Researchers tag and track marine animals to study migration patterns, human impacts on behavior, and behavioral shifts due to climate change. Accurate data collection often relies on tagging individual animals to collect spatio-temporal state estimates of the animal’s geo-position and depth, allowing the measurement of animal motion behaviors and context. Acoustic transmitters are prominent due to their continuous communication without requiring manual retrieval or surfacing to collect data. These transmitters emit underwater acoustic pulses which can be detected by acoustic receivers, or hydrophones. However, the frequent movement of aquatic animals results in high data loss when the animal moves out of the detection range of the stationary hydrophone. Autonomous underwater vehicle (AUV) systems offer a promising solution for localizing acoustic transmitters with higher data resolution over longer periods of time. Such systems deployed in the past have often required multiple hydrophones mounted on a large frame carried by the AUV. This increases AUV drag, limiting the speed at which the AUV can track highly mobile animals such as sharks over large spatial and temporal scales. This work provides an alternative by equipping multiple AUVs with a single compact hydrophone payload, increasing the temporal and spatial resolution and accuracy, with the ability to operate both online and offline. A particle filter algorithm equipped with a hidden Markov model (HMM) behavioral motion model fuses acoustic measurements from multiple AUVs to estimate the acoustic transmitter’s position. Validation using data collected in Santa Elena Bay, Costa Rica, and Long Beach, California, shows a root mean square error (RMSE) of approximately 10 meters for short-term deployments, and a larger simulated dataset shows an RMSE of approximately 15 meters for longer deployments over a much larger geographic area. Using the particle filter with the behavioral motion model fit to historical animal movement data greatly outperforms a baseline random walk motion model. In the absence of such historical data, using the particle filter with a generic velocity motion model also outperforms the baseline model, although not as well as the behavioral motion model. This approach is reasonably robust: it is able to maintain a similar time to convergence with up to 20% of measurements lost.