Animal Language

From PRIMO.ai
Jump to navigation Jump to search

Youtube search... ...Google search


Animal language research asks how nonhuman animals encode and exchange information through sounds, gestures, movements, chemicals, and vibrations, and how far those systems share the properties that make human language special. For most of the twentieth century this was a question for ethologists with notebooks, tape recorders, and patience. In the last decade it has also become a data science problem. Researchers now record animals around the clock with suction-cup tags, hydrophone arrays, drones, collar microphones, and synchronized video, then use machine learning to find units, patterns, and contexts that human ears and hand-drawn spectrograms missed.

The question is old, and much of the foundation predates computers. Karl von Frisch showed that the honeybee waggle dance encodes the direction and distance of food, work that earned him a share of the 1973 Nobel Prize in Physiology or Medicine. In 1980 Robert Seyfarth, Dorothy Cheney, and Peter Marler played recorded vervet monkey alarm calls through hidden speakers and showed that the monkeys responded differently to "leopard," "eagle," and "snake" calls even when no predator was present. Charles Hockett's 1960 list of language "design features" gave everyone a checklist to argue about. Dolphin signature whistles, first described by Melba and David Caldwell in 1965, raised the possibility that some animals carry something like a name.

The AI-era field is built by a mix of biologists, linguists, and computer scientists:

  • Project CETI (Cetacean Translation Initiative), launched in 2020 through TED's Audacious Project and led by marine biologist David Gruber, works with the Dominica Sperm Whale Project founded by Shane Gero in 2005. Its collaborators include MIT computer scientists (Pratyusha Sharma, Jacob Andreas, Daniela Rus, Antonio Torralba), UC Berkeley linguist Gašper Beguš, and theorists Shafi Goldwasser and Michael Bronstein.
  • Earth Species Project (ESP), an independent nonprofit cofounded by Aza Raskin, builds open machine learning models for animal sound across many species.
  • The Wild Dolphin Project (Denise Herzing, studying Atlantic spotted dolphins in the Bahamas since 1985) works with Georgia Tech (Thad Starner) and Google on the CHAT interface and the DolphinGemma model.
  • Google DeepMind's Perch team builds bioacoustic foundation models used for monitoring and Bird Identification.
  • University field and lab groups including Tel Aviv University (Yossi Yovel, bats), the University of Zurich (Simon Townsend, bonobos), the Max Planck Institute for Evolutionary Anthropology (Catherine Crockford and Roman Wittig, chimpanzees), Woods Hole Oceanographic Institution (Laela Sayigh, bottlenose dolphins), UC Berkeley (Julie Elie and Frédéric Theunissen, zebra finches), and the Colorado State, Save the Elephants, and ElephantVoices collaboration (Mickey Pardo, Joyce Poole) on African elephants.

The core problem breaks into four nested questions:

  1. Information: does a signal carry information about who is calling, to whom, and about what?
  2. Structure: is the signal built from reusable parts that combine according to rules?
  3. Meaning: what do those parts do for the animals, as shown by the animals' own behavior?
  4. Conversation: can humans produce signals that the animals treat as meaningful and answer?

Most published work today sits at levels 1 and 2. A growing number of experiments reach level 3. Level 4 remains a research target, not an achievement.

The central mechanism is a loop: pair very large recordings with careful behavioral context, let machine learning propose candidate units and associations, then test those proposals with playback or choice experiments that let the animals "vote" on whether the proposed categories matter to them.

This is a real paradigm shift. The ape-language and animal-language projects of the 1960s to 1990s tried to teach animals human-designed symbol systems: American Sign Language for the chimpanzees Washoe and Nim Chimpsky, lexigram keyboards for the bonobo Kanzi, spoken English labels for Alex the African grey parrot, and artificial gestural languages for Louis Herman's dolphins. Those projects produced genuine findings and bitter disputes about cueing and over-interpretation. Current work flips the direction. Instead of asking whether animals can learn our system, it asks what their own systems are, measured in natural settings at a scale only machine learning makes possible. It also changes the shape of the debate. "Do animals have language?" is giving way to "which design features of language appear in which species, to what degree, and why?"

The Simons Institute described the opportunity this way when it convened the first workshop in what became a recurring series: With advances in machine learning, we now stand a chance to decode communication in non-human animal and plant life, which has fascinated humans for millennia. Cutting-edge technology can be used to benefit not only humankind, but other species on our planet, to better understand and protect the life around us. To make this possible, one needs to bring together and build on the latest discoveries of experts in disparate fields, including machine learning, signal processing, data science, linguistics, and acoustics. The goal of the workshop is to explore the challenges and current state-of-the-art in the study of non-human species communication as well as disciplines and technologies that are likely to provide solutions to these challenges. Decoding Communication in Nonhuman Species | Simons Institute

That first workshop ran virtually in August 2020, organized by Michael Bronstein, Shafi Goldwasser, and David Gruber. The Simons Institute and Project CETI have since held follow-up workshops in 2023, 2024, and July 2025, and the organizers explicitly modeled the series on the 1956 Dartmouth workshop that launched artificial intelligence as a field.

Eras of animal language research
Era Approximate dates Dominant approach Representative work Main limitation
Natural history and ethology 1900s to 1970s Observation and simple field experiments von Frisch's bee dance; early birdsong studies; signature whistles (1965) Small samples, mostly descriptive
Ape and animal "language" projects 1960s to 1990s Teaching human-designed symbols Washoe, Nim Chimpsky, Koko, Kanzi, Alex the parrot, Herman's dolphins Cueing (Clever Hans effect), disputed interpretation, artificial systems
Playback and referential signals 1980 onward Record natural calls, play them back in the wild Vervet alarm calls; prairie dog predator calls; dolphin vocal labels Labor intensive, a few call types per study
Classical machine learning 2000s to mid 2010s Hand-built acoustic features plus classifiers Egyptian fruit bat calls (2016); early commercial "translators" Feature engineering, small labeled sets
Deep and self-supervised learning Late 2010s to 2023 Neural networks on spectrograms; learned embeddings BirdNET, Perch 1.0, AVES; machine-learned elephant addressee prediction Labels scarce; models find structure, not meaning
Foundation models and AI-assisted experiments 2024 onward Audio-language models, generative models, closed-loop playback, behavioral validation NatureLM-audio, Perch 2.0, DolphinGemma, WhAM; sperm whale vowels; zebra finch meaning tests Ground truth, ethics, and hype


Machine learning lets us hear structure in animal communication that humans missed for centuries, but structure is not meaning, and meaning is not conversation. Every step from pattern to translation still has to be earned with evidence from the animals themselves.


