Distributed Kalman Filtering for Speech Dereverberation and Noise Reduction in Acoustic Sensor Networks

In acoustic sensor networks (ASNs), the desired speech signal is commonly corrupted by reverberation and background noise. To solve this problem, the distributed Kalman filtering method for joint dereverberation and noise reduction is proposed in this article. To be specific, the Kalman filter with dereverberation and noise reduction is first introduced for one node in ASNs, where the multichannel linear prediction (LP) and the sidelobe-cancellation (SC) filter are employed, and both filters are jointly estimated by a single Kalman filter. Then, the local distributed Kalman filter (localDKF) for joint dereverberation and noise reduction is presented in ASNs by only exchanging the measurements among nodes, where every node based on the local observation and the neighboring interaction can obtain the speech source estimation. Finally, to enable nodes to communicate with their neighbors in an isotropic manner, the diffusion-based distributed Kalman filter (diffDKF) approach is proposed by fusing all available information among nodes. The proposed method can jointly perform dereverberation and noise reduction in a fully distributed solution by communicating only with neighboring nodes. Experimental results show the validity of the proposed method in noisy and reverberant ASNs.