改进贝叶斯非负矩阵分解的语音增强算法  

Speech enhancement via improved Bayesian nonnegative matrix factorization

作　　者：张立伟[1] 张雄伟[1] 胡永刚[1] 闵刚[1] 李轶南[1] 

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

出　　处：《解放军理工大学学报（自然科学版）》2015年第5期407-412,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(61471394);江苏省自然科学青年基金资助项目(BK20140074)

摘　　要：为了进一步提高增强语音的质量,基于传统的贝叶斯非负矩阵分解语音增强算法,考虑语音帧内原子间的相关性,提出了一种新的改进贝叶斯非负矩阵分解语音增强算法。该算法可分为训练和增强2个阶段:训练阶段利用该算法分别对纯净语音和噪声进行训练,得到纯净语音和噪声字典;增强阶段利用训练得到的纯净语音和噪声字典组成的联合字典结合,计算带噪语音时变增益,并利用最小均方误差估计得到增强语音频谱,进而重构增强语音。实验结果表明,该算法的对数频谱距离值和主观语音质量评估打分均优于非负矩阵分解(NMF)和贝叶斯非负矩阵分解(BNMF)等传统的语音增强算法,特别是在低信噪比条件下,该算法增强的效果更佳。To improve the quality of the enhanced speech,a novel improved speech enhancement algorithm via traditional Bayesian nonnegative matrix factorization（BNMF）was proposed.The correlations of atoms in speech frame were considered in the improved BNMF（IBNMF）.The proposed IBNMF consists of a training stage and an enhancing stage.During the training stage,the training sets of speech and noise were analyzed by IBNMF algorithm,and the dictionaries of speech and noise constructed.In the enhancing stage,combining the dictionaries of speech and noise,the coding matrix of speech was evaluated from the spectrum of noisy speech.Then,the spectrum of enhanced speech was derived by minimum mean square error（MMSE）estimator and the enhanced speech reconstructed.Experimental results show that the scores of log spectral distance（LSD）and perceptual evaluation of speech quality（PESQ）via IBNMF are better than the traditional speech enhancement methods,such as NMF and BNMF,especially in low SNR conditions.

关 键 词：语音增强 贝叶斯非负矩阵 多元Laplace分布 

分 类 号：TN912.3[电子电信—通信与信息系统]