Language, Communication, and the Design Features

Almost every animal communicates. Far fewer are known to have systems with the specific properties linguists associate with language. Keeping the two words apart prevents most of the confusion in popular coverage.

Communication is any signal that evolved to influence the behavior of another individual. A frog's mating call, a peacock's tail, and a skunk's warning stripe all qualify.

Language, in the linguist's sense, is a communication system with a cluster of properties: learned arbitrary symbols, discrete reusable units, rules for combining them, the ability to talk about things that are not present, and the capacity to produce an open-ended number of new messages. Hockett's 1960 list of design features is still the common reference, and in 2002 Marc Hauser, Noam Chomsky, and Tecumseh Fitch argued that recursion (embedding structures inside structures of the same kind) may be the one feature unique to humans.

The modern question is therefore not "do animals have language?" but "which features appear where?" The table below summarizes where the evidence currently stands.

Language design features and current animal evidence
Feature What it means Strongest animal evidence Current status
Semanticity and reference Signals map to objects, events, or categories Vervet and prairie dog predator calls; zebra finch call meanings validated by the birds (2025 to 2026) Well supported for limited repertoires
Arbitrariness The form of a signal is unrelated to its meaning African elephant name-like calls that do not imitate the receiver (2024) Supported in a few species
Vocal labels for individuals Calls that pick out a specific individual Dolphin signature whistles; marmoset phee calls (2024); elephant rumbles (2024) Supported; mechanisms differ by species
Discreteness and combinatorial units A small inventory of units recombines into many signals Sperm whale coda features; humpback song units Supported; how "linguistic" these units are is debated
Duality of patterning Meaningless units build meaningful ones, like sounds into words Sperm whale "phonetic alphabet" and vowel-like features (2024 to 2026) Proposed interpretation
Compositionality The meaning of a combination derives from its parts Japanese tits (2016); bonobos, including nontrivial cases (2025); chimpanzee two-call combinations (2025) Supported for short combinations
Displacement Reference to things distant in space or time Honeybee waggle dance Rare; the bee case is strong
Cultural transmission Learned from others rather than fixed by genes Humpback song; sperm whale clan dialects; songbirds Well supported
Productivity and recursion Unbounded novel messages built hierarchically No convincing demonstration Open question; often argued to be uniquely human

Two clarifications matter throughout this page. First, trivial compositionality means the meaning of a combination is simply the sum of its parts (alarm plus recruitment means "alarm and come here"). Nontrivial compositionality means one part modifies the other, the way "fake" changes "gun" in "fake gun." Until 2025 many researchers argued that only humans showed the nontrivial kind. Second, information is not intention. A signal can carry reliable information about the sender (tired, hungry, large, stressed) without being produced in order to inform anyone. Stressed tomato plants emit informative ultrasonic clicks, but nobody claims the tomatoes are talking.

From Ape Language Projects to Field Decoding

The history explains why current researchers are so careful about controls.

Teaching human systems. Washoe (from 1966) and later Koko the gorilla were reported to use hundreds of signs. In 1979 Herbert Terrace and colleagues published "Can an ape create a sentence?" in Science after analyzing videotapes of Nim Chimpsky; they concluded that Nim's multi-sign sequences were largely imitation and prompting rather than syntax. The episode echoed the much older case of Clever Hans, the horse whose "arithmetic" turned out in 1907 to depend on unconscious cues from human observers. The lesson that survives: any study where a human who knows the expected answer interacts with the animal is vulnerable to cueing.

Other projects produced more durable results. Kanzi learned lexigrams largely by observation. Irene Pepperberg's African grey parrot Alex labeled objects, colors, and quantities. Louis Herman's bottlenose dolphins Akeakamai and Phoenix responded correctly to novel gestural and acoustic "sentences," including changes in word order. In 2011 John Pilley and Alliston Reid reported that the border collie Chaser had learned the names of 1,022 objects. These studies show impressive comprehension and learning capacity. They say less about the animals' own natural communication.

Listening to natural systems. Playback experiments, beginning with the vervet study, tested what natural calls mean by watching responses. Con Slobodchikoff's work on Gunnison's prairie dogs argued that their alarm calls encode predator type and even size and color of an approaching human. Toshitaka Suzuki and colleagues showed in 2016 that Japanese tits respond differently to call sequences depending on order, an early experimental case of compositional syntax in birds.

Measuring at scale. The shift to machine learning came with continuous recording. In a 2016 Scientific Reports study, Yosef Prat, Mor Taub, and Yossi Yovel monitored a colony of Egyptian fruit bats with round-the-clock audio and video for 75 days and analyzed almost 15,000 vocalizations whose emitter, addressee, and context could be identified. What looked like generic squabbling carried information about who was calling, whom they addressed, what the dispute was about, and how the listener responded. That study, featured in the BBC video below, became a template: record everything, attribute every call, then let classifiers measure how much information the signals carry.

Beyond animals. In 2023 a Tel Aviv University team led by Lilach Hadany reported in Cell that tomato and tobacco plants emit airborne ultrasonic clicks when stressed, dozens per hour under drought or cutting compared with fewer than one per hour when unstressed, and that machine learning models could identify the plant's condition from the sounds alone. Whether any organism uses these sounds as signals is unknown, which makes the study a useful reminder of the difference between informative and communicative.

How Animal Communication Is Decoded

The modern workflow has eight stages. Not every project uses all of them, and most of the scientific uncertainty lives in the later stages.

