脉冲噪声环境下语音压缩感知系统的构建与理论研究

Compressed sensing theory, based on sampling and compression, for speech signals, is emerging in recent years. Its purpose is to construct a comprehensive compressed sensing system for speech signals, replacing the traditional speech signal processing system based on the Nyquist sampling theorem. And the robustness is a very important index for the speech compressed sensing system. However, the existing speech compressed sensing system is just robust to the white Gaussian noise and the energy-limited noise. While the system is attacked by the impulsive noise, the original speech signal cannot be effectively reconstructed. Therefore, based on the different subspace theory of noise and speech signals and the Bayesian theory, two kinds of algorithms are proposed in this project using the statistical property and the structure property of the impulsive noise and speech signals, which makes the speech compressed sensing system robust to the impulsive noise in the respect of reconstruction. Moreover, the performance of the system can be analyzed in the theory. The results of the project can further improve the robustness of the speech compressed sensing system, which can provide theoretical support and technical scheme for the development of the next generation speech compressed sensing system.

语音压缩感知理论是近年来新兴的基于语音信号的采样、压缩理论，其目的是构造完备的语音压缩感知系统，取代基于奈奎斯特采样定理的传统的语音信号处理系统。语音压缩感知系统的一个重要指标是系统的鲁棒性。而现有的语音压缩感知系统仅对高斯白噪声和有限噪声具有鲁棒性，当系统受到脉冲噪声干扰的时候，无法有效地恢复出原始语音信号。为此，本项目分别基于噪声和信号的不同子空间理论和贝叶斯理论，利用脉冲噪声和语音信号的统计特性和结构特性，构建脉冲噪声环境下的语音压缩感知系统。提出两类不同的算法模块和特殊的观测矩阵模块，使语音压缩感知系统对脉冲噪声具有鲁棒性，并且从理论上分析系统的性能。本项目的研究成果可以进一步改善语音压缩感知系统的鲁棒性，为下一代语音压缩感知系统的发展提供理论支撑和技术方案。

语音压缩感知理论是近年来新兴的基于语音信号的采样、压缩理论，其目的是构造完备的语音压缩感知系统，取代基于奈奎斯特采样定理的传统的语音信号处理系统。语音压缩感知系统的一个重要指标是系统的鲁棒性。而现有的语音压缩感知系统仅对高斯白噪声和有限噪声具有鲁棒性，当系统受到脉冲噪声干扰的时候，无法有效地恢复出原始语音信号。为此，本项目分别基于噪声和信号的不同子空间理论，利用脉冲噪声和语音信号的统计特性和结构特性，构建脉冲噪声环境下的语音压缩感知系统。基于无监督学习和深度学习理论，提出了多种自适应的字典学习算法来为语音信号构造合适的冗余字典。并且，为噪声环境下提出了多种重构算法， 并将压缩感知理论应用到回放语音检测和单通道语音增强中。本项目的研究成果可以进一步改善语音压缩感知系统的鲁棒性，为下一代语音压缩感知系统的发展提供理论支撑和技术方案