Machine learning algorithms can predict emotional valence across ungulate vocalizations.

Vocalizations can vary as a function of their context of production and provide an immediate measure of an animal's affective states. If vocal expression of emotions has been conserved throughout evolution, direct between-species comparisons using the same set of acoustic indicators should be possible. The present study used a machine learning algorithm (eXtreme Gradient Boosting [XGBoost]) to distinguish between contact calls indicating positive (pleasant) and negative (unpleasant) emotional valence, produced in various contexts by seven species of ungulates. With an accuracy of 89.49% (balanced accuracy: 83.90%), we found that the most important predictors of emotional valence were acoustic variables reflecting changes in duration, energy quartiles, fundamental frequency, and amplitude modulation. This approach is critical in the field of emotional communication, where more information is needed to reach a better understanding of the emotional origins of human language. In addition, these results can help toward the development of automated tools for animal well-being monitoring.