Hyrax voice elements classification using deep learning

Hyrax automatic voice elements detection is a meaningful challenge for researchers. It is a potential monitoring tool, that can assist in learning the daily and seasonally activity patterns or identify associated hyrax behaviors. Manual detection techniques, that require experts’ active listening and labeling are not efficient or scalable. This paper aims to shed light on non-human vocal communication. This research is based on a corpus of thousands of element vocal recordings of female rock hyraxes (Procavia capensis), originated in Ein Gedi, Israel. Two of us (VD and NG) performed the recordings and their manual labeling. Based on this corpus, we developed a vocal-based elements detection method, using a multilayer neural network. The current model consists of five vocal elements. The classification process achieved an F1-score of 0.94. The model can be further extended to additional species.