Lions from different regions appear to develop unique roar patterns. (Photo by Glen Curry on Unsplash)
As with humans, lion's roar “accents” can be shaped by a combination of location, social learning and genetics.
In a nutshell
- Scientists have found that lions produce two distinct roars during growling bouts: a full-throat roar (loud and individually unique) and an intermediate roar (shorter, lower-pitched versions that appear later in the sequence).
- Lions from Tanzania and Zimbabwe produce roars of varying frequencies and durations, and one male from Botswana's roar was so unusual that machine learning algorithms misclassified 83% of his vocalizations.
- AI achieved 87% accuracy in identifying specific lions from their roars alone, offering a new conservation tool that doesn't require cameras or physical capture.
African lions living in widely separated populations in Tanzania and Zimbabwe roar noticeably differently, according to research suggesting that one of nature's most iconic sounds varies by geographic location. This discovery raises intriguing questions about whether lions develop regional vocal patterns similar to human language and birdsong.
Scientists comparing the roars of lions from the two countries found that males from each population produce different vocalizations, with Tanzanian lions roaring at different frequencies and durations compared to their Zimbabwean counterparts. One male lion from Botswana made such an unusual roar that machine learning algorithms had trouble classifying him among other lions at a research site in Zimbabwe.
Results published in Ecology and evolutionarose from broader efforts to improve lion conservation monitoring using acoustic technology. Researchers from the University of Oxford and the University of Exeter installed audio recording devices in Nyerere National Park in Tanzania and analyzed existing recordings from the Bubai Valley Game Reserve in Zimbabwe to better understand the structure of lion roars.
Scientists have discovered two types of lion's roar
Lions have long been known to produce a roar that combines a variety of vocalizations, including moans, roars and grunts. However, new research shows that what scientists previously thought was one type of roar is actually made up of two distinct vocalizations: a loud roar and what researchers recently called an “intermediate roar.”
Loud roars are powerful, sustained vocalizations that reach maximum amplitude and are unique to each lion. Intermediate roars are shorter and of lower frequency and appear later in the roar sequence as the full-throated roar subsides. Using hidden Markov models, the team classified calls within bouts of roaring with an overall accuracy of about 85%, based on the shape of each roar's frequency pattern over time. A simpler method that used just two measurements for each call (its duration and peak frequency) still separated call types with greater than 90% accuracy, even if moans were excluded from the analysis.
This distinction is important because only a loud roar contains enough individual variation to reliably identify specific lions. Previous research has shown that this roar is as unique as a fingerprint: it contains acoustic signatures that allow lions to recognize each other and size up potential rivals.
When the algorithms met an outsider lion
When the researchers applied their classification system to lions in different locations, they encountered an unexpected complication. Maximum frequency and duration varied markedly between the Tanzanian and Zimbabwean populations, suggesting that lions from different regions may develop distinctive roaring patterns.
The most striking incident happened to a man. lion designated A4 in the Zimbabwe study, which is known to originate from the Tuli block in Botswana, more than 60 kilometers from the study site. When the team's algorithm tried to automatically classify its roar, it identified only five of the 30 vocalizations that human experts called loud roars. The algorithm failed, perhaps because A4 did not roar at the top of his lungs in these bouts, or because his loud roars had a lower peak frequency or shorter duration than other males in the Zimbabwean population.
Earlier observations support the possibility of geographic variation. A study conducted in 1988 by researchers P.E. Stander and J. Stander showed that lions in Etosha National Park in Namibia produce shorter roars compared to lion roars. lions are elsewhere in Africa. The current study adds quantitative evidence to these anecdotal observations.
Why Geography May Affect the Roar of Lions
Several factors may explain why a lion's roar varies depending on geographic location. Environmental characteristics such as vegetation density and terrain influence sound propagation, potentially favoring different acoustic properties. in different habitats. Lions may unconsciously adjust their vocalizations to maximize signal transmission in the local environment.
Another possibility involves social learning within prides. Young lions learn behavior from their elders, and subtle differences in roaring technique can accumulate over generations within isolated populations, much as dialects develop in human languages and whale songs. Almost 20% Lion populations include nomadic males who can disperse over 200 kilometers from their birthplace, but if these individuals develop a roaring personality during their formative years, they can carry these characteristics to new territories.
