Roar Data: Redefining a lion's roar using machine learning

For territorial advertisement and intra-pride communication African lions emit a roaring bout, of which one component, is their iconic roar. The full-throated roar of a lion has recently been shown to be a unique and individually identifiable signature. At the same time, the frequency of large-scale passive acoustic monitoring surveys has increased. As such, a lion’s roar may soon become a useful tool to count individuals and estimate population density, to supplement traditional survey techniques. Currently, selecting full-throated roars is heavily dependent on expert inference and is therefore subject to human-induced bias. We propose a data-driven approach to automatically classify lions’ full-throated roars from the other vocalisations that constitute a roaring bout. By using 2- state Gaussian Hidden-Markov Models, we also demonstrate that two types of roars exist within a lion’s roaring bout – a full-throated roar, and a newly named intermediary roar - and these can be classified at an accuracy of 84.7%. We further demonstrate that using simple metrics to describe lion vocalisations - maximum frequency (Hz) and vocalisation length (s) – and K-means clustering is sufficient to classify lion call types, at a high accuracy (95.4%), and that by using data-driven predicted full-throated roars results in an improved ability to identify individuals (F1-score 0.87 vs manual full-throated roar classification 0.80). Here, we establish an easy to understand and implement process that will reduce the knowledge gap and make passive acoustic monitoring more accessible in a field currently dominated by other monitoring techniques (e.g., camera surveys), paving the way for novel research.