Passive acoustic monitoring is firmly established as an effective non-invasive technique for wildlife monitoring. The analysis of animal vocalizations recorded in their natural habitats is commonly used to monitor species occupancy, distribution mapping and community composition. The ability to distinguish between individual animals, however, remains underexplored and presents an exciting opportunity to study individual animal behavior and population demographics in more detail. In this work, we investigate bioacoustic individual-level recognition. In contrast to existing work, we focus on settings where only a subset of the existing population is known and labeled. This is crucial because wildlife populations are constantly changing so that solutions operating only within a known set of individuals are not realistically applicable in the wild. Using two novel datasets, we show that models initially trained to classify only known individuals can also be extended to detect new, previously unseen, individuals that are not part of the training set. We demonstrate that feature extractors pretrained on species classification can be successfully adapted for this task. Extending individual-level recognition to unknown individuals, so-called out-of-distribution classification, is a crucial step towards making individual recognition a realistic possibility in the wild.HighlightsWe show that features learned by models pretrained on bird species data can be transferred to individual classification tasks with minimal effort.We define and explore the out-of-distribution classification problem on individual animal vocalisations and address various subtleties and eco-logical use cases.We compile and contribute two additional datasets to facilitate further research on individual acoustic recognition.