  1. Long-term field context. Decoding depends on knowing the animals: who is related to whom, who travels together, what each individual was doing. The Dominica Sperm Whale Project has followed known families since 2005, the Wild Dolphin Project since 1985, and the Sarasota Dolphin Research Program for decades longer. The 2024 elephant naming study drew on recordings from Amboseli and Samburu spanning 1986 to 2022. Without this backbone, recordings are nearly uninterpretable.
  2. Instrumented recording. Suction-cup tags carry hydrophones, depth sensors, and accelerometers. Hydrophone arrays, buoys, and drones localize callers. Bat colonies are filmed and recorded continuously. The key requirement is attribution: knowing who called, and ideally to whom.
  3. Detection, separation, and attribution. Models first find calls in thousands of hours of noise, then separate overlapping callers (the "cocktail party" problem), then assign each call to an individual. Species-level detectors such as Perch and BirdNET handle the first step well; individual attribution still relies heavily on tags, arrays, and video.
  4. Representation. The waveform becomes either a spectrogram or a learned embedding. Recent systems use pretrained audio encoders: Perch 2.0 learns from labeled recordings of 14,597 species, NatureLM-audio uses a BEATs encoder, DolphinGemma uses Google's SoundStream tokenizer, and WhAM uses VampNet's acoustic tokens. As in human Language Translation and speech processing, tokenization is a design choice that determines what a model can and cannot notice.
  5. Unit and structure discovery. Clustering, sequence models, and statistically motivated segmentation propose candidate units. Project CETI decomposed sperm whale codas into rhythm, tempo, rubato, and ornamentation. The humpback whale study segmented song where transitional probabilities between sound elements dipped, the same cue human infants use to find word boundaries. Beguš used the internal representations of a generative adversarial network trained on codas to flag spectral features that later acoustic analysis confirmed.
  6. Context linking. Candidate units are associated with who, what, where, and what happened next. Researchers build ethograms (catalogs of behavior), borrow distributional semantics from computational linguistics (the bonobo study coded more than 300 contextual features per utterance), and train classifiers to predict addressee or context from acoustics alone.
  7. Experimental validation. The animals get the final say. Playback experiments ask whether animals respond differently to proposed categories, as when elephants responded more strongly to calls originally addressed to them. Operant tests ask whether animals sort signals the way the model does, as when zebra finches pressing buttons confused calls with similar meanings more than calls with similar sounds. Blinded comprehension tests, like the 2024 soundboard dog study, rule out human cueing.
  8. Generation and interaction. Generative models such as WhAM and DolphinGemma produce synthetic calls for controlled experiments. Interfaces such as the Wild Dolphin Project's CHAT system take a different route: they pair novel synthetic whistles with objects dolphins like (sargassum, seagrass, scarves) to build a small shared vocabulary, rather than attempting to translate natural calls.

Dependencies and conditions. The pipeline only works when individual attribution is reliable, behavioral annotations are rich, each context is sampled many times, recordings come from natural settings, and playbacks are ethically acceptable. Failure at any one of these points undermines everything downstream.

Where uncertainty remains.

  • Human-defined categories. Labels like "feeding," "aggression," or "query" are human categories. Jakob von Uexküll's idea of the Umwelt (each species' perceptual world) warns that an animal's categories may cut reality differently.
  • Structure is not function. Statistical regularities such as Zipf's law can emerge from processes that have nothing to do with meaning. George Miller pointed out in 1957 that randomly typed text produces Zipf-like word frequencies.
  • Model artifacts. A generative model can produce sounds that pass expert listening tests yet mean something different, or nothing, to the animals.
  • Cueing. Any test in which a person who knows the expected answer interacts with the animal risks the Clever Hans effect.
  • Population bias. Most results come from a handful of well-studied populations. Sperm whale clans have distinct dialects, so a finding in one Caribbean clan may not transfer to another.

The Physics

The physics of the signal shapes what animals can say, what we can record, and what a model can learn.

Sound in Air and Water

Sound travels at about 343 meters per second in air at 20 °C and about 1,500 meters per second in seawater. Wavelength follows λ = c / f. A 20 Hz elephant rumble has a wavelength of about 17 meters in air; the same frequency in water spans about 75 meters. Long wavelengths diffract around trees and terrain, which helps explain why elephants use low-frequency rumbles (fundamentals in the roughly 10 to 30 Hz range, partly below human hearing) to stay in contact over distances of a few kilometers.

Two loss mechanisms set the range of any call:

  • Spreading loss for a point source in open space grows as about 20 log10(r) decibels with distance r.
  • Absorption grows steeply with frequency. In seawater it is roughly 0.06 dB per kilometer at 1 kHz, about 1 dB per kilometer at 10 kHz, and on the order of 30 dB per kilometer at 100 kHz.

That frequency dependence is why baleen whales use low frequencies for long-range song while toothed whales use high frequencies for short-range echolocation. Human hearing spans roughly 20 Hz to 20 kHz; belugas communicate up to about 120 kHz, and many bats echolocate well above 20 kHz. Much of the animal acoustic world is invisible to unaided human perception, which is one reason automated analysis finds things listeners missed.

Decibel figures need care. Underwater levels use a reference pressure of 1 micropascal, airborne levels use 20 micropascals, and water's much higher acoustic impedance changes the relationship between pressure and intensity. As a rough rule, about 61.5 to 62 dB must be subtracted from an underwater figure to compare it with an airborne one. Sperm whale echolocation clicks, with source levels reported above 230 dB re 1 µPa at 1 meter, are among the loudest biological sounds, but that number cannot be compared directly with a jet engine measured in air.

From Waveform to Representation

A spectrogram slices the waveform into short overlapping windows and computes the frequency content of each (the short-time Fourier transform). Every spectrogram embodies a trade-off: a short window resolves fast events such as clicks but blurs frequency, while a long window resolves pitch but smears timing. This time-frequency uncertainty, formalized by Dennis Gabor, means no single representation suits every species. Sperm whale coda rhythm lives in inter-click timing measured in milliseconds; humpback song lives in slowly changing pitch contours lasting seconds. Learned representations (embeddings from pretrained encoders) partly sidestep the problem by letting the model choose which features to keep, at the cost of making those features harder to interpret.

Production: Source, Filter, and Whale Vowels

Human vowels follow the source-filter model. The vocal folds produce a buzz (the source); the shape of the throat and mouth creates resonances called formants (the filter). Vowels differ mainly in their first two formants: /a/ has a high first formant, /i/ a low first and high second formant.

Sperm whales produce clicks with the phonic lips in their nasal complex, and each click reverberates through the spermaceti organ before leaving the head. The 2025 vowel study reports that coda clicks fall into two spectral classes, one with a single prominent resonant peak and one with two, and that some codas shift between patterns in ways resembling diphthongs. The authors argue these patterns are actively controlled. Because sperm whale codas are produced at a far slower rate than human speech, Beguš described them as very slow vowels. The analogy is acoustic, not semantic: the finding shows that whales modulate frequency as well as timing, not that the spectral classes carry word-like meanings.

