Distributed MVDR Beamformer Based on Distributed Steering Vector Estimation in Acoustic Sensor Network

Distributed beamformers have shown great potential for speech enhancement in acoustic sensor network. In this paper, we propose a novel distributed minimum variance distortionless response (MVDR) beamformer, where the network-wide steering vector used for generating MVDR is estimated in a distributed way. First, each node compresses the multiple speech signals into a single-channel signal and broadcasts it to other nodes. Then, each node reconstructs the new observations and estimates the corresponding part of the network-wide steering vector by using generalized eigenvalue decomposition and a matrix inversion. Finally, the updated steering vectors are compressed into a single-channel signal for computing the local filter coefficients of MVDR. The method is iteratively performed, and in each iteration, we simultaneously estimate the steering vectors and generate the asymptotically optimal centralized MVDR beamformer. Experimental results demonstrate the effectiveness of the proposed method on speech enhancement task.