基于一种改进最短路径法的欠定语音盲分离

为了在欠定条件下实现语音盲源分离，该文首先利用势函数估计源信号的个数和混叠矩阵，在估计势函数曲线时对其做平滑处理，以提高源信号个数估计的准确性.在语音信号分离部分，提出了一种改进的最短路径法，该方法对混叠信号各时频单元进行分类处理，避免了最短路径法中对每个时频单元进行矩阵求逆运算，在大大减少计算量的同时也降低了分离信号的背景噪声.最后给出了仿真实验，实验结果证明了该算法的可行性和优异性.

Underdetermined blind separation of speech based on an improved shortest path method

In order to achieve the underdetermined blind source separation of speech, the potential function is used in this paper to estimate the number of source signals and mixing matrix. Then the potential function curve is smoothed in order to estimate the number of sources more accurately. In the part of speech separation, an improved shortest path method is proposed. The method classifies the time frequency units of mixed signals to avoid the inverse calculation of matrix for each time frequency unit in the shortest path method. The method greatly reduces the quantity of computation cost and background noises. Finally, the simulation results show that the algorithm is practical and has good performance.