Information Theory and Statistical Laws

Claude Shannon's entropy measures the average uncertainty of a signal source:

H = − Σ pi log2 pi (bits per symbol)

The more useful quantity for meaning is mutual information between signal S and context C:

I(S; C) = H(S) − H(S | C)

It measures how much knowing the context reduces uncertainty about the signal (or the reverse). A classifier that predicts context or addressee from acoustics above chance demonstrates that I(S; C) is greater than zero. This is exactly what the bat, elephant, and bonobo studies establish. It does not tell us whether the animals exploit that information.

Zipf's law says the frequency of the r-th most common unit falls roughly as f(r) ∝ 1 / rα with α near 1 in human languages. Transitional probability between units A and B is P(B | A) = P(AB) / P(A). In 1996 Jenny Saffran, Richard Aslin, and Elissa Newport showed that 8-month-old infants segment speech at dips in transitional probability. The 2025 humpback study applied the same segmentation rule to whale song and found the resulting subsequences followed a Zipfian distribution. Information-theoretic laws like these are useful fingerprints of learned, culturally transmitted systems, but as Miller's random-typing example shows, they are not proof of meaning on their own.

Combinatorial Capacity

The 2024 sperm whale study gives a concrete example of how combinatorics expands an inventory. Its analysis identified 18 rhythm types, 5 tempo types, 3 rubato types, and 2 ornamentation states:

18 × 5 × 3 × 2 = 540 possible combinations

Roughly 156 distinct combinations were reported in the 8,719-coda dataset. In information terms, choosing uniformly among the 21 traditional coda types carries at most log2 21 ≈ 4.4 bits; among 156 observed combinations about 7.3 bits; among all 540 about 9.1 bits. If the two vowel-like spectral classes vary independently, they could add up to one more bit per coda. These are capacity ceilings, not measured information transfer. Real usage is far from uniform, and nobody yet knows how much of this capacity whales use to convey meaning.

The Data Gap

The biggest constraint is data. Frontier language models for human text train on trillions of tokens. The best annotated animal datasets are tiny by comparison:

Scale of key annotated datasets
Study Species Approximate data
Sharma et al. 2024 Sperm whales (Eastern Caribbean clan) 8,719 codas
WhAM, NeurIPS 2025 Sperm whales About 10,000 codas collected over two decades
Prat, Taub & Yovel 2016 Egyptian fruit bats Almost 15,000 attributed vocalizations
Elie and Theunissen (2025 to 2026) Zebra finches More than 8,000 calls from over 30 birds
Girard-Buttoz et al. 2025 Wild chimpanzees 4,323 utterances from 53 individuals
Berthet et al. 2025 Wild bonobos About 700 recordings
Pardo et al. 2024 African elephants 469 rumbles
Arnon et al. 2025 Humpback whales (New Caledonia) 8 years of song recordings

That gap of roughly eight to nine orders of magnitude is why bioacoustic models lean on transfer learning from human speech, music, and other species, and why projects such as CETI aim to collect vastly larger corpora.

Why Unsupervised Translation Is Hard

Human-to-human machine translation without parallel text works partly because human languages describe the same world in roughly similar ways, so their word embedding spaces can be approximately aligned. Earth Species Project's early vision drew on that idea. The catch is that animals may not carve the world the way we do, so there is no guarantee of a shared geometry.

Shafi Goldwasser, David Gruber, Adam Tauman Kalai, and Orr Paradise formalized the problem in a NeurIPS 2023 paper. Their framework assumes a prior over plausible translations and proves bounds showing that translation error falls as the source communication system becomes more complex and as the "common ground" between source and target grows. Surprisingly, their bounds suggest unsupervised translation may need about as much source data as a hypothetical supervised setting. The results come from stylized language models and synthetic data, so they show the problem is not hopeless in principle rather than showing that whale translation will work.

Why This Matters

Scientific importance. Animal communication is the only comparative evidence we have about how language evolved. If bonobos and chimpanzees both combine calls compositionally, the shared ancestor of all three species may already have had that capacity at least 7 million years ago. If whale song shows the same statistical structure as human language because both are culturally transmitted, then learning and transmission, not a uniquely human mental module, may explain some "universal" properties of language. These are live debates, and the new data are reshaping them.

Conservation. The most immediate payoff is monitoring, not translation. Bioacoustic models count species, track populations, and detect threats across forests and oceans faster than human analysts; a University of Hawaiʻi lab reported finding endangered honeycreeper sounds nearly 50 times faster using Perch. Understanding what calls mean could also support targeted interventions, such as warning elephants away from conflict zones, reducing ship strikes, or understanding how noise pollution disrupts whale communication. This connects directly to Animal Ecology and Collective Animal Intelligence.

Animal welfare. Classifiers that estimate distress or emotional state from vocalizations could help monitor farms, shelters, zoos, and laboratories. The same tools raise questions about who controls the data and how it is used.

Law and ethics. Evidence of language-like structure has legal consequences. NYU's More-Than-Human Life (MOTH) Program and Project CETI argue that such findings could strengthen protections for whales and oceans. The same capabilities could be misused: geographer Karen Bakker, author of The Sounds of Life (2022), warned that sufficiently good synthetic calls could lure animals to their deaths. Even well-intentioned playbacks can cause harm, as described under Recent Developments.

Technology. Animal communication is the hardest possible low-resource language problem: no parallel data, few labels, alien perceptual worlds. Methods developed here (self-supervised audio models, unsupervised translation theory, interpretability for audio models) feed back into human speech technology, Bio-inspired Computing, and Neuroscience, where decoding vocalizations from neural recordings (as in the Heelan et al. primate study above) links behavior to the brain. Research on Connecting Brains and brain interfaces faces a parallel challenge: reading meaning from signals that were never designed for us.

What would change if it works. A validated, even partial, translation of a nonhuman communication system would be one of the most consequential scientific results imaginable. It would force legal and moral systems to reckon with nonhuman perspectives, transform conservation, and change how humans understand their place among other minds.

What remains uncertain. Nobody has yet demonstrated open-ended two-way communication with any wild species. It is unknown whether any animal system has the productivity of human language, whether rich structure in whale codas carries correspondingly rich meaning, and whether human-defined behavioral categories match the animals' own. The honest current position is that machine learning has made animal communication far more measurable, while translation remains a hypothesis to be tested.

Current Research

Recent Developments

The period from mid-2024 to late 2026 produced more significant results than the preceding decade. The table summarizes the main studies; the subsections explain what each does and does not establish.

