基于二阶统计量的语音信号时域卷积盲分离算法

A Time-Domain Blind Convolutive Separation Algorithm for Speech Signals Based on Second-Order Statistic

该文针对语音信号卷积混迭模型,利用信号间近似不相关和短时平稳等二阶统计特性,根据盲分离固有的排列和尺度不定性,改进最小二乘拟合函数,将关于混迭矩阵的四次函数转化为3组待定参数的二次函数。提出优化该代价函数的非正交联合块对角化算法,分3个子步,每个子步求解一个最小二乘问题,交替估计3组待定参数,逼近代价函数最小点。与ZJBD等同类方法相比,该方法具有计算复杂度低,估计精度高且对初始参数选择不敏感等特点,可直接在时域实现语音卷积盲分离。

For blind convolutive separation of speech signals, a cost function improved from the least squares fitting function is developed based on second-order statistic, which simplifies the quartic-function with respect to the mixture matrix into three quadratic functions. The iternative approach with three sub-steps is proposed to perform the non-unitary joint block-diagonalization. In each sub-step, a closed solution is derived by minimizing the cost function associated with one parameter-group while fixing the others. Furthermore, the feature of low computational complexity is analytically proven. Compared with ZJBD, the simulations illustrate that the proposed algorithm has the merits of robust initialization selection and better estimate accuracy.