自训练过完备字典和稀疏表示的语音增强  

Speech enhancement of self-training over-complete dictionary and sparse expression

作　　者：崔晓[1] 

机构地区：[1]郑州师范学院,河南郑州450044

出　　处：《现代电子技术》2015年第13期56-58,62,共4页Modern Electronics Technique

基　　金：郑州市普通科技攻关计划项目(141PPTGG365);河南省教育厅科学技术研究重点项目(14A510023)

摘　　要：提出的算法利用带噪信号进行训练以获得过完备字典,通过设定较大的字典训练阈值,训练过程只对语音信号进行,使得自训练字典与语音信号之间相关性较强。利用该字典和较小的阈值对语音信号进行稀疏表示,进而实现语音增强。仿真实验表明,增强后的信号表示系数稀疏度更强,增强效果在信噪比(SNR)和感知语音质量评估(PESQ)得分方面均有较大改进。In the proposed algorithm, the dictionary is trained by the signal with noise to obtain over-complete dictionary. By setting the bigger dictionary training threshold, the correlation between self-training dictionary and speech signal stronger is enhanced. The training process aims to speech signal only. The dictionary and the smaller threshold are used to conduct sparse representation of speech signal, and then speech enhancement is realized. Simulation experiment results show that the expression coefficient sparsity of the enhanced signal is stronger, the enhancement effects are greatly improved in scoring aspect of signal-to- noise ratio （SNR） and perceptual evaluation of speech quality （PESQ）.

关 键 词：正交匹配追踪 迭代阈值 字典训练 语音增强 

分 类 号：TN912.3[电子电信—通信与信息系统]