Online estimation of time-variant microphone utility in wireless acoustic sensor networks using single-channel signal features

Rating the usefulness of individual microphones for subsequent signal processing is a key problem in Wireless Acoustic Sensor Networks (WASNs). This contribution expands a general-purpose approach for microphone ranking and selection in WASNs to facilitate online selection in time-variant scenarios. A Kalman Filter (KF) is employed to robustly estimate the time-varying inter-channel correlation of single-channel signal features. The information contained in multiple features is efficiently combined by recursively updating the dominant singular vector of each channel-wise matrix of feature correlation coefficients to obtain the similarity of the recorded signals while accounting for all features simultaneously. Capturing the resulting pairwise similarity of microphone channels by a graph structure, the individual microphones' utility is obtained as the Fiedler vector, which is an eigenvector of the graph Laplacian. A method to resolve the inherent sign ambiguity of the Fiedler vector using the entropy of the observed microphone signals is proposed. Experiments using synthesized and recorded data demonstrate the efficacy of the proposed approach.