数字多媒体内容安全的研究对于确保公共秩序、维护司法公正，乃至保障国家安全都具有重要意义，数字语音被动取证是其中亟待解决的难点问题之一。本项目围绕数字语音被动取证研究中存在的载体失配、未知伪造操作无法感知等关键科学问题，拟从剖析伪造语音与自然语音的本质差异为切入点，建立“基础取证特征+深度学习”的数字语音被动取证研究框架。在构建数字语音被动取证标准数据库的基础上，研究基础取证特征的跨域提取和规整方法。充分挖掘深度学习在特征自学习方面的能力，研究基于瓶颈特征的被动取证新方法。构建基于卷积深度置信网络的深度学习模型，探索一类适用于数字语音被动取证的新方法。力求在已有研究工作的基础上，形成新的研究思路，为解决数字语音被动取证难题探索出一套新的理论和技术方案。

The security of digital multimedia has great influence on public safety, judicial justice and even national security. Passive forensic for digital speech is still a challenge problem need to be solved. This project aims to solve these key problems: cover source mismatch, unknown forgery detection. We will analyze the essential difference between the spoofing and nature speeches and establish a novel passive speech forensic framework based on “Basic forensic feature + Deep learning”. A new standard database for passive speech forensic will be constructed and the methods of cross-domain feature extraction and regulation will be proposed. We will study the novel speech forensic algorithm based on bottleneck feature by exploring the self-learning ability of deep learning. Moreover, a new class of passive speech forensic algorithms will be proposed based on the convolutional deep belief network. We hope that we can make some new research ideas and explore a new technical scheme for solving the key problems of passive speech forensic.