一种新的基于稀疏表征的二阶段欠定语音盲分离方法

提出了一种新的混叠语音盲分离方法，即在欠定的情况下基于信号的稀疏表征，通过两个阶段估计出混叠矩阵和源信号。在混叠矩阵估计阶段，利用类拉普拉斯窗口函数构造出一个新的势函数，根据基于势函数的聚类算法估计出混叠矩阵。在源估计阶段，针对1^1-范数方法的不足，提出了一种新的基于高阶统计特性的稀疏表征来进行源信号的估计——统计稀疏分量分析。仿真实验表明，和同类其他二阶段估计方法相比，本文所提方法分离结果的重构信噪比更高，分离性能也更加优越。

Two-Stage Approach to Underdetermined Blind Speech Separation Based on Sparse Representation

A novel two-stage approach to underdetermined blind speech separation is presented, which is based on sparse representation. The approach contains two stages, which estimate the mixing matrix and the source signals respectively, using sparse representation in the underdetermined case. In mixing matrix estimation stage, a new potential function is constructed by Laplacian-like window function, and a corresponding potential-function-based clustering is performed to estimate the mixing matrix. In the source recovery stage, against the disadvantage of 1^1-norm solution, we propose a new sparse representation based on high-order statistics, which is called statistically sparse component analysis （SSCA）, to estimate the sources. Simulation based on the proposed approach exhibits very good separating effect, and compared with other existing two-stage methods, the proposed approach results higher reconstruction signal-to-noise ratios （SNR） and better separation performance.