一种新的基于稀疏表示的单通道盲源分离算法

Novel Single Channel Blind Source Separation Algorithm Based on Sparse Representation

该文针对稀疏表示应用于单通道盲源分离中存在字典间互干扰的问题,通过在常规联合字典中引入一个新的子字典——"共同子字典",提出一种新的基于稀疏表示的单通道盲源分离算法。新的字典学习目标函数中单个源的保真度由对应子字典和共同子字典构成,共同子字典的存在可以有效避免某一源信号在其他子字典上寻求成份而带来的互干扰问题。目标函数的求解通过交替执行稀疏表示、字典更新和比例系数优化3个步骤来实现。在测试阶段,通过收集单个源所对应子字典和共同子字典上的分量可以估计出混合信号中的单个源信号,从而达到盲源分离的目的。在语音数据库上进行的对比实验发现,所提算法较传统算法和前沿算法在两个通用评价指标上最高有近1 dB的提高。

The main drawback of sparse representation based Single Channel Blind Source Separation （SCBSS） is the interference between sub-dictionaries. To alleviate this drawback, an extra sub-dictionary, named common sub-dictionary, is proposed to add into traditional union dictionary. The single source is reconstructed by linear combining sparsely activity atoms of its corresponding sub-dictionary and common sub-dictionary. The common sub-dictionary can pure discriminative information in each source＇s specified sub-dictionary since the common information different sources shared together is gathered in common sub-dictionary. The optimization of objective function involves three steps： sparse representation, dictionary updating and weight coefficients optimization, the three steps are iteratively performed for a specified number of times or until convergence. In test stage, single source separation is achieved by combining atoms in source corresponding sub-dictionary and common sub-dictionary with the sparse coefficients of single mixed signal over union dictionary. Experimental results on speech dataset show that, when compared with traditional and state of art algorithms, the proposed algorithm can improve the performance 1 dB at most.