A humpback whale off New Caledonia can hold a song for 30 minutes or longer, cycling through phrases that shift and evolve across breeding seasons. To the human ear, the performance sounds alien. But when researchers fed nearly a decade of those recordings into algorithms originally designed to decode how infants learn words, something unexpected emerged: the songs obey the same mathematical rule that governs every known human language.
Two peer-reviewed studies published in 2025 now make the case in detail. Together, they suggest that whales and humans have independently arrived at the same statistical shortcut for organizing sound. The finding is striking, but it comes with a sharp caveat: sharing a pattern is not the same as sharing a purpose.
The mathematical fingerprint inside whale song
The first study, published in Science, applied infant-speech segmentation techniques to humpback whale recordings collected over roughly eight years off New Caledonia. The analysis identified statistically coherent, word-like segments whose frequency distribution follows a Zipfian power-law pattern. Zipf’s law is one of the most reliable regularities in linguistics: the most common word in a language appears about twice as often as the second most common, three times as often as the third, and so on down the list. English, Mandarin, Arabic, Swahili, every language ever tested follows this curve. Now humpback whale songs do too.
The second study, published in Science Advances, widened the scope dramatically. Researchers examined vocal sequences from 16 cetacean species and compared them against data from 51 human languages, testing two specific compression laws. One is Zipf’s law of abbreviation, which predicts that the most frequently used signals tend to be the shortest. The other is Menzerath’s law, which predicts that as a sequence grows longer, its individual components get shorter. According to the published analysis, 11 of those 16 whale species showed clear adherence to Menzerath’s law. Both laws describe efficiency shortcuts: ways to transmit more information with less effort. Their presence across such a wide taxonomic range suggests the pattern is not a quirk of humpbacks alone.
These results build on earlier groundwork. A 2019 study using network analysis on long-term humpback recordings had already detected small-world structure and Zipf-Mandelbrot rank-frequency distributions in humpback song elements, as reported in Royal Society Open Science. That work established that the statistical regularities are durable across years of recording, not artifacts of a single season or population. The newer papers extend the finding across species and, for the first time, benchmark whale vocalizations directly against dozens of human languages.
A separate theoretical thread helps explain why such patterns might arise in the first place. A 2025 perspective paper in Scientific Reports uses information-theoretic modeling to show how learning constraints, memory limits, and transmission noise can push any vocal system toward efficient coding, whether the signalers are humans, birds, or marine mammals. The argument is not that whales are linguistic, but that the physics of communication may favor the same solutions regardless of the communicator.
Why structure does not automatically mean language
The presence of Zipf’s law and Menzerath’s law in whale song does not, on its own, prove that whales are “speaking” in any sense humans would recognize. Statistical structure is not the same as semantic content. A system can follow Zipf’s law without carrying referential meaning. The classic demonstration, first described by mathematician George Miller in 1957, showed that randomly generated text can produce Zipfian distributions under certain conditions. Pattern alone is not proof of purpose.
One line of experimental work points toward a simpler explanation. In a controlled study published in Scientific Reports, researchers Inbal Arnon and Simon Kirby demonstrated that iterated cultural transmission alone can generate language-like statistical structure. In their setup, participants learned and reproduced artificial signals over repeated “generations,” with each group passing its version of the system to the next. Over time, Zipfian distributions emerged even without any explicit pressure to communicate meaning. If the same process operates in whale populations, where songs are learned, modified, and passed between individuals across breeding seasons, then the statistical overlap with human language could be a byproduct of social learning rather than evidence of linguistic cognition.
No study in the current body of evidence has linked specific Zipfian segments of whale song to identifiable behavioral outcomes or referential content. The recordings are rich in acoustic detail but lack the kind of contextual annotation that would let researchers say, for example, that a particular high-frequency unit corresponds to a specific social interaction or environmental cue. Without that behavioral grounding, the statistical findings remain descriptive. They tell us that whale songs are organized efficiently, but not what that efficiency is being used to convey.
The gaps researchers are racing to fill
Several open questions now define the frontier of this work. One involves population-level variation. The 16-species dataset in the Science Advances study is broad, but the degree to which social network size, migration overlap, or group isolation affects adherence to these compression laws has not been tested directly. Whale populations with larger, more fluid social networks might display measurably different Zipf exponents than smaller, isolated groups. That comparison could help distinguish whether the pattern is driven by the scale of cultural transmission or by something more fundamental about how cetacean brains organize sound.
Another gap involves the definition of the units themselves. In human language, linguists draw on well-established categories: phonemes, syllables, morphemes, words. For whales, the segmentation into notes, units, phrases, and themes is partly conventional and can differ between research groups. The Science paper used machine-learning methods inspired by infant speech processing to infer likely boundaries, but other approaches could carve the same acoustic stream into different pieces. Until there is stronger agreement on what constitutes a meaningful unit for whales themselves, any comparison to human words or syllables carries an asterisk.
As of June 2026, no published follow-up has yet bridged the gap between statistical pattern and behavioral meaning. The next critical experiments will likely involve pairing acoustic analysis with detailed behavioral observation: tracking whether specific song segments correlate with mating success, group coordination, or responses to environmental change. Several research groups, including teams behind the original Science and Science Advances papers, have signaled that this behavioral linkage is a priority.
What the convergence actually tells us
What the research does establish, firmly, is that the mathematical architecture of whale song is not random. Humpbacks and other cetaceans arrange their vocal elements in ways that mirror some of the deepest regularities known in human language. That convergence makes whale communication one of the most powerful natural test cases for theories about how structure emerges in complex signaling systems.
The parallel is genuinely remarkable. Humans compress language because they need to convey complex ideas quickly under the constraints of attention and memory. Whales may compress their songs for entirely different reasons: the physical demands of underwater sound transmission, the need to signal across kilometers of open ocean, or the dynamics of cultural copying in large, mobile populations. The outcome looks the same on a graph, but the pressures behind it could be fundamentally different.
For now, the most defensible reading of the evidence is this: whales and humans have converged on a common solution to the problem of organizing sound efficiently. Whether they have also converged on anything resembling shared meaning remains one of the most tantalizing open questions in animal cognition, and one that no statistical test, however elegant, can answer alone.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
