Global Coherence Field and Distributed Particle Filter-Based Speaker Tracking in Distributed Microphone Networks

Based on the combination of global coherence field (GCF) and distributed particle filter (DPF) a speaker tracking method is proposed for distributed microphone networks in this paper. In the distributed microphone network, each node comprises a microphone pair, and its generalized cross-correlation (GCC) function is estimated. Based on the average over all local GCC observations, a global coherence field-based pseudo-likelihood (GCF-PL) function is developed as the likelihood for a DPF. In the proposed method, all nodes share an identical particle set, and each node performs local particle filtering simultaneously. In the local particle filter, the likelihood GCF-PL for each particle weight is computed with an average consensus algorithm. With an identical particle set and the consistent estimate of GCF-PL for each particle weight, all individual nodes possess a common particle presentation for the global posterior of the speaker state, which is utilized by each node for an estimated global speaker position. Employing the GCF-PL as the likelihood for DPF, no assumption is required about the independence of nodes observations as well as observation noise statistics. Additionally, only local information exchange occurs among neighboring nodes; and finally each node has a global estimate of the speaker position. Simulation results demonstrate the validity of the proposed method.