Key studies at a glance (evidence labels: Demonstrated, Proposed interpretation, Hypothesis)
Year Species Team Method Main finding Evidence level
2024 Sperm whales Project CETI and MIT Feature decomposition of 8,719 codas Codas combine rhythm, tempo, rubato, ornamentation into a large inventory Demonstrated structure; "phonetic alphabet" label disputed
2024 African elephants Colorado State, Save the Elephants, ElephantVoices Random forest on 469 rumbles plus playback Calls contain non-imitative, receiver-specific labels Demonstrated; name not yet isolated
2024 Dogs UC San Diego and collaborators Blinded soundboard presses with 59 dogs Trained dogs respond to words themselves, not owner cues Demonstrated for a few words
2025 Humpback whales Hebrew University, Edinburgh, St Andrews, others Infant-inspired segmentation of 8 years of song Subsequences follow Zipf's law Demonstrated statistics; meaning not addressed
2025 Bonobos University of Zurich, Harvard Distributional semantics on ~700 recordings Widespread compositionality, including three nontrivial cases Demonstrated correlationally
2025 Chimpanzees Max Planck Institute for Evolutionary Anthropology 4,323 utterances, 16 two-call combinations Four combinatorial mechanisms that expand meaning Demonstrated correlationally
2025 Bottlenose dolphins WHOI, Sarasota Dolphin Research Program, St Andrews Shared whistle catalog plus wild playbacks Shared non-signature whistles may act like words (alarm, query) Hypothesis with experimental support
2025 Sperm whales UC Berkeley and Project CETI Spectral analysis of coda clicks Vowel-like and diphthong-like spectral patterns Demonstrated acoustics; function unknown
2025 Sperm whales UC Berkeley and Project CETI WhAM generative model (NeurIPS 2025) Synthetic codas; useful learned representations Tool, not a finding about meaning
2025 to 2026 Zebra finches UC Berkeley Ethogram, machine learning, operant tests Birds categorize calls by meaning more than by sound Demonstrated in captivity

Sperm Whales: From Codas to Vowels

What was tested. Whether sperm whale codas (short click sequences, usually under two seconds) are a small fixed set of types or a combinatorial system.

Method and findings (2024). Pratyusha Sharma, Shane Gero, Roger Payne, David Gruber, Daniela Rus, Antonio Torralba, and Jacob Andreas analyzed 8,719 codas from about 60 whales of the Eastern Caribbean clan, recorded by the Dominica Sperm Whale Project between 2005 and 2018. They identified two context-independent features (rhythm and tempo) and two context-dependent ones: rubato (gradual stretching or compressing of codas across an exchange) and ornamentation (an extra click). Freely combined, these produce an inventory nearly an order of magnitude larger than the 21 traditional coda types.

Method and findings (2025 to 2026). Gašper Beguš and colleagues analyzed more than a thousand tagged coda recordings and reported in Open Mind (November 2025) that clicks also vary in spectral shape, forming an "a-like" class and an "i-like" class plus diphthong-like transitions. The patterns were consistent across whales and coda types, and whales exchanging codas sometimes alternated vowel classes while keeping the rhythm. A follow-up paper in Proceedings of the Royal Society B (April 2026) extended the phonological analysis.

What it establishes and what it does not. The work establishes that sperm whale codas are more structured and carry more potential information than the old click-count view assumed. It does not establish what any coda means. CETI's own researchers stress that describing the architecture of a communication system is not translation.

Objections. Whale biologist Luke Rendell of the University of St Andrews criticized the "phonetic alphabet" label, arguing that the way whales vary tempo works quite differently from how alphabet letters build linguistic expression. Both studies come from one research consortium and largely one clan; independent replication on other populations is still needed.

Language-Model Tools for Animal Sound

NatureLM-audio (Earth Species Project, ICLR 2025). The first audio-language foundation model built for bioacoustics. It connects a BEATs audio encoder to the Llama 3.1 8B language model, trained on curated text-audio pairs spanning animal sounds, human speech, and music. Researchers query it in plain English for species classification, detection, call-type and life-stage classification, captioning, and individual counting. On the new BEANS-Zero benchmark it set state-of-the-art results on several tasks, including zero-shot classification of species it never saw in training, and showed that knowledge learned from speech and music transfers to animal sounds. ESP released weights and code; a May 2025 update traded some bioacoustic accuracy for prompt flexibility, and the demo moved to version 1.1 in April 2026.

Perch 2.0 (Google DeepMind, August 2025). A supervised model pretrained on 14,597 species of birds, mammals, amphibians, and insects. It achieved state-of-the-art results on the BirdSet and BEANS benchmarks and, in a paper presented at the NeurIPS 2025 workshop on AI for Non-Human Animal Communications, outperformed specialized marine models on whale and underwater tasks despite almost no marine training data. The team's hypothesis is that fine-grained species classification forces the model to learn unusually general acoustic features.

DolphinGemma (Google, Georgia Tech, Wild Dolphin Project, April 2025). A model of roughly 400 million parameters built on Gemma technology and trained on the Wild Dolphin Project's decades-long archive of Atlantic spotted dolphin recordings. It uses the SoundStream tokenizer and works audio-in, audio-out: given a sequence of dolphin sounds, it predicts likely next sounds, much as a text model predicts the next word. It is small enough to run on the Pixel phones researchers carry into the water and supports the CHAT system (upgraded from a Pixel 6 to a Pixel 9 based rig), which plays synthetic whistles associated with objects. Google announced it would release DolphinGemma as an open model in summer 2025; public sources as of 2026 indicate the model and data had not been broadly released, so independent evaluation remains limited.

WhAM (Project CETI and UC Berkeley, NeurIPS 2025). The Whale Acoustics Model fine-tunes VampNet, a masked acoustic token model pretrained on music, on about 10,000 codas. It generates synthetic codas from any audio prompt, can restyle other sounds into coda form, and its internal representations perform well on rhythm, social unit, and vowel classification even though it was trained only to generate. Evaluation used Fréchet Audio Distance and perceptual tests with expert marine biologists. WhAM is infrastructure for controlled playback experiments; it is not a translator.

What these tools establish. Foundation models now make detection, classification, and generation of animal sounds dramatically cheaper and more general. None of them, on its own, assigns meanings to animal signals.

