语音声学参数篡改检测的关键理论及算法研究

Rapid development in speech synthesis and deep neural networks has enabled ordinary people to tamper the acoustic parameters of speech signals without leaving perceptual clues. However, the abuse of these technologies not only increases the risk of speech tampering, but also enables more sophisticated way of tampering——acoustic parameter-based speech tampering. Since tampering motivated by malicious intends may destroy important information of speech, it is very important to investigate whether there is tampering happened to the speech signal, where the tampering is, and how to recover the speech from tampering. In the field of tampering detection, most existing algorithms focus on the detection of basic type of speech tampering, such as clipping and splicing. This research focuses on the key issues in the acoustic parameter-based speech tampering and tries to provide a theoretical basis for tampering detection. Four main algorithms will be designed: 1) feature extraction algorithm based on Primal-dual algorithm; 2) multi-category discrimination algorithm for identifying the type of tampered acoustic parameters; 3) tampering localization algorithm based on attention mechanism; 4) tampering recovering algorithm (end-to-end) based on enhanced U-Net. In addition, the loss functions and implementation framework concerning the above algorithms are also designed. This research can provide effective theoretical and technical accumulation for acoustic parameter-based tampering detection.

语音合成技术及神经网络的发展使人们可以精确的操纵语音的声学参数，生成极具自然度的合成语音。然而，滥用这些技术不但增加了语音被篡改的风险，也催生了一种更加精良隐蔽的语音篡改方式——声学参数篡改。利用声学参数篡改，不法分子可以伪造出任何人的声音，引发严重的语音安全问题。由于声学参数篡改在篡改方式上更加精良复杂，篡改后的语音也更加流畅、清晰，检测此类篡改存在很大的技术难度。本课题重点研究声学参数篡改检测中的关键问题，包括如何提取声学参数篡改特征并在语音中定位声学参数篡改；如何区分不同声学参数篡改的独特属性以识别篡改类型；如何依据篡改类型及原始语音特点对篡改区域进行恢复。在算法方面，课题将研究基于近邻算子的篡改特征抽取算法，基于多分类判别器的篡改特征匹配算法，基于注意力机制的篡改定位算法，以及基于增强U-Net的篡改恢复算法。本课题工作将为声学参数篡改检测提供新的理论和技术积累。

语音合成技术及神经网络的发展使人们可以精确的操纵语音的声学参数，生成极具自然度的合成语音。然而，滥用这些技术不但增加了语音被篡改的风险，也催生了一种更加精良隐蔽的语音篡改方式——声学参数篡改。由于声学参数篡改在篡改方式上更加精良复杂，篡改后的语音也更加流畅、清晰，检测此类篡改存在很大的技术难度。本课题重点研究了声学参数篡改检测中的关键问题，包括篡改特征提取、篡改特征识别、篡改定位、以及篡改恢复等内容。本项目的研究成果已经发表多篇国际权威期刊和会议上，完成任务书的全部要求。本课题工作为声学参数篡改检测提供新的理论和技术积累。