鲁棒性语音压缩感知重构技术研究

发布时间：2018-10-11 10:44

【摘要】：压缩感知是一种全新的信号处理技术,它可以边采样边压缩,打破了奈奎斯特采样定理的约束。它的采样频率远低于奈奎斯特采样频率,同时实现了对信号的压缩,这大大节约了采样资源、传输带宽以及存储空间。压缩感知关键技术有三个部分:稀疏表示,观测矩阵的构建以及重构算法的设计。应用压缩感知的前提条件是信号具有稀疏性或者是可压缩的,而语音信号是近似稀疏的,所以可以应用压缩感知对语音信号进行处理。本文研究了语音与压缩感知的结合,并重点研究了语音压缩感知的鲁棒性重构算法的设计,因为鲁棒性的重构算法是压缩感知技术能否被实际应用的关键。本文的主要研究内容和创新如下:首先,本文详细介绍了压缩感知基础理论以及语音信号与压缩感知的结合,验证了语音信号的稀疏性,并且通过实验仿真讨论了现有的具有代表性的语音压缩感知的观测矩阵与重构算法的性能。然后探讨了噪声对语音压缩感知的各个部分的影响。其次,研究了一种新型的快速重构算法,它与其他的算法不同,它借助了离散余弦变换(DCT)基与确定性观测矩阵的特性,使得重构算法的复杂度大大的降低。但是,通过实验发现,这种快速重构算法对噪声的鲁棒性能不好。因此,本文提出了一种自适应快速重构算法,该算法根据输入语音信号的信噪比,自适应的选择最优的重构参数。实验仿真表明,自适应的快速重构算法具有较好的抗噪声能力,提高了语音信号的重构信噪比且重构速度也有所提升。最后,分析了前向后向追踪(FBP)算法,发现其固定了前向步长和后向步长,即每次迭代时支撑集增加的元素个数是固定的,这会导致算法的收敛速度不理想。因为在重构的过程中,残差中含有的信号分量越来越少,因此应该增大迭代的步长以加快算法的重构速度。所以,本文提出了快速的前向后向追踪(FFBP)算法,它根据两次相邻迭代的残差的变化率,动态的调整前向步长,最终提高了重构语音信号的速度。实验仿真表明,FFBP算法具有和FBP算法同等的重构信噪比,但是,FFBP算法的重构速度明显快于FBP算法。
[Abstract]:Compression sensing is a new signal processing technology, which can compress while sampling, breaking the constraint of Nyquist sampling theorem. The sampling frequency is far lower than the Nyquist sampling frequency, and the signal compression is realized, which greatly saves the sampling resources, transmission bandwidth and storage space. There are three key technologies in compressed sensing: sparse representation, the construction of observation matrix and the design of reconstruction algorithm. The precondition of compression sensing is that the signal is sparse or compressible, while the speech signal is nearly sparse, so the compression perception can be used to process the speech signal. This paper studies the combination of speech and compression perception, and focuses on the design of robust reconstruction algorithm for speech compression perception, because robust reconstruction algorithm is the key to whether compression sensing technology can be applied in practice. The main contents and innovations of this paper are as follows: firstly, the basic theory of compression perception and the combination of speech signal and compression perception are introduced in detail, which verifies the sparsity of speech signal. The performance of the existing representative speech compression sensing observation matrix and reconstruction algorithm is discussed by experimental simulation. Then the effect of noise on speech compression perception is discussed. Secondly, a new fast reconstruction algorithm is studied, which is different from other algorithms. It uses the properties of discrete cosine transform (DCT) and deterministic observation matrix to reduce the complexity of the reconstruction algorithm. However, it is found that the fast reconstruction algorithm is not robust to noise. Therefore, an adaptive fast reconstruction algorithm is proposed, which adaptively selects the optimal reconstruction parameters according to the signal-to-noise ratio of the input speech signal. The experimental results show that the adaptive fast reconstruction algorithm has better anti-noise capability and improves the signal to noise ratio of speech signal reconstruction and the reconstruction speed. Finally, the forward and backward tracking (FBP) algorithm is analyzed, and it is found that the forward step size and the backward step size are fixed, that is, the number of elements added to the support set is fixed during each iteration, which will lead to the unsatisfactory convergence rate of the algorithm. Because in the process of reconstruction, the signal components in the residuals are less and less, so the step size of the iteration should be increased to speed up the reconstruction of the algorithm. Therefore, a fast forward backward tracking (FFBP) algorithm is proposed, which dynamically adjusts the forward step size according to the change rate of the residuals of two adjacent iterations, and finally improves the speed of speech signal reconstruction. Experimental results show that the FFBP algorithm has the same SNR as the FBP algorithm, but the reconstruction speed of the FFBP algorithm is obviously faster than that of the FBP algorithm.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN912.3

【参考文献】