基于约束型卡尔曼滤波的最大似然无失真波束形成器  

Maximum Likelihood Distortionless Response Beamformer Based on the Constrained Kalman Filter

作　　者：王劲夫 杨飞然[1,2,3] 梁兆杰 杨军 WANG Jinfu;YANG Feiran;LIANG Zhaojie;YANG Jun(Institute of Acoustics,Chinese Academy of Sciences,Beijing 100190,China;University of Chinese Academy of Sciences,Beijing 100049,China;IOASONIC(Su zhou)Technologies Co.Ltd.,Suzhou,Jiangsu 215163,China;Shenzhen Power Supply Bureau Co.Ltd,Shenzhen,Guangdong 518000,China)

机构地区：[1]中国科学院声学研究所,北京100190 [2]中国科学院大学,北京100049 [3]中科新声(苏州)科技有限公司,江苏苏州215163 [4]深圳供电局有限公司,广东深圳518000

出　　处：《信号处理》2022年第5期938-945,共8页Journal of Signal Processing

基　　金：国家自然科学基金(62171438,11804368);中国科学院青年创新促进会基金(2018027);中国科学院声学研究所自由探索项目(QYTS202111)。

摘　　要：自适应波束形成技术可以有效地拾取高质量的语音信号。近期提出的最大似然无失真(Maximum Likelihood Distortionless Response,MLDR)波束形成器不需要盲估计噪声协方差矩阵,具有很好的应用前景。本文提出了一种基于约束型卡尔曼滤波的MLDR波束形成器并给出了其低复杂度实现算法。将MLDR波束形成器设为卡尔曼滤波器的内部状态变量,采用一阶马尔科夫过程对其建模,而卡尔曼滤波器的观测方程则由MLDR波束形成器的代价函数构成。此外,通过对角化近似,进一步地降低了波束形成器的计算复杂度。在CHiME-3数据集上的测试结果表明,所提的对角化近似的波束形成器在计算复杂度更低的情况下取得了和已有在线实现的MLDR波束形成器相近的性能。Adaptive beamforming is an effective technique for high-quality sound acquisition. The recently proposed maximum likelihood distortionless response(MLDR)beamformer is promising because it does not require an explicit noise covariance matrix as input. In this paper,based on the constrained Kalman filter,an MLDR beamformer is proposed and its low-complexity implementation is also presented. The measurement equation is constructed using the cost function of the MLDR,and the beamformer weights are described by a first-order Markov process. Additionally,a diagonal form of the constrained Kalman filter is presented to further improve the computational efficiency. Experimental results on CHiME-3 indicate that the proposed beamformer has a similar performance to the existing online MLDR beamformer,but the former is computationally more efficient.

关 键 词：自适应波束形成器 语音增强 卡尔曼滤波 

分 类 号：TN912[电子电信—通信与信息系统]