Slow and steady: universal learnable structure in animal vocalizations

Abstract

The discovery of universal structures in natural stimuli has allowed neuroscientists to make powerful normative statements about the organization of sensory systems. We sought such structure in animal vocalizations -- stereotyped sound patterns that are received, recognized, and responded to by conspecific and allospecific individuals. Across humans, macaques, and four evolutionarily divergent species of birds we found that: (a) Vocalizations have the majority of their power in low-frequency temporal modulations, (b) Most of this power arises from 9-10 ``slow'' auditory features, (c) The acoustic space spanned by the leading slow features is universal, up to a small rotation, across species, (d) The leading slow features span the same space as the features most predictive of the future time course, (e) A classifier can learn to discriminate vocalizations with high accuracy based on the 10 leading features. These results suggest a universal temporal structure in vocalizations that may be exploited in animal communication.