A method for identifying sounds used in the classification of alarm calls

Abstract

In this study, we present a methodology that identifies acoustic units in Gunnison's prairie dog alarm calls and then uses those units to classify the alarm calls and bouts according to the species of predator that was present when the calls were vocalized. While traditional methods measure specific acoustic parameters in order to describe a vocalization, our method uses the variation in the internal structure of a vocalization to define possible information structures. Using a simple representation similar to that used in human speech to identify vowel sounds, a software system was developed that uses this representation to recognize acoustic units in prairie dog alarm calls. These acoustic units are then used to classify alarm calls and their associated bouts according to the species of predator that was present when the alarm calls were vocalized. Identification of bouts with up to 100% accuracy was obtained. This work represents a first step toward revealing the details of how information is encoded in a complex nonhuman communication system. Furthermore, the techniques discussed in this paper are not restricted to a database of prairie dog alarm calls. They could be applied to any animal whose vocalizations include multiple simultaneous frequencies.