Names and Vocal Labels

Elephants (June 2024). Michael Pardo and colleagues analyzed 469 rumbles from wild female-offspring groups in Kenya's Amboseli National Park and Samburu and Buffalo Springs reserves, involving 101 callers and 117 receivers. A random forest classifier predicted the intended receiver from the call's acoustic structure 27.5% of the time, well above control levels. Crucially, calls did not resemble the receiver's own calls, so unlike dolphins and parrots, which address others by imitating their signature calls, elephants appear to use arbitrary labels, as humans do. In playbacks, elephants responded more energetically to calls originally addressed to them. Labels were more common in long-distance calls and from adults addressing calves. Limits: the "name" has not been isolated within the rumble, and the prediction accuracy shows the label is a partial component of each call. Joyce Poole of ElephantVoices, whose talk appears below, coauthored the study.

Marmosets (2024). A study published in Science reported that common marmosets use specific phee calls to address particular individuals, and that family members use similar labels, suggesting the labels are socially learned.

Dolphins. Bottlenose dolphins develop individually distinctive signature whistles and copy the whistles of others to address them, a finding supported by playback studies such as Stephanie King and Vincent Janik's 2013 work. The new frontier is the other half of their repertoire, covered below.

Compositionality in Great Apes

Bonobos (Science, April 2025). Mélissa Berthet, Martin Surbeck, and Simon Townsend recorded wild bonobos in the Kokolopori Bonobo Reserve in the Democratic Republic of the Congo, coded more than 300 contextual features for each of about 700 utterances, and used methods from distributional semantics to build what Berthet described as a bonobo dictionary of the most common calls. Every call type appeared in at least one compositional combination, and three combinations showed nontrivial compositionality, where one call modifies the other. That directly challenges the claim that nontrivial compositionality is uniquely human. Limits: the evidence is correlational (meaning inferred from contexts, not tested by playback). Townsend himself emphasized that the system is orders of magnitude less complex than human language, and some linguists remain unconvinced that the mathematical criterion captures linguistic compositionality.

Chimpanzees (Science Advances, May 2025). Cédric Girard-Buttoz, Catherine Crockford, Roman Wittig, and colleagues analyzed 4,323 utterances from 53 wild chimpanzees in Taï National Park, Ivory Coast, and found four ways in which 16 different two-call combinations change meaning. Some add meanings (feeding plus resting), some narrow them (an ambiguous feeding-or-travel call plus an aggression-related call is associated with travel), and some are idiom-like, creating meanings unrelated to the parts (resting plus affiliation is associated with nesting). Order also mattered for some pairs. Unlike earlier studies limited to predator alarms, these combinations appeared across everyday life. Limits: again correlational; a 2023 playback study showed chimpanzees process at least one combination differently from its parts, and the new catalog supplies many candidates for similar tests.

Statistical Laws in Whale Song

Inbal Arnon, Simon Kirby, Ellen Garland, and colleagues applied infant-inspired segmentation to eight years of humpback whale song from New Caledonia (Science, February 2025). Cutting the song wherever transitional probability between sound elements dipped produced subsequences whose frequencies followed a Zipfian distribution, the statistical signature found in every human language.

The authors' interpretation is that these properties make systems easier to learn and therefore emerge in any culturally transmitted signal, human or whale. That is a claim about learnability and transmission, not about meaning. Humpback song may be closer to music than to language, and the authors note that music shows Zipfian patterns too. The prediction worth watching is the inverse: species that do not learn their signals should not show these properties.

Meaning Tested by the Animals Themselves

Sarasota dolphins (2025). About half the whistles produced by the Sarasota Bay bottlenose dolphin community are non-signature whistles that had received little study. Laela Sayigh of WHOI, with Peter Tyack, Vincent Janik, and the Sarasota Dolphin Research Program, built a catalog of shared non-signature whistle types (reporting 22 types used by dozens of dolphins) using suction-cup hydrophones during catch-and-release health assessments and digital acoustic tags. In wild playbacks, one whistle (NSW-A, used by more than 25 individuals) typically produced avoidance, suggesting an alarm function. Another (NSW-B) appeared when dolphins encountered something unexpected, suggesting a query function, though responses to its playback were variable. Sayigh describes both functions as working hypotheses. If they hold, dolphins share at least a small set of word-like signals beyond names.

Zebra finches (2025 to 2026). Julie Elie, working in Frédéric Theunissen's lab at UC Berkeley, combined more than a decade of behavioral observation with machine learning to sort more than 8,000 calls from over 30 birds into 11 core call types, with meanings such as distress, hunger, greeting, aggression, and bonding. A classifier using acoustics alone struggled to separate calls that sounded alike but meant different things. Elie then let the birds weigh in. Finches pressed buttons that played call types, with only one type rewarded with seed each day. They learned quickly which type paid off and interrupted the others. The decisive evidence came from their mistakes: birds confused calls with similar meanings more often than calls with similar acoustics, indicating that they categorize calls by meaning. Philosopher Jonathan Birch highlighted that Elie asked the finches themselves whether her dictionary was right. Limits: a captive laboratory population and a repertoire of 11 categories; the meaning labels still come from a human-built ethogram, though the birds' behavior validates their grouping.

The Coller Dolittle Challenge

The Jeremy Coller Foundation and Tel Aviv University launched the Coller Dolittle Challenge for Two-Way Interspecies Communication in 2024, explicitly modeled on the Turing test and chaired by bat researcher Yossi Yovel. It awards US$100,000 each year to the most promising research and offers a grand prize of either a US$10 million equity investment or US$500,000 in cash to a team that achieves genuine two-way communication. Entries are judged on three criteria: non-invasiveness, versatility (communication in more than one context using the animal's own signals), and responsiveness (a measurable response from the animal).

  • 2025 winner: Laela Sayigh's WHOI and Sarasota team, for the non-signature whistle work above.
  • 2026 winner: Julie Elie of UC Berkeley, for the zebra finch work above. The 2026 finalists also included teams working on African striped mice, chimpanzees, and bonobos.

The prize shapes the field by rewarding experiments where animals respond, not just correlations. It also reveals how far away the grand prize remains: the winning entries demonstrate understanding of a handful of signal categories, not conversation.

Soundboard Dogs and Augmentative Interspecies Communication

Viral videos of dogs pressing recorded-word buttons raised obvious Clever Hans worries. A 2024 PLOS ONE study led by Amalia Bastos and Federico Rossano tested 59 soundboard-trained dogs in two experiments. In the first, researchers visited 30 homes; one experimenter hid button labels under colored stickers and another, unaware of which word each button played, pressed them. In the second, owners either pressed buttons or spoke words. Dogs responded appropriately to "play" and "outside" words whether the words came from a button or a person, and not to a nonsense word, showing they attend to the words themselves. Food words produced less consistent responses.

