基于联合块对角化的语音盲源分离算法

Speech Blind Source Separation Algorithm based on Joint Block Diagonalization

针对语音信号的卷积混叠盲源分离问题,提出了一种时域非正交联合块对角化的改进算法。对观测信号进行滑窗处理,使其转化为瞬时混叠模型,然后利用信号的二阶统计量构造目标矩阵组,使其具有可联合块对角化结构。改进代价函数,并采用循环最小化方法进行优化,求取代价函数的最优解,得到混叠矩阵的估计,从而实现语音信号的盲源分离。与传统算法相比,该算法无需进行预白化处理,避免了白化误差的引入;无需进行频域的相关变换,使得算法更加简单;引入了非奇异性约束项,使得算法不易收敛到退化解。仿真结果验证,算法收敛速度快,且分离效果好。

A modified algorithm of time-domain non-orthogonal joint block diagonalization is proposed for convolutive mixtures blind source separation of speech signals. The observation signal is processed by sliding window, thus to convert it into instantaneous aliasing model. Then the target matrix is constructed by using the two order statistics of the signal so that it can have a joint-block diagonalization structure. The cost function is improved and optimized with the cycle minimization method. The estimation of the mixed matrix could be acquired by obtaining the best solution of cost function, thus to realize blind source separation of the speech signal. Compared with the traditional algorithm, the proposed algorithm requires no pre-whitening, thus avoiding the introduction of whitening error; no frequency domain correlation transform, thus making the algorithm simpler, and with the introduction of nonsingular constraint term, the algorithm becomes difficult in converging to the degenerate solution. Simulation results indicate that this algorithm has fast convergence speed and good separation effect.