L_(1/2)稀疏约束卷积非负矩阵分解的单通道语音增强方法  

A single-channel speech enhancement approach using convolutive non-negative matrix factorization with L_(1/2) sparse constraint

作　　者：路成 田猛[1,2] 周健[1,2] 王华彬[1] 陶亮[1] LU Cheng TIAN Meng ZHOU Jian WANG Huabin TAO Liang(Key Laboratory of Intelligent Computing and Signal Processing of Ministry of Education, Anhui University Hefei 230031 Institute of Media Computing, Anhui University Hefei 230601)

机构地区：[1]安徽大学计算智能与信号处理教育部重点实验室,合肥230031 [2]安徽大学媒体计算研究所,合肥230601

出　　处：《声学学报》2017年第3期377-384,共8页Acta Acustica

基　　金：国家自然科学基金项目(61301295;61372137);安徽大学博士科研启动经费项目;安徽省自然科学基金项目(1708085MF151)资助

摘　　要：为了刻画语音信号帧间相关性和使用更少的语音基表示语音特征,提出一种采用L_(1/2)稀疏约束的卷积非负矩阵分解方法进行单通道语音增强。首先,进行噪声学习得到噪声基;然后,以噪声基为先验信息结合L_(1/2)稀疏约束卷积非负矩阵分解方法学习含噪语音中的语音基成分;最后,利用学习到的语音基和系数重建出干净语音信号。在不同噪声环境下进行的实验结果表明,本文方法优于采用L_1稀疏约束的卷积非负矩阵方法及传统的统计语音增强方法。A single-channel speech enhancement approach is presented, where a novel convolution non-negative matrix factorization algorithm with L1/2 sparse constraint is proposed, aiming at characterizing the inter-correlation of the speech signal and using less basis to present the speech signal. The noise basis is obtained firstly by training the noise, the speech basis is learnt from noisy speech by using the proposed approach combined with pre-trained noise basis. Then, the enhanced speech is reconstructed by the speech basis and its corresponding coefficients. Experimental results in different noise environments show that the proposed approach outperforms the convolution non-negative matrix factorization algorithm with L1 sparse constraint and conventional statistical speech enhancement algorithms.

关 键 词：语音增强 非负矩阵分解 含噪语音 可懂度 稀疏性 正则化 语音质量 幅度谱 无监督 帧间 

分 类 号：TN912.35[电子电信—通信与信息系统]