全空间相关矩阵广义加权预测误差算法去混响研究

Research on dereverberation by generalized weighted prediction error algorithm of full-space correlation matrix

广义加权预测误差(generalized weighted prediction error, GWPE)算法是一种有效的多通道语音去混响算法,但在麦克风间距较近时去混响效果变差。文章分析了GWPE算法在小间距麦克风阵列下的局限性,即未充分考虑麦克风信号间的空间相关性。据此,研究了全空间相关矩阵的广义加权预测误差(generalized weighted prediction error of full-space correlation matrix, GWPE-FCM)算法,该算法考虑了不同输入通道之间的空间相关性,计算复杂度比GWPE算法略增加。仿真结果表明,GWPE-FCM算法与GWPE算法相比在整体去混响方面效果更好,特别是在麦克风距离较近的情况下。全空间相关矩阵能够更准确地估计输入信号的相关性,使得GWPE-FCM算法在去混响效果上表现更出色。

The generalized weighted prediction error(GWPE)algorithm is an effective multi-channel speech dereverberation algorithm,but its performance deteriorates when the microphone spacing is small.This paper analyzes the limitations of the GWPE algorithm in small microphone arrays,i.e.,it does not fully consider the spatial correlation between microphone signals.Then the generalized weighted prediction error using full-space correlation matrix(GWPE-FCM)algorithm is introduced,which takes into account the spatial correlation between different input channels and slightly increases computational complexity in comparison with the GWPE algorithm.The simulation results show that compared to the GWPE algorithm,the GWPE-FCM algorithm performs better in overall dereverberation,especially when the microphone distance is close.PESQ increased by about 0.2 and STOI increased by about 0.1.