Identifying Distinguishing Acoustic Features in Felid Vocalizations Based on Call Type and Species Classification

The cat family Felidae is one of the most successful carnivore lineages today. However, the study of acoustic communication between felids remains a challenge due to the lack of fossils, the limited availability of audio recordings because of their largely solitary and secretive behaviour, and the underdevelopment of computational models and methods needed to address these questions. This study attempts to develop a machine learning-based approach which can be used to identify acoustic features that distinguish felid call types and species from one another through the optimization of classification tasks on these call types and species. A felid call dataset was developed by extracting audio clips from diverse sources. Due to the limited availability of samples, this study focused on the Pantherinae subfamily. The audio clips were manually annotated for call type and species. Time–frequency features were then extracted from the dataset. Finally, several multi-class classification algorithms were applied to the resulting data for classifying species and call types. We found that duration, mean mel spectrogram, frequency range, and amplitude range were among the most distinguishing features for the classifications.