基于总体最小二乘算法的平稳声信号二阶盲分离方法

Second-order statistics method based on total least-squares for blind separation of stationary sound signals

利用声源信号相互独立的已知条件,提出分离平稳声信号的总体最小二乘二阶盲分离算法,该算法在联合近似对角化相关函数矩阵的约束下估计分离矩阵,在同时考虑估计矩阵与观测信号误差的情况下给出修正模型,在修正模型总体误差最小时分离源信号。提出的算法具有无需选择非线性函数,不存在收敛到局部最小值与计算速度快的特点,分离的声源信号具有较小的失真。半消音室电机噪声分离实验验证了上述算法的有效性。

A blind separation method using second-order statistics based on total least-squares is presented for sound signal separation. The method identifies the mixing matrix through jointly approximate diagonalisation of correlation fun ction matrices. The existence of error in mixing matrix and observation is considered and the modified blind separation model is given. The source signals are recovered while the total least-squares error is minimum. The smaller distortions exist in the reconstructed signals. The proposed method doesn't select nonlinear function and possesses fast computational speed. Finally, the results of experiment analysis made in semi-anechoic chamber demonstrate the effectiveness of the presented method.