Proceedings of the Annual Meeting of the Cognitive Science Society

We propose here a set of machine-learning algorithms to produce a generative low-dimensional and visually-understandablespace of the communicative repertoire of vocal species such as songbirds. As opposed to human speech, where individualelements are well defined and grounded in principled ways, the methods for defining units of animal communication sys-tems are often more varied and rely on human-centric heuristics. Using our method, we can automatically discover latentstructure in the vocal repertoire of individuals and use these to define-well principled categorical boundaries between vocalelements in communicating species. Further, we can sample from latent representations to generate novel vocal units thatcan be used to probe perceptual and physiological representations of communication. We demonstrate two use cases: (1)automated labeling of songbird vocal repertoires showing novel structure in vocal communication, and (2) a perceptualtask demonstrating that behavioral and physiological representational spaces can be biased by contextual information.GitHub.com/timsainb/AVGN