基于字典学习的LSP参数稀疏表示及性能分析  

Sparse representation and performance analysis for LSP parameters via dictionary learning

作　　者：闵刚[1,2] 张雄伟[1] 杨吉斌[1] 陈砚圃[2] 

机构地区：[1]解放军理工大学指挥信息系统学院,江苏南京210007 [2]西安通信学院,陕西西安710106

出　　处：《解放军理工大学学报（自然科学版）》2014年第2期121-126,共6页Journal of PLA University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61072125);江苏省自然科学基金资助项目(BK2012510)

摘　　要：为了研究LSP的稀疏表示方法,高效量化LSP参数,基于字典学习对LSP参数进行稀疏表示,并采用MOD和K-SVD算法训练参数字典,以平均谱失真和均方根误差为准则,通过仿真实验分析了算法的有效性,得出了字典学习时的稀疏度、原子个数等关键参数选取的原则。对比训练和测试LSP参数均方根误差性能曲线发现:随着稀疏度的增加,LSP参数字典外推能力增强,对训练集外参数稀疏表示性能恶化逐步减弱。To achieve the sparse representation of line spectrum pair（LSP） parameters and quantize the LSP parameters efficiently,the sparse representation of LSP parameters was studied based on dictionary learning while the dictionary was learned by MOD and K-SVD algorithm. Experimental results show that the algorithm is effective via the ASDM and RMSE criteria. The principle for choosing the key parameters such as sparsity and the number of atoms was also derived. Comparing the RMSE curve of the training and test LSP parameters, it is found that the extrapolation performance for LSP dictionary is improved and the degrading performance for outside the training LSP set is decreased gradually with the increase of sparsity.

关 键 词：线谱对 字典学习 稀疏表示 语音编码 

分 类 号：TN912.3[电子电信—通信与信息系统]