Who is that rumble for? Inside the debate over elephant names

Joyce Poole had noticed a curious selectivity in elephant conversations. One contact call would draw a response from a particular family member. Another would bring a different elephant into the exchange, while the rest paid little attention. In Save the Elephants’ account of the research, she described years of wondering whether the caller was addressing someone by name.[1]

A study published in June 2024 turned that field observation into an acoustic analysis and a playback experiment. Its authors interpreted the results as evidence for name-like calls. The interpretation is now contested: a 2026 reanalysis challenges the patterns detected by machine learning, while an October data-and-code release accompanies a reply still under review.[1][5][7]

Listening for the receiver

The archive comprised 469 calls from wild African savannah elephants in Kenya, with 101 callers and 117 receivers identified. Some recordings lacked a certain caller identity; the main receiver-classification test therefore used 437 calls. A random-forest model identified the intended listener in 27.5% of these, against 8.0% for models given shuffled acoustic features.[2]

An elephant’s rumble already carries information about its speaker and the circumstances of the call. Colorado State University’s account describes how researchers followed elephants by vehicle and recorded their vocalisations. That observational work supplied something a microphone alone cannot: information about who was present and what the animals were doing.[3]

For a greeting, the apparent receiver was the elephant being approached. For a caregiving call, it was the calf being attended to. Calls with uncertain or multiple recipients were excluded. The acoustic model then searched for a relationship between sound and that field-assigned identity.[2]

The question is subtle. A recognisable voice tells a listener who is speaking. A name-like signal would help identify whom the speaker is addressing. Predicting the second from the first can be misleading if a particular caller usually talks to the same few companions.[5]

Taking the recording back to the elephant

The field experiment changed the setting. Researchers played recordings to 17 wild elephants; 15 experienced both conditions. One recording had originally been addressed to the subject. The comparison came from the same caller but had been addressed to another elephant.[2]

The animals responded more strongly to calls originally meant for them, including approaching the speaker and calling back. Holding the caller constant made familiar-voice recognition alone a less satisfying explanation. Something about the two recordings mattered to the listener.[3]

A July 2025 correction repaired five playback-data errors. It altered numerical estimates for vocalisation timing and counts but left statistical significance and the qualitative findings unchanged.[4]

Those responses made the naming proposal compelling: the acoustic pattern appeared to matter outside a computer. But a selective response still needs an explanation. An animal may learn that a particular sound concerns it without using a name in the human sense.

The disagreement is about where the pattern comes from

Guha Dharmarajan’s reanalysis, published in Animal Behaviour’s May 2026 volume, argues that caller identity and similarities between calls made close together in time explain the classifier’s success. It also attributes the apparent similarity of calls directed to the same receiver mainly to temporal dependence and repeated observations being treated as more independent than they are. His conclusion is that name-like calls remain unproven.[5]

Imagine, purely as an illustration, several recordings of one caller during a single encounter with one companion. A model might recognise the encounter’s acoustic fingerprint. That could help it predict the companion without detecting a reusable sound that refers to that individual. The distinction is between recognising a situation and recognising an address.

Dharmarajan also proposed an alternative account of the playback responses in a February 2026 conference abstract: elephants’ social relationships and learned associations could produce selective reactions without arbitrary names. This is an offered explanation of the experiment, not a new field trial.[6]

The reply has reached a concrete stage. A Zenodo deposit published on 10 October supplies reanalysis code and corrections to timestamps in archived Amboseli recordings. Its description identifies it as material for a manuscript submitted to Animal Behaviour, with authors temporarily redacted for double-blind review. The public deposit contains data and code rather than the manuscript or saved result tables.[7]

The timestamp problem is consequential: elapsed time on a tape had been treated as if recording continued through pauses. The deposited script corrects that timing and examines receiver-related call similarity after excluding closely spaced calls. These materials make the methodological response inspectable; they do not yet establish that the published challenge has been resolved.[7]

What would make a rumble a name?

The original study found mixed evidence that different callers use a shared label for the same elephant. Its classifier could not reliably carry receiver identification across callers. Researchers also have not isolated a discrete name-bearing part of a rumble.[2][3]

That leaves a more interesting question than whether a machine can attach an elephant’s identity to an audio file. What relationship connects the sound, the caller and the listener—and does that relationship persist when the surrounding encounter changes?

Poole’s observation remains the starting point: an elephant calls, and one family member answers. The field response is real. Whether the animal heard an individual address, recognised a familiar situation, or combined several cues is the question the continuing research has to separate.[1][6]

Report an error or suggest a correction