一种有效的语音信号盲分离方法

Effective Method of Blind Separation for Speech Signal

频域盲语音信号分离存在着排序模糊问题,提出一种基于相邻频点幅度相关和DOA估计相结合的解排序模糊方法,并且通过对一系列预处理(白化)、独立分量分析和后处理算法的优化和有机组合,很好地实现了卷积混合语音信号的盲分离。用真实录制的语音信号进行了仿真实验,恢复出来的源信号的信干比较分离之前提高了约13 dB,证明了算法的有效性。

Since the permutation problem exists in frequency domain blind separation of speech signals, a method based on the amplitude correlation of adjacent frequency points and DOA estimation are proposed for solving the problem. The blind separation of the convolution mixing speech signals is implemented by the optimization and conbination of a set of preprocess- ing method （whitening）, independent component analysis and post-processing algorithm. Simulation results with real record- ing speech signals show that the signal interference of the recovered signals can be increased by 13 dB than before. It proves the validity of the proposed method.