Robust Distributed MVDR Beamforming in Wireless Acoustic Sensor Networks With Subnetwork Selection

Wireless acoustic sensor networks (WASNs) based on the Internet of Things (IoT) have gradually showcased significant advantages in speech processing tasks. Since audio signals captured by acoustic sensors are inevitably contaminated by environmental noises, exploring effective speech enhancement techniques is essential in WASNs. To this end, a robust distributed minimum variance distortionless response (MVDR) filtering method for speech enhancement is proposed in this paper. Specifically, a novel signal model for distributed networks is established by jointly exploiting both interchannel and interframe correlations of speech. Then, a centralized MVDR filter for WASNs is presented by utilizing information from all nodes to achieve good speech enhancement performance. Next, the distributed MVDR optimization problem is proposed and converted into its dual form, and further solved in a distributed manner using the primal-dual method of multipliers. Finally, to reduce computational cost and power consumption, the most efficient subnetwork for distributed speech enhancement, which balances the input signal-to-noise ratio and transmission power, is selected via a data-driven comparison algorithm. The proposed method can effectively suppress interference noise and improve speech quality and is robust to network topology changes and sound source movement. In addition, it does not need to activate all nodes during consensus iterations, thereby mitigating the communication and computational load. Experimental results under different conditions validate the effectiveness of the proposed method.