自适应语音压缩感知方法.pdfVIP

第 42 卷第 6 期 Vo1. 42 No.6 东南大学学报(自然科学版) 2012 年 11 月 JOURNAL OF SOUTHEAST UNIVERSITY (Natura1 Science Edition) Nov. 2012 doi: lO. 3969/j. issn. 1001 -0505.2012.06.001 自适应语音压缩感知方法 罗武骏陶文凤左加阔赵力 (东南大学水声信号处理教育部重点实验室，南京 210096) 摘要:针对固定正交基下语音信号稀疏化程度低、适应性差的问题，提出了一种自适应的语音稀 疏化方法，并将其应用到语音压缩感知理论中.该方法首先采用线性预测系数的加权线性组合对 语音信号进行线性预测，并以线性预测残差基作为信号基.然后，按照稀疏约束条件训练出稀疏 表示的过完备字典，并交替应用 1- 范数稀疏约束的追踪和奇异值分解算法，达到字典与稀疏系 数同步更新.该方法从信号特征入手，学习并提取特征或纹理信息，能较好地实现语音信号的稀 疏化，提高语音压缩感知的重构性能.实验结果显示，与其他正交基方法相比，该方法的语音稀疏 化程度高.语音质量的主客观评价结果显示，该方法具有良好的重构性能， 关键词:压缩感知;稀疏性;语音;线性预测 中图分类号: TN912 文献标志码:A 文章编号: 1001 - 0505 (2012 )06-1027 -04 Adaptive compressed sensing method for speech La Vu Tuan Dao Van Phuong Zuo Jiak:uo Zhao Li (Key Laboratory of Underwater Acoustic Signa1 Processing 0 1 Ministry of Education. Sou由east University , Nanjing 21∞96 ， China) Abstract: To overcome the problem that the method of sparsification for sp臼ch signal based on fixed orthogonal base has a low sparsity and is not adaptive , a new adaptive sparsification algorithm is de- veloped for speech signal compression. First , sp四ch signal is predicted by linear predication using weighted linear combination of linear predictive coefficients , and the linear prediction residual are used as the signal bases. 白len ， the adaptive training dictionary is trained under the sparsity constraint , and the dictionary and sp缸sity coefficients are updated by altematively using l-norm sparsity constraint p盯suit and singu1ar value decomposition (SVD) algorithm. By analyzing the feature of speech sig- nals , the new scheme can exactly extract