语音压缩感知及其重构算法  

Compressed sensing in speech and its reconstruction algorithm

作　　者：梁瑞宇[1,2] 邹采荣[1] 赵力[1] 奚吉[2] 张学武[2] 

机构地区：[1]东南大学信息科学与工程学院,南京210096 [2]河海大学计算机及信息工程学院,常州213022

出　　处：《东南大学学报（自然科学版）》2011年第1期1-5,共5页Journal of Southeast University：Natural Science Edition

基　　金：国家自然科学基金资助项目(60472058;60975017);江苏省自然科学基金资助项目(BK2008291);中央高校基本科研业务费专项资金资助项目(2009B32614)

摘　　要：在研究语音信号在小波域的稀疏性的基础上,提出双正交小波变换的方法,与一维小波变换方法相比稀疏度提高10%～25%.此外,提出基于自适应次梯度投影算法(ASPM)进行压缩感知(CS)语音信号重构的方案.ASPM算法首先根据压缩感知重构模型建立包含稀疏重构信号并具有随机属性的凸集,然后运用次梯度投影的思想将该凸集的投影转化为对多个闭合半平面的投影,最后将更新后的稀疏重构信号投影到限定集合上.同时,该算法设计了自适应调节膨胀系数的机制以获得快速收敛性.理论分析和仿真结果表明,该算法具有快速收敛性和较低的重构误差,在不同的噪声强度下具有较高的鲁棒性.Based on the research of sparseness of speech signals in wavelet domain,a method based on biorthogonal wavelet transform is presented.Compared with 1-dimension wavelet transform method,the sparseness can be improved 10% to 25% at least.Furthermore,Adaptive subgradient projection method（ASPM） is proposed in this paper for speech reconstruction in compressed sensing.Stochastic property convex set which contains the sparse reconstruction signal is established by the CS（compressed sensing） reconstruction model firstly.Then subgradient projection is adopted to convert projection onto convex sets to projection into multiple closed halfspaces.Finally,the updated sparse reconstruction signal vector is projected onto the constrained set.Meanwhile,mechanism which adaptively adjusts inflation parameter in different iterations has been designed for fast convergence.Theoretical analysis and simulation results conclude that this algorithm has fast convergence,lower reconstruction error,and exhibits higher robustness in different noise intensity.

关 键 词：次梯度投影 双正交小波 压缩感知 稀疏重构 

分 类 号：TN912[电子电信—通信与信息系统]