稀疏卷积非负矩阵分解的语音增强算法  

Speech Enhancement Based on Convolutive Nonnegative Matrix Factorization with Sparseness Constraints

作　　者：张立伟[1] 贾冲[1] 张雄伟[1] 闵刚[1,2] 曾理[1] 

机构地区：[1]解放军理工大学指挥信息系统学院,南京210016 [2]西安通信学院基础部,西安710106

出　　处：《数据采集与处理》2014年第2期259-264,共6页Journal of Data Acquisition and Processing

基　　金：江苏省自然科学基金(BK2012510)资助项目

摘　　要：鉴于卷积非负矩阵分解在语音增强算法中的成功应用,进一步考虑语音信号在时频域的稀疏性,提出了稀疏卷积非负矩阵分解(Sparse convolutive nonnegative matrix factorization,SCNMF)的语音增强算法。该算法包括训练和增强两个阶段。训练阶段通过SCNMF算法分别对纯净语音和噪声的频谱进行训练,得到纯净语音和噪声字典,并将其作为增强阶段的先验信息。增强阶段首先通过SCNMF算法对带噪语音的频谱进行分解,然后利用纯净语音和噪声联合字典以及相应的迭代公式对语音编码矩阵进行估计,重构增强语音。通过实验仿真分析了稀疏因子对增强语音质量的影响。实验结果表明,在非平稳噪声和低信噪比条件下,本文算法增强效果均优于多带谱减、非负矩阵分解和卷积非负矩阵分解等传统的算法。In recent years, sparse convolutive nonnegative matrix factorization（SCNMF） algo- rithm has been well used for speech enhancement. Considering the sparsity of speech signals in the frequency domain, a speech enhancement approach based on SCNMF is proposed. The ap- proach consists of a training stage and a denoising stage. During the training stage, the prior information about the spectrum of speech and noise is modeled by SCNMF algorithm and the dictionaries of speech and noise are constructed. In the denoising stage, the spectrum of noisy speech is analyzed by SCNMF algorithm, and the dictionaries of speech and noise and the itera- tive formulation are combined to evaluate the coding matrix of speech, then the enhanced speech is reconstructed. The impact of sparse factor on enhanced speech quality is analyzed through simulation experiments. Experimental results show that the proposed method outper- forms traditional speech enhancement algorithms, such as multi-band spectral subtraction （MSS）, nonnegative matrix factorization（NMF）, convolutive nonnegative matrix factorization （CNMF）, in non-stationary noise and low SNR conditions.

关 键 词：语音增强 稀疏卷积 非负矩阵 字典训练 稀疏因子 

分 类 号：TN912.3[电子电信—通信与信息系统]