The study establishes word comprehension for a few trained associations. It does not show that dogs compose sentences with their buttons; Rossano's group says that question is the subject of future work.

Ethics and Governance

As capabilities grow, researchers have begun writing rules before the technology outruns them. In November 2024 NYU's More-Than-Human Life (MOTH) Program and Project CETI convened scientists, animal welfare experts, research ethicists, environmental lawyers, data governance specialists, and AI researchers. The resulting principles treat nonhuman animals as subjects rather than objects and reject technology that aims to dominate nature. In late 2025 MOTH published the voluntary PEPP framework (Prepare, Engage, Prevent, Protect), which asks researchers to anticipate harms, monitor them continuously, and respond quickly when they occur.

The concerns are concrete. Gruber has noted that even routine recording and playback can stress animals. In one documented case, researchers studying elephant communication played the recorded call of an individual who had died; the family searched and called frantically, and the dead elephant's daughter kept calling for days afterward. Beyond welfare, the framework addresses data ownership, dual use (hunters, poachers, or fishing fleets exploiting decoded signals), and the risk of flattening rich, contextual expression into human categories. Most current projects deliberately focus on listening rather than talking back. Supporters note that many human rights norms began as voluntary principles before becoming enforceable.

Research Infrastructure and Community

The field now has its own venues. The Simons Institute and Project CETI workshop series reached its fourth edition in July 2025, with talks on dolphins (Denise Herzing and Thad Starner on DolphinGemma), birds (Tim Gardner), sperm whales (Orr Paradise on WhAM), machine translation of human languages (Markus Freitag of Google), Perch and marine bioacoustics (Tom Denton and Lauren Harrell), validation methods (Adam Kalai), and even professional constructed-language creators David and Jessie Peterson. NeurIPS 2025 hosted a dedicated workshop on AI for Non-Human Animal Communications. Public benchmarks such as BEANS, BEANS-Zero, and BirdSet are beginning to give the bioacoustics community shared yardsticks.

Open Questions

  • Meaning at scale. Can validation methods like Elie's operant tests or Sayigh's playbacks scale beyond a dozen signal categories, and to wild animals that cannot be trained?
  • How much of whale coda capacity carries meaning? A large combinatorial inventory may encode identity, clan, emotion, or turn-taking rather than propositional content.
  • Is nontrivial compositionality widespread? Townsend suspects it may be common in primates but doubts it is extensive outside them. Systematic comparisons across species are needed.
  • Do statistical laws diagnose cultural learning? The humpback result predicts that non-learned systems should lack Zipfian structure; that prediction is testable.
  • Will translation without parallel data work in practice? The NeurIPS 2023 theory says it is possible in principle for complex systems with enough common ground. Whether sperm whale communication meets those conditions is unknown.
  • How should playback be governed? Generative models make it easy to produce convincing synthetic calls. The field has not settled when such playbacks are acceptable.
  • Open models and data. Reproducibility depends on access. NatureLM-audio, Perch, and WhAM have open code or weights; other headline models have not been independently evaluated.

Talks and Explainers

Load video
YouTube

Load video
YouTube
Decoding Communication in Nonhuman Species
With advances in machine learning, we now stand a chance to decode communication in non-human animal and plant life, which has fascinated humans for millennia. Cutting-edge technology can be used to benefit not only humankind, but other species on our planet, to better understand and protect the life around us. To make this possible, one needs to bring together and build on the latest discoveries of experts in disparate fields, including machine learning, signal processing, data science, linguistics, and acoustics. The goal of the workshop is to explore the challenges and current state-of-the-art in the study of non-human species communication as well as disciplines and technologies that are likely to provide solutions to these challenges.

Load video
YouTube
How Close Are We to Talking With Animals?
We know about apes learning sign language, but what about dolphins typing on an underwater keyboard? Or A.I. machines translating prairie dog chirps? We want to answer how close we are to conversing with the animal kingdom.

Load video
YouTube
What Information Do the Voices of African Savanna Elephants Communicate
Joyce Poole, ElephantVoices. Poole later coauthored the 2024 Nature Ecology & Evolution study showing that wild African elephants address one another with individually specific, name-like calls.

Load video
YouTube
How artificial intelligence is helping scientists talk to animals - BBC News
Scientists are harnessing the power of artificial intelligence (AI) to decode animal languages. Scientists in Israel are closer than ever to making two-way communication with another species more likely - by using AI to understand the language of bats. “I’ve always dreamt of a Dolittle machine that will allow me to talk with animals”, said Professor Yossi Yovel of Tel Aviv University. The team at Tel Aviv University created a large database of bat noises and videos, teaching AI how to differentiate between the different sounds. The Algorithm then matched the sounds with specific social interactions captured on film. Adi Rachum from Tel Aviv University said, “in the end the computer will be able to speak the language, to understand what they say to each other”. Yovel now chairs the Coller Dolittle Challenge, which rewards progress toward exactly this kind of two-way communication.

Load video
YouTube
Load video
YouTube

Animal Translators

Consumer "animal translators" have a long history and a mixed record. The Bowlingual, a Japanese gadget from Takara that claimed to read emotion from a dog's bark, won the 2002 Ig Nobel Peace Prize and has been relaunched since. Smartphone apps now claim to translate meows and barks. Con Slobodchikoff, whose research on prairie dog alarm calls is described in Chasing Doctor Dolittle, has pursued a dog translator commercially, and the 2018 NBC article above predicted a pet translator within a decade.

As of late 2026, no consumer product has published peer-reviewed evidence that it translates animal vocalizations into specific meanings. The research reality is narrower and more interesting: classifiers can detect species, individuals, contexts, and arousal levels from calls, and in a few species carefully controlled experiments have validated a small number of signal meanings.

