多点同步激光侦听系统语音重构理论研究及实验研究

Due to advantages in strong concealment and long-distance voice detection, laser listening systems have very important potential applications for national security purpose. Our recent experimental study reveals that there are two critical problems resulting in both decreasing speech quality and reducing speech intelligibility for multi-point synchronous laser listening systems, where one is missing middle and high frequency components for the speech signal and the other is the noise and the reverberation in adverse environments. To solve these two problems, we need to study in three scientific fields. First, to physically model the system, we introduce a statistical acoustic model to study the impact of different types of sound sources on both the acoustic transfer function (ATF) and the vibration transfer function (VTF). This can give us some clear guidelines on choosing detection objects and developing suitable speech processing methods. Second, to reduce the noise and the reverberation in a robust way, we study how to calibrate the non-acoustic sensor array automatically and propose new speech dereverberation and new noise reduction algorithms under mismatched signal models. We further study the robustness and the stability of these new proposed algorithms by using experimental data. Third, in order to dramatically improve both speech quality and speech intelligibility, we propose to extend the recently proposed phase-optimized K-SVD (PO-K-SVD) algorithm to reconstruct speech from the enhanced speech. To solve the phase continuity problem in the speech reconstruction stage, we focus on investigating the impact of spectral estimators and the frame length/overlap on the overcomplete dictionary of multi-channel spectra and that of instantaneous phases, respectively. This study can significantly improve both the theory and the method of multi-point synchronous laser listening systems, and we expect to make them become more applicable in our lives.

激光侦听系统具有隐蔽性强和拾音距离远的突出优点，对国家安全具有潜在的重要价值。本项目研究先期实验数据分析表明：语音中高频能量缺失和干扰噪声混响引起的语音声品质和可懂度下降是多点同步激光侦听系统的焦点问题，因此需重点解决三个科学问题。首先，针对系统的物理建模这一首要问题，引入统计声学方法研究不同特性声源的声学传递函数和振动系统响应，为优选探测目标和研究语音信号处理方法提供物理依据；其次，针对去噪去混响这一核心问题，研究信号模型失配的校准方案和稳健去噪去混响新理论和新方法，同时建立语音数据库验证去噪去混响新方法的有效性和稳定性；最后，针对语音中高频能量缺失这一关键问题，拓展基于相位优化的K-SVD理论以实现语音重构，研究谱估计对过完备幅度字典训练的影响并研究分帧策略实现连续相位。开展本项目研究将极大的完善多点同步激光侦听系统的理论和方法，推动其逐步走向实用化。

激光侦听系统具有隐蔽性强和拾音距离远的突出优点，对国家安全具有潜在的重要价值。本项目立项之前的研究先期实验数据分析表明：语音中高频能量缺失和干扰噪声混响引起的语音声品质和可懂度下降是多点同步激光侦听系统的焦点问题，因此需重点解决三个科学问题。首先，针对系统的物理建模这一首要问题；其次，针对去噪去混响这一核心问题；最后，针对语音中高频能量缺失这一关键问题。通过对这三个问题的研究，将极大的完善多点同步激光侦听系统的理论和方法，推动其逐步走向实用化。 . 本项目主要在以下三个方面进行了为期四年的项目研究：首先，研究不同特性声源的声学传递函数和振动系统响应，为优选探测目标和研究语音信号处理方法提供物理依据；其次，研究信号模型失配的校准方案和稳健去噪去混响新理论和新方法，同时建立了语音数据库验证了去噪去混响新方法的有效性和稳定性；最后，研究了基于深度学习的语音重构和去噪去混响方法。主要成果包括：1)搭建了双点激光语音侦听系统平台，可用于实验和实际应用；2) 建立了激光侦听系统大型数据库，包括语音库和噪声库；3)研究了适用于激光语音侦听系统的传统去噪去混响方法，并得到了应用；4)研究了适用于激光语音侦听系统的深度学习去噪去混响方法和语音重构方法，并得到了应用。本项目的理论研究和算法研究都在实验平台上获得了实验验证，同时发表了11篇杂志论文和6篇会议论文，其中SCI论文5篇，EI论文11篇；授权发明专利4项，还有多项专利已提交申请；算法研究成果应用于国家安全领域，得到了专家组的一致肯定。