Acoustic communication is fundamental to animal social behavior and ecology, yet remains challenging to study in large, mobile carnivores across expansive landscapes. Bioacoustics—the study of sounds produced by living organisms—offers a non-invasive, cost-effective approach to wildlife monitoring with applications in ecosystem assessment, species detection, and behavioral research.

Gray wolf (Canis lupus) vocalizations have been documented for decades, but recent advances in autonomous recording technology and artificial intelligence have enabled unprecedented research scale. Since 2023, we have deployed autonomous recording units (ARUs) in Yellowstone National Park, now maintaining 50 sites. We have recorded over 400,000 hours of continuous audio data, from which we identified over 8,000 wolf vocalization events. In parallel, we equipped 15 wolves from five packs with collar-mounted audio biologgers, enabling short-term, continuous recording of near-field vocalizations. Integration of biologger data with direct field observations provides ground-truth for investigating pack-specific signatures, individual identification, quantifying detection probability of wolf howls, evaluating feasibility of acoustic-based abundance estimation, and investigating functional contexts of wolf vocal behavior.

Existing annotation platforms and AI classifiers present significant challenges for large-scale bioacoustics; we are conducting comparative tests of ARU hardware, wolf howl and chorus howl classifiers, and annotation platforms to decrease costs of these types of animal behavior studies.

The scale of this growing dataset, combined with the exceptional visibility of Yellowstone's wolves and long-term demographic monitoring, provides unique opportunities to link acoustic patterns to known individuals and social structures—advancing canid communication understanding while developing scalable technologies for non-invasive wildlife monitoring, predator-livestock conflict mitigation, and wildlife crime prevention.

​

As we prepare for fully functional AI model, below are the test results of a very limited subset of data, as we seek advice from AI researchers on what we should consider. A report of a comprehensive tests against our entire dataset will be published at a later date.