基于NMF和FCRF的单通道语音分离  

Single-channel speech separation with non-negative matrix factorization and factorial conditional random fields

作　　者：李煦[1] 屠明[2] 吴超[1] 国雁萌[1] 纳跃跃[1] 付强[1] 颜永红[1] 

机构地区：[1]中国科学院声学研究所,语言声学与内容理解重点实验室,中国北京100190 [2]亚利桑那州立大学,信号分析与感知实验室,美国坦佩85281

出　　处：《清华大学学报（自然科学版）》2017年第1期84-88,共5页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金资助项目(11461141004,91120001,61271426);中国科学院战略性先导科技专项(XDA06030100,XDA06030500);国家“八六三”高技术项目(2012AA012503);中科院重点部署项目(KGZD-EW-103-2)

摘　　要：近年来,非负矩阵分解(non-negative matrix factorization,NMF)被广泛应用于单通道语音分离问题。然而,标准的NMF算法假设语音的相邻帧之间是相互独立的,不能表征语音信号的时间连续性信息。为此,该文提出了一种基于NMF和因子条件随机场(factorial conditional random field,FCRF)的语音分离算法,首先将NMF和k均值聚类结合对纯净语音的频谱结构以及时间连续性进行建模,然后利用得到的模型训练FCRF模型,进而对混合语音信号进行分离。结果表明:该算法相比没有考虑语音时间连续特性的基于NMF的算法如激活集牛顿算法(active-set Newton algorithm,ASNA),在客观指标上有明显提高。Non-negative matrix factorization （NMF） has been extensively used for single channel speech separation. However, a typical issue with the standard NMF based methods is that they assume the independency of each time frame of the speech signal and, thus, cannot model the temporal continuity of the speech signal. This paper presents an algorithm for single-channel speech separation based on NMF and the factorial conditional random field （FCRF） method. A model is developed by combining NMF with the k-means clustering method. This model can concurrently describe the spectral structure and the temporal continuity of the speech signal. Then, the model is used to train the FCRF model, which isused to separate the mixed speech signal. Tests show that this algorithm consistently improves the separation performance compared with the active-set Newton algorithm, an NMF based approach that dose not consider the temporal dynamics of the speech signal.

关 键 词：单通道语音分离 因子条件随机场 非负矩阵分解 K均值聚类 

分 类 号：TN912.3[电子电信—通信与信息系统]