Subnetwork-Specific Distributed Speech Enhancement for Multiple Sources Using Acoustic Sensor Networks

Existing subnetwork-based distributed speech enhancement (DSE) algorithms merely consider single target source scenarios and do not exploit inter-subnetwork information exchange in multi-source cases, which could be beneficial to enhancement performance. To overcome this limitation, this letter proposes a subnetwork-specific DSE (SS-DSE) method for multiple sources using an acoustic sensor network consisting of array nodes. The proposed algorithm first constructs source-specific subnetworks according to the estimated signal-to-interference-plus-noise ratio (eSINR) and MFCC-based feature consistency across local enhanced outputs of each array node, where each subnetwork aims to enhance individual desired source. Then, source-specific local enhanced signals in each subnetwork are fused to obtain preliminary enhancement for individual desired source. Finally, for each subnetwork, preliminary results exchanged from other subnetworks are employed as improved interference references for further cancellation, leading to the final subnetwork-specific output. Experimental results validate the effectiveness of our algorithm, which significantly outperforms existing distributed and subnetwork-based methods.