A Coherence-based Clustering Method for Multichannel Speech Enhancement in Wireless Acoustic Sensor Networks

Speech enhancement constitutes a great challenge in unknown noisy environments. Many studies have addressed this problem for both single-channel and centralized multichannel cases. However, in a real-world scenario, the effect of the reverberation and interference sounds degrades the performance of the state-of-the-art methods. In this sense, speech and signal processing with wireless acoustic sensor networks (WASNs) is becoming more and more popular, since they are able to physically cover a larger space and capture more spatial information mitigating the effect of the reverberation and the interference. In this paper, we present an unsupervised clustering method to cluster the nodes in a WASN into subnetworks, which detect different speakers. Thus, each subnetwork will be interested in detecting one of the multiple speakers in the acoustic scene. The proposed node clustering is based on the estimation of the magnitude-squared coherence between microphones observations, which measures the degree of their linear dependency. Then, a nonnegative matrix factorization (NMF) based approach is developed and applied to find the optimal clustering. Simulation results show that the proposed clustering method can assign nodes into subnetworks based on the microphones observations obtaining promising results.