Exploring within- and across-individual variation of the bonobo vocal repertoire with state-of-the-art classification approaches

Abstract

Vocalizations in bonobos (Pan paniscus) present within- and across-individual variation along a graded repertoire, which makes difficult for us to understand how information in encoded. Recent studies have shown that these vocalizations convey information on the emitter’s identity, with varying reliability according to the degree of arousal induced by the production context. Still, whether this individual signature is stable from one vocalization type to another is unknown. Through a fine-grained acoustic analysis of 1,300+ productions by ten captive bonobos from Apenheul Zoo, The Netherlands, and Planckendael Zoo, Belgium, we assessed the reliability and consistency of individual signature across the repertoire with state-of-the-art machine learning approaches (Support Vector Machines, Extreme gradient boosting and Neural networks). We also compared three parameter sets of manual – mostly durations and spectral slopes – and automatic features (Mel-Cepstral coefficients and vocalization shape modelling with Discrete-Cosinus-Transform). First, we show that while the shortest vocalizations (peep) occupy a distinctive area of the acoustic space, overlaps between the other categories are frequent across the individuals, revealing differences in their weighting between the temporal and spectral dimensions. Secondly, automatic classification confirms that reliable information on the emitter is present. Moreover, by training classifiers on short vocalization types and testing them on longer types, we show that individual signature remains stable across the repertoire. Thirdly, we showed that different parameter sets can successfully be combined to improve the classification performances. Together, these results suggest the existence of reliable idiolectal differences that can be exploited by the bonobos in their social interactions.