语音伪装的安全问题研究

The development of audio/speech processing algorithms and tools has been gaining a broader application for audio/speech as well as some emerging security problems. Among them, it is the voice disugise that is open and urgent to solve. Exsiting researches and crime records have shown that with voice disguise, it is easy to hide speaker’s identity and to deceive human beings and speaker recogntition systems, and thus it presents serious threats to personal, econmic and even military security. However, there is still a lack of researches on the countermeasures of voice disguise. Hence, in this project, we are to conduct the following researches including the detection of voice transformation (VT) disguise, voice conversion (VC) disguise and replay disguise, and the speaker recognition systems robust to VT and VC. The change rules introduced by voice disguise to signals will be studied systematically, based on which the models of voice disguise will be built. Combined with the modles, the techniques of speaker recognition and machine learning methods such as deep learning and support vector machine will be employed to implement the novel detection and recognition methods mentioned above. These researches are to provide with theoretical and technical supports to social security and thus sovle the relevant problems.

随着语音处理算法及相应实现工具的成熟，语音在获得更广阔应用的同时，亦面临不少新的安全问题。其中，语音伪装便是亟待解决的重要安全问题之一。已有的研究及犯罪记录均表明：语音伪装能轻易骗过人耳及说话人识别系统，从而隐藏说话人身份，对人身、经济乃至军事等领域均带来严重威胁。然而，目前国内外的相关研究仍然不足，有效的解决方法仍然缺乏。为此，本项目将针对语音伪装的安全问题展开研究，具体内容包括：对语音变形伪装的检测及对变形伪装鲁棒的说话人识别方法、对语音转换伪装的检测及对转换伪装鲁棒的说话人识别方法，以及重播伪装的检测方法。本项目将全面研究各类伪装为信号引入的变化规律，以此进行建模，并综合说话人识别方法以及深度学习、支持向量机等机器学习方法，实现以上具有创新性的检测和识别方法。通过本项目的研究，能够为目前尚未解决的语音伪装安全问题提供理论及技术支持，解决它们所带来的威胁公共安全的问题。

随着语音处理算法及相应实现工具的成熟，语音在获得更广阔应用的同时，亦面临不少新的安全问题。其中，语音伪装便是亟待解决的重要安全问题之一。已有的研究及犯罪记录均表明：语音伪装能轻易骗过人耳及说话人识别系统，从而隐藏说话人身份，对人身、经济乃至军事等领域均带来严重威胁。然而，目前国内外的相关研究仍然不足，有效的解决方法仍然缺乏。为此，本项目针对语音伪装的安全问题展开研究，具体内容包括：对语音变形伪装的检测及对变形伪装鲁棒的说话人识别方法、对语音转换伪装的检测及对转换伪装鲁棒的说话人识别方法，以及重播伪装的检测方法。本项目全面研究各类伪装为信号引入的变化规律，以此进行建模，并综合说话人识别方法以及深度学习、支持向量机等机器学习方法，实现了以上具有创新性的检测和识别方法。通过本项目的研究，能够为目前尚未解决的语音伪装安全问题提供理论及技术支持，解决它们所带来的威胁公共安全的问题。