一种变步长的自适应特征值分解时延估计方法

在房间声场环境下基于传声器阵列的说话人定位中，时延估计算法是其中的关键步骤。与其它常用时延估计方法相比，自适应特征值分解（Adaptive Eigenvalue Decomposition，AED）时延估计算法因其优越的抗混响性能受到越来越多的关注。但在受噪声和混响干扰的语音条件下，传统的自适应特征值分解算法收敛速度较慢，对初值敏感。通过引进动量因式，提出了一种变步长的特征值分解算法，通过理论分析和仿真实验，证实了新算法的收敛性能要优于传统的特征值分解算法，节省收敛时间，使算法的整体性能有所提高。

Variable step-size AED for time delay estimation

To realize speaker localization in a room, time delay estimation （TDE） based on microphone arrays is important. Compared to other TDE algorithms, adaptive eigenvalue decomposition （AED） algorithm gets attention for it is robust under reverberation environment. However, the conventional AED algorithm has sluggish convergence when speech signal is with noise and reverberation. Through analyzing the convergence of AED and introducing a momentum factor, a variable step-size AED algorithm is proposed. Theoretical analysis and simulation experiment show that the new algorithm is superior to original AED in saving time of convergence, which improves performance of the algorithm.