基于压缩感知的汉语语音稀疏表示研究

发布时间：2018-02-26 03:05

本文关键词： 压缩感知语音信号稀疏表示循环观测小波树模型　出处：《江南大学》2014年硕士论文　论文类型：学位论文

【摘要】：压缩感知理论是近年来信号处理领域的研究热点，它能够突破奈奎斯特采样定律的限制，实现一种全新的边压缩边采样的采样方式。而信号的稀疏表示是压缩感知理论中的重要部分，能否得到原始信号更稀疏的表示直接影响到压缩感知对信号观测值的恢复效果。本文以汉语语音信号为研究对象，分别对DCT域、线性预测分析的残差域和小波树模型三种稀疏表示方法进行研究，将其应用于语音压缩感知框架中，改善重构语音的质量，具体的研究工作如下： 1.研究了基于DCT域稀疏预处理的语音压缩感知方法。针对语音信号在DCT域的近似稀疏性导致重构误差较大的问题，提出了基于稀疏度和基于阈值两种稀疏预处理的方法提高变换域的稀疏性，预处理后的信号牺牲了部分精度但得到了绝对的稀疏性，对预处理后的信号进行压缩感知观测，仿真实验验证了预处理方法的有效性，重构语音的质量得到了提高。 2.研究了基于循环观测改进的线性预测语音压缩感知方法。针对语音信号在线性预测残差域的稀疏表示时，需要信号的线性预测系数来构造稀疏变换矩阵，从而增加了预测系数传输数据量的问题，引入循环矩阵提出将线性预测系数存入对角阵向量中构造循环矩阵，由此得到循环观测矩阵再对语音信号观测，同时提取该循环矩阵中的线性预测系数构造残差域稀疏变换矩阵，从而间接地减少线性预测系数的传输，仿真实验表明预测系数循环观测矩阵有稳定的重构性能，线性预测压缩感知方法有更好的重构效果，改进方法减少的数据量比例达到2.4%以上。 3.研究了基于小波树稀疏性适应观测的语音压缩感知方法。针对语音信号在小波树中节点数目与观测数目不匹配的问题，根据小波树节点数修改了小波树模型重构算法中的初始支撑集，在大量实验的基础上得出固定观测数下能够最佳重构的小波树节点数，对于语音在小波树中的稀疏性好的帧分配较多的观测数目，较差的分配较少的观测数，然后根据观测数目调整小波树的节点个数。仿真实验结果表明，小波树模型能保证较好的稀疏性，对不同稀疏性的语音帧采用不同观测数，并选取最佳的小波树节点数，平均重构信噪比得到一定的提高，最后与前两章方法比较了重构语音质量和重构耗时，在两者间得到一个较好的平衡。
[Abstract]:The compressed sensing theory of signal processing in recent years is a hot research field, it is able to break the laws of the Nyquist sampling limit, to achieve a new edge edge sampling method. The compression and sparse representation of signals is an important part in the theory of compressed sensing, can get a more sparse representation of the original signal directly affects the perception of compression the signal observations recovery. This paper takes Chinese speech signal as the research object, the DCT domain of linear prediction residual domain and wavelet tree model three sparse representation method is studied and applied to speech compressed sensing framework, improve the quality of the reconstructed speech, the main research is as follows:
1. the voice DCT domain sparsity pretreatment method based on compressed sensing. According to the approximate sparsity of speech signal in the lead to greater reconstruction error in DCT domain, we propose a sparse degree and threshold two methods based on sparse pretreatment to improve sparsity based on transform domain signal preprocessing after sacrifice some accuracy but the sparsity of the absolute, the preprocessed signal is compressed sensing, the simulation results verify the validity of pretreatment method, improves the quality of reconstructed speech.
2. of the linear circulation observation improved prediction method based on perceptual speech compression for sparse speech signal in the domain of linear prediction residual representation, signal of the linear prediction coefficients to construct the sparse transform matrix, thus increasing the prediction coefficient of transmission data, introducing a cyclic matrix, put forward linear prediction coefficients in constructing cycle matrix array vector, the observation matrix of speech signal cycle observation, the simultaneous extraction of the linear prediction coefficients to construct circular matrix residual domain sparse transform matrix, and thus indirectly reduce the transmission coefficient of linear prediction, simulation results show that the prediction coefficient of cyclic observation matrix reconstruction of stable performance, linear prediction method with compressed sensing reconstruction effect better, the amount of data to improve the methods to reduce the proportion of more than 2.4%.
3. research on wavelet tree sparse adaptive observations of speech compression method based on perception. For the voice signal in the wavelet tree node number and the number of observations does not match the problem, according to the number of tree nodes to modify the initial wavelet wavelet tree model reconstruction algorithm of the support set, based on a large number of experiments that the number of wavelet tree node optimal reconstruction the number of observations can be fixed, the number of observations for speech in the wavelet tree sparse good frame allocation more, less the number of observations is allocated according to the number of nodes, and then adjust the number of observed wavelet tree. Simulation results show that the wavelet tree model can guarantee the sparsity of speech frame is good, different the sparsity of the observation with different number, and select the number of wavelet tree node is the best, the average reconstruction signal-to-noise ratio can be improved, and finally the first two chapters compared with the method of reconstruction of speech quality and weight It takes time to get a better balance between the two.

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

【参考文献】