Genetic differences Between populations may also play a role, although the researchers note that this remains speculative without further studies comparing the genetics and acoustics of different lion populations.
Machine learning brings conservation to the wild
The research team developed their classification system primarily to improve acoustic monitoring of wild lion populations. Traditional methods for estimating lion numbers rely on camera traps or track counts, both of which have limitations in dense vegetation or large areas.
Because the loud roar is unique, researchers could theoretically use audio recorders to identify specific lions without physically catching or observing them. A new study found that using data-driven loud roar classification improved the balanced performance score (called F1 score) from 0.80 to 0.87 when algorithm selecting which roar to analyze, rather than having people make that selection manually.
However, geographic diversity presents a potential complication. If roaming males from distant populations produce roars with unusual acoustic characteristics, automated systems may fail to detect them or misclassify them. their vocalizations. Researchers will need to account for regional differences when developing conservation monitoring programs that cover large geographic areas.
The study deployed custom-built stand-alone recording devices called CARACAL at 50 locations in the Matambwe sector of Tanzania. These devices recorded continuously for 62 days, recording audio data that the researchers manually reviewed to identify single bouts of roaring from individual lions. The team analyzed 1,416 vocalizations from nine recording stations. in Tanzania and compared them to 1,733 vocalizations from five male lions carried by acoustic biologists in Zimbabwe.
By extracting the contour of the fundamental frequency (the lowest frequency of the sound wave), they could model the temporal pattern of different types of calls. Using hidden Markov models and K-means clustering (relatively simple machine learning techniques), the researchers successfully automated the classification of vocalizations during bouts of roaring. This approach requires minimal computational resources compared to more advanced deep learning methods, making it accessible to conservation organizations with limited technical infrastructure.
Paper notes
Restrictions
The study's conclusions about geographic variation are based on samples from only two populations separated by significant distance. The researchers note that intermediate populations between Tanzania and Zimbabwe were not sampled, leaving uncertainty as to whether the differences represent a gradual continuum or distinct regional types. Sample sizes were relatively modest, with recordings from nine stations in Tanzania and five individuals in Zimbabwe, although this reflects the difficulty of collecting acoustic data from lions from wild populations. The Zimbabwean recordings were obtained exclusively from male lions, as the female lions in this study did not produce bouts of roaring during the recording period, due to the presence of small cubs. The researchers manually classified the moans before using their automated system, bringing a dose of human judgment to the data-driven approach. One of the key findings regarding individual A4 from Botswana is based on a single lion known to have originated outside the study population, providing limited statistical power to draw conclusions about inter-population differences in roaring.
Funding and Disclosure
The lead author, JG, received funding from a doctoral training grant provided by the UKRI AI Center for Doctoral Training in Environmental Intelligence (UKRI grant number EP/S022074/1). The camera trap study received funding from the Wildlife Conservation Network's Lion Recovery Fund (TZ-RC-02), the Darwin Initiative Opportunity and Opportunity Fund (DARCC009), and WWF Germany (213/10143411). The Zimbabwean survey data analyzed in this study were collected from a previous project by Wijers et al. (2020), in which specially designed biologists adapted to eight lions were installed in Bubaye Valley Nature Reserve in November 2014. The authors declared no conflict of interest.
Publication details
The research article, “Roar Data: Redefining Lion's Roar Using Machine Learning,” was written by Jonathan Growcott, Alex Lobora, Andrew Markham, Charlotte E. Searle, Johan Wahlström, Matthew Wyers, and Benno I. Simmons. It appears in the magazine Ecology and evolutionpublished November 2025 (vol. 15, e72474). The research was carried out at several institutions: the Center for Ecology and Conservation at the University of Exeter, the Wildlife Conservation Research Unit at the University of Oxford, the Tanzania Institute of Wildlife Research, the Department of Computer Science at the University of Oxford and Lion Landscapes in Tanzania. Research permits were granted by the Tanzania Wildlife Research Institute (TAWIRI), Tanzania National Parks Authority (TANAPA) and the Commission on Science and Technology (COSTECH) under research permits 2023-780-NA-2023-879 and 2023-665-ER-2021-287. The published article is open access under the terms of the Creative Commons Attribution License, DOI: 10.1002/ece3.72474.
