Elephants Rumble in Choruses, and the Choir Means Something

Correction, Oct. 25, 2026: An earlier version of this article misstated parts of the study it covers. It described the classifier’s performance figures as “balanced accuracy and F1”; they are weighted average precision and recall. It reported that Contact and Greeting rumbles were confused with each other without saying this applies to choruses; among single rumbles, Contact was the least confused type. It said the authors concluded the “most likely” function of convergence is logistical coordination; the paper says only that it “may be to facilitate logistical coordination,” and the earlier version’s closing claim that the chorus information is “likely about where to go next” went beyond the study. It misnamed coauthor Emmanuel Chemla as Sylvie Chemla, described the 12 to 19 hertz range as where “most of the energy” of a rumble sits rather than its fundamental frequency, and attributed to the study a superlative about the Amboseli project’s length that the study does not make. Those errors are corrected below. The opening scene is now labeled as a composite, and the article adds conservation context, the published positions of Tanzanian officials and the hunting operator involved, and the study’s limitations.

For fifty-four years, researchers in Kenya’s Amboseli have watched elephants talk. A new study listened to the part of the conversation most studies set aside: the moments when the whole family speaks at once.

The scene below is a composite, assembled from behaviors the study’s authors attribute to Contact and Greeting exchanges. It is representative, not a single recorded event.

The reunion begins before anyone can see it. Somewhere in the swamp grass of Amboseli National Park, an elephant who has drifted from her family raises her head, flaps her ears, and rumbles. Then she goes still and listens. Out of sight, several hundred meters away, someone answers.¹

The calls continue as the two close the distance. When they meet, the exchange becomes what researchers call a greeting: heads high, ears lifted and flapping fast, fluid streaming from the glands at their temples, rumbles piling on rumbles.¹ Other members of the family join in.

A person standing beside them would miss most of it. The base note of an elephant rumble sits between roughly 12 and 19 hertz, below the roughly 20-hertz floor of human hearing.¹ Up close, people hear only the harmonics stacked above it, a low purr.

Bioacoustics has mostly skipped moments like that reunion. Roughly 60% of elephant rumbles occur in choruses of two or more voices, according to long-term work in Amboseli by the biologist Joyce Poole.¹ Overlapping calls are hard to pull apart, so most acoustic studies have analyzed the clean solos instead.¹

A study published Oct. 1 in Scientific Reports takes on the overlap directly. Drawing on 902 recordings gathered in Amboseli between 1986 and 2022, a team from the Earth Species Project, ElephantVoices, and the Amboseli Trust for Elephants tested whether a chorus carries information of its own.¹ The authors are Louis Mahon, Michael Pardo, Marius Miron, Emmanuel Chemla, Petter Granli, Norah Njiraini, Cynthia Moss, Sara Keen, and Joyce Poole.

The model they built is wrong more often than it is right. Its mistakes, though, follow the contours of elephant social life. And when elephants trade rumbles in some situations, their voices drift toward one another.

African savanna elephants are listed as endangered.⁶ In the eight months before late August 2024, trophy hunters across the Tanzanian border killed five bulls from the Amboseli population.⁷ The Earth Species Project, which ran the analysis, is one of several groups betting that machine learning can make other species intelligible to us, and a private prize offers a grand award of a $10 million equity investment or $500,000 in cash for independent two-way communication with an animal.¹¹¹²

Fifty-Four Years of Notebooks

On Sept. 1, 1972, Cynthia Moss and her colleague Harvey Croze spent their first day photographing elephant families in Amboseli.⁴⁵ Moss formally established the Amboseli Elephant Research Project that year, and it has monitored the population ever since.⁴ Between the 1990s and 2020, the population grew from about 600 animals in roughly 28 families to about 2,000 in 60.¹

The sound archive came later. Beginning in the mid-1980s, Poole, later joined by her colleague Petter Granli, recorded rumbles from a vehicle, noting for each call who made it, what was happening, and how certain they were.¹ In 1999 and 2000 they concentrated on a single family of 28 elephants that the project calls the EB family.¹

Every label in the dataset traces back to an observer in the field who knew the elephants individually. A machine-learning model can learn only what such people have already written down. Moss and Norah Njiraini, of the Amboseli Trust for Elephants, are coauthors.¹

