkage Networks,DRSN)的多特征联合语音欺骗检测方法,首先DRSN利用基于深度注意力机制的自适应阈值学习模块和软阈值模块提高了在复杂声学环境下的特征学习能力,再选取合适的声学特征构建单类特征-DRSN检测模型,最后执行多模型联合检测以实现互补,进一步提升整体性能。使用ASVspoof2019数据集的实验结果表明,相较于最佳基线系统,本方法在t-DCF和EER性能指标上分别降低47%和53%。With the upgrading of technology,the latest speech synthesis and speech conversion models can generate deceptive speech that is perceptually indistinguishable from real speech,which seriously threatens privacy security of public.Although the existing detection methods have made considerable progress,problems like simplex detection acoustic environment or poor generalization ability against unknown deception attacks are still untackled.In response to the above problems,a multi-feature fusion speech deception detection method based on Deep Residual Shrinkage Network(DRSN)is proposed.Firstly,DRSN improves the feature learning ability in complex acoustic environments through an adaptive threshold learning module which is based on deep attention mechanism,as well as a soft threshold module.Then the appropriate acoustic features are selected to build single-type feature-DRSN models.Finally,the models are jointly used for detection so as to achieve complementarity.Thereby,the overall performance of the detection method is improved.The result of detection on the ASVspoof 2019 dataset shows that the t-DCF and ERR performance indicators of the proposed method are reduced by 47%and 53%respectively,compared with the best baseline system.

关 键 词：伪装语音检测 欺骗对策 复杂声学环境 DRSN 

分 类 号：TN912.3[电子电信—通信与信息系统] TP183[电子电信—信息与通信工程]