Maximum a Posteriori Trajectory Estimation for Acoustic Source Tracking

This paper presents an acoustic source tracking framework based on a combination of the Forward-Backward and Viterbi algorithms. Browsing forward and backward the lattice of states, the proposed approach is robust against faulty observations and pauses in the target speech activity, reducing the number of outlier estimates. The most likely trajectory, i.e. best sequence of states, is obtained by applying the Viterbi decoding to the state posteriors estimated using the Forward-Backward algorithm. A batch off-line implementation is adopted to limit the estimation latency. Experiments on a real data set collected on purpose and on the NIST RT-05 source localization task show the potential of the proposed approach in real reverberant environments.