基于压缩感知的语音信号压缩重构算法研究

发布时间：2018-01-04 19:12

本文关键词：基于压缩感知的语音信号压缩重构算法研究　出处：《中北大学》2014年硕士论文　论文类型：学位论文

【摘要】：传统香农采样定理要求采样率必须高于信号最高频率的两倍，就可以较好恢复原信号，虽然实现了信号的采集、压缩和恢复，但随着采集数据和频率的急剧增加，使得目前通信系统越来越难以承受。Candès等人提出的压缩感知理论很好地解决了这个难题。压缩感知将可稀疏的信号通过观测从高阶矩阵线性投影为低阶，，信号的采集和压缩在此过程同时进行，最后高概率精确地重建原始信号。压缩感知跳出了传统采样的思维模式，所以必然会改变未来信号处理的方式。首先，本文综述了压缩感知以及在语音信号处理领域的研究现状，系统概述了压缩感知数学模型，围绕信号稀疏，设计观测矩阵和选择重构算法三个关键技术进行分类比较，并且分析了设计观测矩阵的约束条件，讨论了压缩感知与香农采样的区别和联系。其次，概述了传统语音信号处理的主要过程和语音信号特征，因为语音具有良好的可压缩性，所以压缩感知理论可以实现语音信号的压缩重构。选择DCT为稀疏基并验证了在清浊音的稀疏性，最后进行实验分析，选取一段采样率为22.05K的语音信号，通过OMP算法和BP算法分别实现语音信号重构，并对恢复的语音信号进行主观和客观评价分析，最后得出结论：（1）语音信号的压缩比值和帧长对重构信号的质量都有影响；（2）BP算法重构语音质量比OMP算法的高，但恢复信号时间较长。最后，结合普通压缩感知中稀疏基、观测矩阵和重构算法的缺点引入自适应算法，利用自适应冗余字典KSVD算法、自适应观测矩阵和SAMP重构算法，提出自适应压缩感知的概念，介绍了上述算法的具体实现步骤，并分别进行仿真对比，验证了KSVD算法具有更好的稀疏性，根据每帧语音能量自适应分配观测个数，显著提高了重构语音的质量，SAMP缩减了重构信号所用时间，最后对自适应压缩感知进行仿真分析和主客观评价，对比了普通压缩感知重构信号，自适应压缩感知具有重构语音质量显著提高并且运行时间明显减少等优点，从而验证了自适应压缩感知的可行性。
[Abstract]:Two times the traditional Shannon sampling theorem requires that the sampling rate must be higher than the highest frequency of the signal, you can restore the original signal is good, although the realization of signal acquisition, compression and recovery, but with a sharp increase in data acquisition and frequency, the communication system becomes more and more difficult to bear.Cand s proposed the theory of compressed sensing is very good to resolve this problem. The compressed sensing will be sparse signal by observing from the matrix of high order linear projection for low order, signal acquisition and compression in this process at the same time, the high probability of accurately reconstruct the original signal. Compressed sensing sampling out of the traditional mode of thinking, so will change the future way of signal processing.
First of all, this paper reviews the research status of compressed sensing and processing of speech signal in the field of system overview of the compressed sensing model, on the design of the observation matrix and sparse signal, select the reconstruction algorithm of three key techniques for classification and analysis of constraints, the design of the observation matrix, the relation and difference between the compressed sensing and Shannon sampling the discussion.
Secondly, summarizes the main process and characteristics of the traditional speech signal processing of speech signal, because the speech has good compressibility, so the theory of compressed sensing technology can reconstruct the speech signal compression. Select DCT for sparse matrix and verify the sparsity in voicing, and finally the experimental analysis, select a sampling rate of speech the 22.05K signal, the speech signal reconstruction respectively through the OMP algorithm and BP algorithm, and the voice signal recovery to analyze the subjective and objective evaluation, finally draws the conclusion: (1) effect of mass ratio and frame compression of speech signal length on signal reconstruction; (2) the BP algorithm of speech quality than OMP algorithm the high, but to restore the signal for long time.
Finally, combined with the ordinary compressed sparse basis perception, observation matrix and reconstruction algorithm of the shortcomings of the adaptive algorithm is introduced, using adaptive redundant dictionary KSVD algorithm, adaptive observation matrix and SAMP reconstruction algorithm, put forward the concept of adaptive compressed sensing, introduces the specific implementation steps of the algorithm, and the simulation results were verified, sparse KSVD algorithm better, according to each frame of speech energy adaptive allocation of the number of observations, significantly improve the quality of reconstructed speech, SAMP reduced the reconstructed signal with time, finally the adaptive CS was simulated and the subjective and objective evaluation, compared with the common compressed sensing signal reconstruction, adaptive compressed sensing has significantly improved the quality of reconstructed speech and operation less time etc, which verifies the feasibility of adaptive compressed sensing.

【学位授予单位】：中北大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN912.3

【引证文献】