By Science Blog
Machine learning found receiver-specific information in 469 Kenyan elephant rumbles, and 17 wild elephants responded faster to calls originally directed at them.
A deep elephant rumble is not an empty tone. Its structure can carry the caller’s identity, age, sex, emotional state and the situation in which the call was made. In a 2024 study, researchers asked whether one more identity was hidden inside: the elephant the caller meant to reach.
They assembled 469 contact, greeting and caregiving rumbles from wild African savannah elephants in Kenya. The study record maintained by Save the Elephants lists 101 identified callers and 117 identified receivers across Samburu and Buffalo Springs National Reserves in the north and Amboseli National Park in the south.
A machine-learning model found enough structure to identify the intended receiver correctly 27.5% of the time. Models given randomly scrambled acoustic features averaged only 8%.
That result alone would be interesting but incomplete. The stronger evidence came when the scientists played old calls to 17 wild elephants. A subject heard either a rumble originally directed at her or a rumble from the same caller meant for somebody else. When the old call had originally been addressed to the listening elephant, she approached the speaker and called back significantly sooner.
The calls therefore contained receiver-specific information that mattered to the elephants themselves. “Name-like” is a reasonable description. It is also a deliberately cautious one.
How researchers decided who a call was for
The study did not begin by recording every sound in a herd and assigning a recipient after the fact. The team restricted its analysis to calls for which one intended receiver could be identified from the social setting.
In a contact rumble, the receiver was the only adult family member more than 50 metres from the caller, or the one elephant that answered or approached. A greeting receiver was the individual who approached or was approached. A caregiving receiver was the calf being suckled, comforted or roused by an adult or adolescent female.
Calls with uncertain or multiple recipients were excluded. The final set comprised 371 rumbles from Samburu and 98 from Amboseli. The recordings spanned different periods between 1986 and 2022, drawing on long-running field knowledge that allowed researchers to recognise individual animals by their ears and other physical features.
The newer phase required 14 months of intensive fieldwork within a four-year project, according to Colorado State University’s account. Researchers followed known elephants by vehicle, recording calls while also watching closely enough to determine who was speaking and who responded.
The study in Nature Ecology & Evolution then measured two sets of acoustic features. A random-forest model tried to predict the receiver from those features, with its performance compared against 10,000 versions in which the acoustic information had been randomly permuted.
The 27.5% accuracy came from 437 calls for which the caller as well as the receiver was known. It was far from perfect, but decisively above the 8% average for the shuffled models. Calls from one caller to one receiver were also more similar to each other than that caller’s calls to different receivers, even after the analysis accounted for relationship type, behavioural context and recording date.
Names may be used only when they help
The model did much better in some circumstances than others. It correctly classified 42% of the contact calls and 46.8% of the caregiving rumbles considered in that analysis, but only 3.9% of greeting calls. Calls made by adult females were also much more likely to be classified correctly than those made by juveniles.
That variation argues against imagining a name placed neatly inside every sentence. A receiver label may be useful when a particular elephant is out of sight, or when an adult is directing care towards a calf. It may add little during a greeting at touching distance, when the recipient is already obvious.
Even for humans, names appear in only a fraction of utterances. An elephant rumble also conveys several kinds of information simultaneously. The features that helped predict the receiver accounted for only a small portion of the sound’s total variation.
Researchers have not isolated a discrete acoustic fragment and translated it as one elephant’s name. The relevant cue could arise from a subtle combination of frequencies and timing that is hard to see in a spectrogram. The model’s failure on most calls is consistent with labels being used selectively, but it also shows how much remains undecoded.
The playback removed a simpler explanation
A statistical model can discover a pattern without proving that an animal notices it. The playback experiment asked the elephants directly, through behaviour.
The team tested 17 animals in Samburu: 15 adult females, one nine-year-old female and one adolescent male. Fourteen received one test playback and one control. One subject received two pairs, while one heard only a test and another only a control.
In each comparison, the caller’s identity stayed the same. The test recording had originally been directed to the subject. The control came from that same caller but had been intended for another elephant. This mattered because a stronger response could otherwise reflect nothing more than recognising the voice of a close relative or preferred companion.
A single rumble was broadcast from a loudspeaker placed roughly 50 metres away. Paired trials were separated by at least seven days, with an average gap of nearly a month. The observers scoring the animals’ responses did not know whether a test or control call had been played.
The elephants approached the speaker faster after calls addressed to them, with a hazard ratio of 8.77. They vocalised sooner, with a hazard ratio of 7.45, and produced 2.41 times as many calls during the following ten minutes. Only one subject gave an unambiguously stronger response to the control.
Those numbers should not be read as a claim that every elephant reacted immediately. Among trials in which an approach or vocalisation occurred, the average latency was about 100 seconds and varied widely. Nor could the researchers guarantee that every archived playback call contained a vocal label. Despite that noise, the paired responses differed in the predicted direction.
Name-like is not the same as a human name
ScienceBlog’s earlier overview of the elephant-name finding placed it alongside better-known examples from dolphins and parrots. Those animals can address another individual by imitating that individual’s signature sound.
The elephant result may be different. Receiver identity remained predictable in calls that did not resemble the receiver’s own vocalisations, and the broader analysis found no clear evidence that the label depended on imitation. If that holds up, an elephant’s label for a companion may be closer to an arbitrary learned association than to calling someone by copying their voice.
There is an important complication. Calls made by different elephants to the same receiver were slightly more similar than calls aimed at different receivers. Yet when the model was trained on some callers and tested on entirely different ones, it could not identify the receiver better than chance. The cross-caller accuracy was 1.1%, compared with 1.4% for shuffled models.
That mixed result leaves several possibilities. Family members may share a label but pronounce it within highly distinctive voices. They may use partially overlapping versions. Or each caller may have a somewhat private way of addressing the same individual. More recordings per caller and receiver will be needed to separate those possibilities.
For the same reason, this study does not yet establish that elephants possess nouns, combine names with other word-like units or label water, food and places. It supports a narrower conclusion: their rumbles contain individually specific addressing information that listeners recognise.
Why elephant society may favour individual addressing
African elephant families live in social networks that repeatedly split and reunite. Close associates can be widely separated while foraging, then seek one another over distance. Relationships remain differentiated across many years and many individuals.
In that setting, a call that says more than “someone in my family should answer” has obvious value. A mother may need one calf. A separated adult may need one particular companion. Receiver-specific information can attract the right animal’s attention without requiring every listener to respond.
Rumbles also help elephants coordinate group decisions. In related work, male elephants were found using patterned “let’s go” calls before departures, extending a ritual previously associated with female-led family groups. These findings show that the same low-frequency channel can organise both individual relationships and collective movement, though they do not yet amount to a translated elephant language.
The naming study is persuasive because two different forms of evidence converge. Acoustic analysis found information about the intended receiver, and the playback animals behaved as though that information was meaningful.
Seventeen subjects remain a small field sample, dominated by adult females and confined to one of the two recording regions. The model did not reveal the label itself. Replication with more elephants, populations and call types will determine how general the system is.
What the study establishes is quieter and more interesting than the claim that scientists can now read an elephant’s address book. When one elephant rumbles, another can apparently hear not only who is speaking, but whether the message is meant for her.