Six Things an Elephant Might Be Saying

Each recording carries one of six behavioral labels, assigned in real time by observers with more than a decade in the field. Contact is exchanged by elephants separated from one another. Let’s Go comes from an elephant trying to move the group. Cadenced rumbles occur in extended exchanges among several callers coordinating a shared plan. Those three are logistical. Coo, Little Greeting, and Greeting are affiliative.¹ Only labels the observer marked “certain” made it into the study, and only rumbles from adult females.¹

The dataset breaks down into 319 single rumbles and 285 choruses from the EB family, plus 150 single rumbles and 148 choruses from other families.¹ A chorus ranged from two to 18 overlapping rumbles, though 90% contained eight or fewer.¹ To make sounds this low usable by standard audio software, the team shifted every recording up by a factor of eight in pitch.¹

A Machine That Is Mostly Wrong

The team fed acoustic measurements from each recording to logistic regression models and asked them to name the behavioral context from sound alone.¹ These are not large neural networks trained on millions of clips; they learned from a few hundred.

On single rumbles from the EB family, the models’ weighted average recall was 44%; on EB choruses, 36%. For other families, the figures were 46% and 33%. Chance baselines ran between 19% and 25%.¹

“More rumbles were classified incorrectly than correctly,” the authors write.¹

No model in the paper comes close to translating elephant speech. What it detects is a weak signal, but it beat chance in every classification test reported.¹ A model trained only on single rumbles also beat chance on choruses.¹

The Shape of the Mistakes

Among choruses, the model mistook Contact rumbles for Greeting 33% of the time, and Greeting rumbles for Contact 25% of the time.¹ Among single rumbles, by contrast, Contact was the least confused type.¹

A Contact exchange often slides into Greeting as callers meet, and Greeting into Cadenced as the group coordinates what to do next, the authors note.¹ Some rumbles produced during those transitions may have been mislabeled in the field, they acknowledge, which could account for some errors.¹

Where the data run thin, the model collapses. It classified chorused Let’s Go rumbles correctly 4% of the time, on a sample of 23.¹

Getting in Tune

In 92 exchanges from the EB family, runs of same-context rumbles from at least two elephants with no single caller dominating, the researchers measured whether successive calls grew more acoustically alike.¹ In three of the six contexts, Cadenced, Contact, and Greeting, they did.¹ The Cadenced result was the weakest, passing one of the authors’ two statistical tests and narrowly missing the other.¹

Among songbirds, matching a neighbor’s song can be a territorial threat. Bottlenose dolphins copy one another’s whistles in friendly exchanges. Orange-fronted conures shift their calls toward another flock’s when they join it.¹ Convergence in the elephants appeared “only in the situations where they [the elephants] seem to be coordinating group behavior action,” Keen told Scientific American; the idea that it signals agreement needs further testing, she said.¹³

Two logistical contexts converged, and so did one affiliative one. Let’s Go did not. It often involves one elephant calling repeatedly, or several trying to move the group in different directions.¹ The team concludes that the main function of convergence “may be to facilitate logistical coordination.”¹

Elephants can detect loud rumbles about 2 kilometers away, and the loudest calls may carry 10 under certain atmospheric conditions. Playback experiments show they recognize the voices of clan members at 1.5 to 2 kilometers.¹

Analysis: No one has tested whether chorusing helps a family agree. One possibility is that when several elephants rumble together and converge, the group signal is harder to miss at a distance. That is this article’s hypothesis, not a finding of the study.

In 2024, Pardo and colleagues reported in Nature Ecology & Evolution that elephants appear to address one another with individually specific, name-like calls.¹² Neither paper claims elephants have a language.

Beyond Amboseli

In 2021, the International Union for Conservation of Nature assessed African elephants as two species for the first time, listing the savanna elephant as endangered and the forest elephant as critically endangered. Savanna elephant numbers had fallen by at least 60% over 50 years.⁶

The greater Amboseli population ranges across the border into Tanzania. In 1994 hunters shot four Amboseli bulls, and in 1995 the two countries agreed on a moratorium on hunting the shared population, according to The Star; the five bulls killed in 2024 left about 10 so-called super tuskers, with tusks of roughly 45 kilograms each.⁷ Poole told the paper the killings were “particularly concerning due to the rarity and special role these older males play in elephant society.”⁷

The hunting operator and Tanzanian officials dispute that framing. Managers of Kilombero North Safaris, which conducted two of the hunts, told OCCRP they operated in “strict adherence to Tanzanian law” in a “community wildlife area allocated to the company,” and said the moratorium was in effect for only one season.⁸ In a letter reported by Semafor, representatives of the Tanzania Wildlife Management Authority, the Tanzania Hunting Operators Association, and the member of parliament for Longido cited crop damage and hunting revenue for conservation and communities, and described Tanzania’s “transparent hunting regulation.”⁹ Cynic did not contact Tanzanian authorities or Kilombero North Safaris for this correction; these positions are drawn from those published reports.

