Researchers from the UK and Tanzania have developed an innovative method for automatically classifying lion vocalizations using machine learning technologies. This approach not only clarified the structure of the lion’s roar—revealing a previously unknown element—but also significantly improved the accuracy of identifying individual animals, reaching up to 87%. Additionally, the study found that lions from different regions of Africa “speak” in distinct ways.

The Uniqueness of the Lion’s Roar

Every lion has a unique roar, which serves both to defend its territory and to communicate within the pride. Biologists use recordings of these roars to estimate population sizes. Previously, specialists had to manually listen to hours of recordings to isolate individual segments for each animal. This process was labor-intensive and depended on the expertise of the analysts, who could make mistakes or interpret the same sounds differently.

Machine Learning in Vocalization Analysis

The authors of the study, published in the journal Ecology and Evolution, automated this process with machine learning. It was previously believed that a lion’s roar consisted of three parts: introductory moans, a loud roar, and a concluding growl. However, analysis of sound patterns showed the structure is more complex: within the loud roar, an additional element was identified—an intermediary roar.

To sort the sounds, the researchers used the K-means clustering method. The algorithm, relying on just two parameters—sound duration and maximum frequency—learned to distinguish types of vocalizations without human input. The classification accuracy reached 95.4%.

Improving Identification Accuracy

For biologists, the most valuable part of the recording is the loud roar, which encodes the individuality of each animal. Manual analysis often led to confusion between true roars and similar sounds, reducing identification accuracy. The algorithm outperformed human experts: using data selected by the neural network increased the recognition accuracy of specific lions from 80% (with manual labeling) to 87%. The program applies consistent evaluation criteria and does not suffer from fatigue, eliminating human error.

Geographical Features and “Accents”

During the study, scientists discovered interesting details about lion ecology. Comparing recordings from Tanzania and Zimbabwe, they noticed differences in the frequency characteristics and duration of the sounds. This points to the existence of geographical “accents”: populations from different parts of the continent sound different. Such variation could complicate the creation of a universal program for all of Africa—a neural network trained on the “Tanzanian dialect” would require additional training to work in Zimbabwe.

The emergence of accents may be linked to the transmission of roaring skills from adults to younger lions, as well as the influence of terrain and vegetation, which affect how sound travels. Lions may unconsciously adapt their roar to local conditions to be heard over greater distances.

Social Context and Monitoring

The study also confirmed that acoustic monitoring is highly dependent on social context. In Tanzania, microphones did not record a single roar from females—local lionesses had young cubs during the research period, so the mothers remained silent to avoid attracting the attention of competitors and other predators.

Conservation Prospects

This new method transforms passive acoustic monitoring into an accessible and effective tool for wildlife conservation. Deploying autonomous microphones in the savanna is simpler and cheaper than organizing expeditions or maintaining a network of camera traps. Delegating data processing to algorithms will allow biologists to more accurately estimate predator populations across vast and hard-to-reach territories.