基于卡尔曼滤波的低复杂度去混响算法*

在电话会议、智能音箱等应用场景下，传声器往往处在声源的远场。混响信号的存在会掩蔽后续到达的直达声信号，降低传声器接收信号的语音质量，以及语音识别系统的准确识别率。多通道线性预测算法是一种经典的盲去混响算法，但该算法往往具有较高的计算复杂度。本文提出了一种简化的卡尔曼滤波更新算法，通过对角化卡尔曼滤波器状态向量误差协方差矩阵，降低了自适应多通道线性预测去混响算法的复杂度。通过与现有分块对角简化算法对比发现，本文提出的简化算法在保证语音质量的同时，进一步降低了原卡尔曼滤波算法的复杂度。

Kalman filter based low-complexity dereverberation algorithm

Microphones are always far away from the speech source in the video-conference systems and intelligent loudspeakers applications. Reverberation signal will smear successive direct signal, which severely degrades the audible speech quality of the captured signals and the performance of automatic speech recognition (ASR) system. The multi-channel linear prediction (MCLP) algorithm is one of the classical blind dereverberation methods, but it suffers from high computational cost. We propose a simplified Kalman filter algorithm, which reduces the complexity of adaptive MCLP dereverberation method by diagonalizing the state error correlation matrix. Compared with the original Kalman filter, the complexity of the proposed algorithm is reduced considerably without significant performance degration.