Passive acoustic monitoring of marine mammals faces fundamental challenges when multiple animals vocalize simultaneously, generating complex mixtures of overlapping direct and reflected signals across hydrophone arrays. We present a location-informed source separation approach that leverages 3-D tracked animal positions to extract individual vocalizations from multi-channel multi-source recordings. We extend our MAMBAT (Multiple-Animal Model-Based Acoustic Tracking) framework beyond localization: After using multi-target Bayesian methods and sound propagation models to localize and track all vocalizing animals, we apply these positions in reverse to separate individual vocalizations from the recorded mixtures. Using the 3-D positions and sound propagation model, we compute time-of-arrivals for each source at each hydrophone, calculate relative delays with respect to a reference sensor, and shift binary maps from MAMBAT accordingly. Consensus-based fusion across multiple sensors isolates direct-path signals from individual sources while eliminating multipath arrivals. We demonstrate the method on killer whale whistles recorded at the US PMRF Navy range, where manual inspection of surface-reflected delays confirms successful separation of multiple simultaneous vocalizers. Our approach handles a priori unknown numbers of animals and provides clean, individual records of vocalizations that could enable behavioral studies and abundance estimation for odontocetes. Work supported by the ONR Marine Mammals and Biology program.
Gruden et al. (Wed,) studied this question.