Claims versus evidence
Claim you will see What the research actually supports
"AI can translate what your dog or cat is saying" Classifiers can estimate context or arousal from some calls; no validated word-level translation exists
"Whales have an alphabet" Sperm whale codas have combinatorial and vowel-like structure; meanings are unknown, and the "alphabet" label is disputed
"Elephants have names" Elephant calls contain receiver-specific, non-imitative labels; the label has not been isolated within the call
"Scientists talked to dolphins" CHAT teaches dolphins human-made synthetic whistles for objects; WHOI playbacks suggest shared whistles with alarm and query functions
"Soundboard dogs speak in sentences" Trained dogs understand some button words without cueing; sentence production is untested
"Plants scream when hurt" Stressed plants emit informative ultrasonic clicks; there is no evidence these are signals aimed at anyone

Load video
YouTube
AI 'Google Translate' for Animal Speech! - Next Big Thing in AI
AI is being used to both decode and replicate animal sounds and behaviors, aiming to enable communication between us and our furry, scaly, or feathery friends. Pioneering researchers are developing AI translators for animals like prairie dogs and dolphins, translating their unique sounds into English. On the flip side, AI is also generating animal-esque sounds and behaviors, with projects like David Hanson's "Animatronic" for Sophia the robot and Stanford's "Earth Species Project" creating realistic animal avatars that learn and evolve. (Editor's note: Earth Species Project is an independent nonprofit, not a Stanford project, and its published research centers on audio models such as NatureLM-audio rather than animal avatars.)

Load video
YouTube
Translating canine communication with computer science
Researchers at North Carolina State University are inventing technology to decode dog talk. Hari Sreenivasan visits a computer science lab that has designed a harness to monitor physiological and emotional changes and send wireless commands through vibrations, which could be used with guide animals or search and rescue dogs.

Load video
YouTube
Apps That Can "Translate Animal Sounds"
Looking apps that can do things that phones literally can't do like translating cat meows into English, Xray scans and more. No hate to these apps they all say in the description that they can't actually do these things, other than the cat one. It’s a actually a good translate app other than that though

Load video
YouTube
Woof! Dog translator gets upgrade
The Bowlingual, a gadget that analyzes a dog's bark to detect its emotion, is being relaunched. The gadget can supposedly recognize and convey 6 basic emotions from a dog's bark.


Translator - Saturday Night Live (SNL)

Three scientists (Scarlett Johansson, Kyle Mooney, Mikey Day) receive a shock when they debut their invention, a machine that translates for pets.

Load video
YouTube
Translator - SNL
Three scientists (Scarlett Johansson, Kyle Mooney, Mikey Day) receive a shock when they debut their invention, a machine that translates for pets.

Dr. Dolittle

Doctor John Dolittle is the central character of a series of children's books by Hugh Lofting starting with the 1920 The Story of Doctor Dolittle. He is a physician who shuns human patients in favour of animals, with whom he can speak in their own languages. He later becomes a naturalist, using his abilities to speak with animals to better understand nature and the history of the world. Doctor Dolittle | Wikipedia

The character remains the field's shorthand for its ultimate goal. Researchers routinely invoke a "Dolittle machine," and the largest prize in the field, the Coller Dolittle Challenge, is named after him. The fictional version and the scientific version differ in one important way: Dolittle simply understood animals, while real researchers must prove, experiment by experiment, that the categories they find matter to the animals.

Load video
YouTube
Doctor Dolittle (1967) ORIGINAL TRAILER
The original trailer of Doctor Dolittle directed by Richard Fleischer and starring Rex Harrison, Samantha Eggar, Anthony Newley and Richard Attenborough.

Load video
YouTube
Dolittle - Official Trailer
He's just not a people person

Load video
YouTube
Dr. Dolittle (1998)
Successful physician and devoted family man John Dolittle (Eddie Murphy) seems to have the world by the tail, until a long-suppressed talent he possessed as a child - the ability to communicate with animals - is suddenly reawakened with a vengeance! Now every creature within squawking distance wants the good doctor's advice, unleashing an outrageous chain of events that turn his world upside down!

Load video
YouTube
Richard Dawkins - Science in the Soul
Richard Dawkins talked about his book Science in the Soul: Selected Writings of a Passionate Rationalist, in which he makes an argument for the “scientific way of thinking”.

Further Learning

Research Sources

Sperm whales and Project CETI

AI models for animal sound

Names, compositionality, and statistical structure

Meaning tests and the Coller Dolittle Challenge

Ethics, governance, and community

Additional Viewing

Load video
YouTube

Long-Term Work and Machine Learning Approaches to DolphinGemma | Denise Herzing (Wild Dolphin Project) and Thad Starner (Georgia Tech and Google Research), Simons Institute, July 31, 2025. Herzing and Starner explain how four decades of underwater observation of Atlantic spotted dolphins became the training data for DolphinGemma and how the CHAT wearable uses synthetic whistles to explore a shared vocabulary. Recorded in an hour-long workshop slot, it is the most direct source on what DolphinGemma does and how it is being used in the field. Watch on YouTube

Load video
YouTube

WhAM: Towards A Translative Model of Sperm Whale Vocalization | Orr Paradise (UC Berkeley and Project CETI), Simons Institute, July 30, 2025. Paradise presents the Whale Acoustics Model, which fine-tunes a music-trained masked acoustic token model to generate and restyle sperm whale codas. The talk is useful for understanding how generative models support playback experiments and why generation is not the same as translation. Watch on YouTube

Load video
YouTube

Overview and Marine Bioacoustics and Perch Model | Tom Denton (Google DeepMind) and Lauren Harrell (Google Research), Simons Institute, July 31, 2025. A roughly half-hour session on the Perch bioacoustic foundation models, agile modeling for building new detectors from a few examples, and why a model trained mostly on birds transfers to whale sounds. It shows where machine learning already delivers real value for conservation monitoring. Watch on YouTube

Load video
YouTube

2021 Wild Ocean Science: Technology | Wild Dolphin Project, hosted by PBS NewsHour correspondent Miles O'Brien. A multi-speaker event covering underwater localization of dolphin sounds with the ASPOD device, 360-degree camera work, and an update from Thad Starner and Denise Herzing on machine learning pattern discovery and the underwater computer interface for two-way exploration. It provides the pre-DolphinGemma background for the CHAT project. Watch on YouTube

Load video
YouTube

Dolphins: Breaking the Code (Full Episode) | Changing Seas, South Florida PBS, 2018. This documentary follows Denise Herzing's in-water research on Atlantic spotted dolphins in the Bahamas and the technology used to correlate their sounds with behavior. It is a clear introduction to why long-term field context is the foundation of every decoding effort. Watch on YouTube

Retrieved from "https://primo.ai/w/index.php?title=Animal_Language&oldid=48528"