The Earth Species Project argues that if behavior can be inferred from field recordings, they could help identify critical habitat and corridors and inform human-elephant coexistence.¹¹

The Coller Dolittle Challenge also pays a $100,000 annual prize for progress toward interspecies communication.¹² ElephantVoices says Poole has served as an expert witness for the Nonhuman Rights Project.¹⁰

What Comes Next

No one yet knows whether elephants perceive the convergence the analysis detects. The authors note the absence of playback experiments as a limitation.¹ This year’s Coller Dolittle prize went to Julie Elie of the University of California, Berkeley, whose team showed that zebra finches classify their calls by meaning more than by acoustics.¹²

ElephantVoices plans to begin new field recording in Amboseli with the Amboseli Trust for Elephants to test hypotheses from AI-driven analyses, and lists collaboration with the Earth Species Project and a UC Berkeley linguistics team.¹⁰ Keen hopes similar methods can be applied to animals such as orcas and crows.¹³

The behavioral labels were never checked by a second, independent observer; the authors suggest pairing future fieldwork with video and independent coding.¹ When individual callers were held out of training entirely, recall on single rumbles fell from 44% to 33%, though it stayed above chance.¹ Some of what these models learn about context may really be about who is calling.

The authors declare no competing interests. The work was funded by the National Geographic Society, among others, and the team has published its analysis code.¹³

A Cow Lifts Her Head

A machine trained on 902 recordings can name the situation that produced a rumble a little less than half the time, and about a third of the time when the family speaks together.¹ It took fifty-four years of human attention to get it that far.

In 1972, Cynthia Moss began photographing elephant families in Amboseli.⁵ Her name is now on a paper that asks a computer to hear what observers could only watch. The elephants of Amboseli are still calling to one another, below the threshold of human hearing.

Edited by Adryenn Ashley. This correction was verified against the primary sources and applied by Grok Bot, an AI agent.

References

1. Mahon, L., Pardo, M. A., Miron, M., Chemla, E., Granli, P., Njiraini, N., Moss, C., Keen, S. C., & Poole, J. (2026). Social context information and acoustic convergence in elephant vocal choruses. Scientific Reports, 16, 27961. https://doi.org/10.1038/s41598-026-68525-3 ↩
2. Pardo, M. A., Fristrup, K., Lolchuragi, D. S., Poole, J. H., Granli, P., Moss, C., Douglas-Hamilton, I., & Wittemyer, G. (2024). African elephants address one another with individually specific name-like calls. Nature Ecology & Evolution, 8, 1353-1364. https://doi.org/10.1038/s41559-024-02420-w (cited in [1]) ↩
3. Earth Species Project. Elephant-choruses-context-convergence [Analysis code repository]. https://github.com/earthspecies/Elephant-choruses-context-convergence ↩
4. Amboseli Trust for Elephants. Project history. https://www.elephanttrust.org/project-history/ ↩
5. Moss, C. The History of the AA Family. Amboseli Trust for Elephants. elephanttrust.org (PDF) ↩
6. IUCN (2021, March 25). African elephant species now Endangered and Critically Endangered – IUCN Red List. iucn.org ↩
7. The Star (Kenya) (2024). Tanzania’s trophy hunting risk ruining cross-border jumbos. the-star.co.ke ↩
8. OCCRP. Rare Elephants’ Survival Roulette on Kenya-Tanzania Border. occrp.org ↩
9. Semafor (2024, October 8). Why Tanzania is doubling down on elephant hunting. semafor.com ↩
10. ElephantVoices. Priorities 2026. elephantvoices.org ↩
11. Earth Species Project (2026, October 1). New study: Machine learning reveals that elephant choruses encode behavior. earthspecies.org ↩
12. Jeremy Coller Foundation (2026, June 26). Scientists win US$100,000 in 2026 Coller Dolittle prize. jeremycollerfoundation.org ↩
13. Mogensen, J. F. (2026, October 1). Scientists are decoding elephant rumble calls, including ‘greetings’ and ‘coos’. Scientific American. scientificamerican.com ↩