一种基于非负矩阵分解的语音增强算法  

Speech Enhancement Based on Nonnegative Matrix Factorization

作　　者：隋璐瑛[1] 张雄伟[2] 黄建军[3] 董军涛[4] 

机构地区：[1]解放军理工大学指挥自动化学院研究生1队 [2]解放军理工大学指挥自动化学院信息作战系 [3]解放军理工大学指挥自动化学院研究生2队 [4]中国人民解放军73689部队

出　　处：《军事通信技术》2012年第1期18-22,30,共6页Journal of Military Communications Technology

摘　　要：文章提出了一种基于非负矩阵分解的语音增强算法。该算法包括两个阶段,训练阶段和增强阶段。训练阶段通过非负矩阵分解算法对纯净的噪声频谱进行训练,得到噪声字典矩阵,保存其作为增强阶段的先验信息。增强阶段首先通过非负矩阵分解算法对带噪语音的频谱进行分解,然后联合噪声字典矩阵和推导得到的相应迭代公式对语音字典矩阵和语音编码矩阵进行估计,重构增强语音。仿真结果表明,文中增强方案在抑制背景噪声,提高信噪比和减少语音失真方面要优于传统的语音增强算法。A speech enhancement approach based on nonnegative matrix factorization algorithm was proposed to enhance the speech contaminated by additive noise. The technique for speech denoising consists of a training stage and a denoising stage. During the training stage, the prior information about the spectrum of noise was modeled by nonnegative matrix factorization algorithm and the noisedictionary constructed. In the denoising stage, the spectrum of noisy speech was analyzed by nonnegative matrix factorization algorithm, then, the noisedictionary was combined with iterative formulation to evaluate the speechdictionary and the coding matrix of speech, and to reconstruct the enhanced speech. Experimental results show that the proposed speech enhancement project yields less residual noise and better speech quality than the traditional speech enhancement algorithm.

关 键 词：语音增强 非负矩阵分解 字典训练 迭代规则 

分 类 号：TN912.3[电子电信—通信与信息系统]