Automatic acoustic identification of individual animals in the wild

Automatic acoustic identification of individuals in the wild concerns the development of machine learning methods that can identify specific individuals based on their vocalizations. Furthermore, the developed methods should be able to operate across a multitude of species and environments that mimic the unrestrained conditions of the real-world.

From the zoology field it has been shown how several species are able to produce vocalizations containing distinct characteristics that inform about the individual, these characteristics, are often referred to as acoustic signatures.

Given the variety of species (from across all taxa) that use acoustic signatures and especially the variety of vocal systems or the different communication strategies animals employ, it is expected that these signatures won't manifest in a simple consistent way, that could be detected by standard signal processing. This is the main reason why automatic systems for acoustic identification of individuals are seldom developed to work with different species, and are instead highly
specialized in the identification of individuals within a single species and often within a small group of individuals.

An important point of this topic is to address the machine learning challenges that the application of such a system in a real-world scenario bring. Not only being able to work with different species but also being able to cope with varying environment conditions, different habitats etc. In the wild we cannot impose hard boundaries regarding the species that we are going to record, the type of calls, or the individuals that will enter the targeted space. Ideally, we want our system to work equally well across all these different conditions, but it is not desirable to fine tune the model or define specific instructions for every potential situation

In recent years, deep learning methods have been applied very successfully on several bioacoustic related tasks. Especially, multi-task learning is an interesting methodology that contributes to the generalisation capabilities of the models and thus an interesting approach to be explored here.