A lion's roar is one of the most studied animal vocalizations. Biologists have recorded it, spectrographed it, measured its propagation distance (up to 8 km), analyzed its social function (territory, mate attraction, group cohesion), and used it as a census tool. The roar was singular. Lions roar. One vocalization type, well understood.
Growcott et al. (Ecology and Evolution, 2025) applied K-means clustering to lion vocalizations recorded in Tanzania and found two distinct roar types: the known full-throated roar and a previously unrecognized “intermediary roar.” The classification used two simple features — maximum frequency and vocalization length — and achieved 95.4% accuracy. Two metrics. Two clusters. One had been invisible for the entire history of lion research.
The intermediary roar wasn't rare. It appeared consistently in roaring sequences alongside the full-throated version. The reason it was missed isn't that human ears can't hear it or that it's acoustically subtle. The reason is that the classification system had one category where it needed two. When you listen to a lion's roaring sequence with the assumption that there is one roar type, you hear one roar type. The variation between individual roars — the shorter ones, the higher-pitched ones — gets filed under “individual variation” or “warm-up” or “the lion isn't trying very hard.” The intermediary roar was absorbed into the noise around the known signal.
The machine learning system didn't know the prior literature. It had no expectation of one roar type. It just measured frequency and duration and asked: how many clusters are in this data? The answer was two. The simplest possible quantitative analysis — two dimensions, unsupervised clustering — was enough to break a classification that decades of expert listening had failed to question.
This matters for conservation. Acoustic monitoring counts lions by identifying individuals from their roars. If you're conflating two vocalization types, you're introducing systematic error into your individual identification. A lion producing both roar types might look like two different animals, or one animal's intermediary roars might be discarded as noise. The population estimate depends on the classification being right.
Growcott's point is broader: “We believe there needs to be a paradigm shift in wildlife monitoring and a large-scale change to using passive acoustic techniques.” The automated system outperformed expert listeners at individual identification. Not because it heard something humans couldn't, but because it measured what humans didn't think to count.
The structural lesson is about the relationship between classification and observation. Expert listeners are better than machines at many tasks. They can identify species in complex soundscapes, detect recording artifacts, judge behavioral context. But expertise comes with priors. The expert knows what a lion roar sounds like and fits what they hear to that template. The machine doesn't know what anything sounds like. It just clusters. And sometimes the clusters are right and the template is wrong.
Growcott, Lobora, Markham, Searle, Wahlstrom, Wijers, and Simmons, "Roar Data: Redefining a lion's roar using machine learning," Ecology and Evolution 15, e72474 (2025). University of Exeter / Tanzania Wildlife Institute.