Machine learning reveals adaptive maternal responses to infant distress calls in wild chimpanzees

Abstract

Distress calls are an acoustically variable group of vocalisations found in many groups of animals. Their function is most likely to recruit help but it is uncertain whether caregivers can base their reaction by inferring the nature of the disturbance from the acoustics of the calls alone, or whether they must also rely on context. The current view is that distress calls represent an acoustically graded system, with no clear acoustic boundaries. However, the fact that a call system is graded does not automatically disqualify it from categorical perception. For humans, similarly, there has been an ongoing debate as to whether infant distress call variants convey discrete information about the needs of callers, suggesting that distress calls are an interesting system to study how graded systems are deployed to communicate discrete events. To investigate this, we studied distress calls in wild chimpanzees (Pan troglodytes schweinfurthii), by taking advantage of state-of-the-art machine learning to classify natural distress calls produced by infant chimpanzees in various discrete situations. We found that vocalizations could be classified reliably by their context of occurrence and distance between mother and caller and that specific acoustic features predicted maternal responses. We concluded that infant chimpanzee distress calls convey information about discrete events, which guide maternal intervention decisions by inferring meaning from highly graded acoustic features of their infant distress calls. These results suggest a reexamination of infant distress calls in other species, including humans, and of other acoustically graded systems.