Project CETI Is Using Machine Learning to Decode Sperm Whale Language

Sperm whales have the largest brains of any animal that has ever lived, six times heavier than ours. They live in matrilineal family units that speak in dialects, and they exchange information using bursts of clicks called codas. For decades, biologists suspected these vocalizations carried meaning, but the data was too vast and the patterns too subtle for human analysts to crack. Project CETI, a nonprofit research collaboration launched in 2020 and funded in part by TED's Audacious Project, is using machine learning to do what unaided ears never could: work out the structure of how another species communicates, and maybe one day what it is saying.

The stakes reach beyond biology. Understanding how a non-human intelligence communicates would reshape how humans think about minds unlike our own, and conservationists argue it would strengthen the case for protecting the ocean's deepest divers. Sperm whales face entanglement in fishing gear, ship strikes, plastic pollution, and noise from shipping and seismic surveys that interferes with the very channel they use to talk.

Listening Machines in the Caribbean

CETI's field work is anchored off Dominica, a small island nation in the Eastern Caribbean that is home to a resident population of sperm whales. The project deploys an unusual toolkit: hydrophone arrays that record the whales continuously, soft robotic fish that glide alongside pods collecting audio and video, and biologging tags that suction onto individual whales and capture their clicks, depth, and movement during deep dives.

The recordings are then processed by machine learning systems developed with researchers at institutions including Harvard University and the University of California, Berkeley. One of the core engineering problems is simply hearing the whales at all. The open ocean is loud, and sperm whale clicks are short pulses buried in waves, rain, boats, and the crackle of snapping shrimp. CETI's detection models solve this with 99.5 percent accuracy, pulling individual whale clicks out of background noise. Classification models then sort the clicks into known pattern types with 97.5 percent accuracy, and further models identify which vocal clan a whale belongs to with 95.3 percent accuracy, effectively reading whale dialects.

A Phonemic Alphabet for Whales

The project's landmark scientific result arrived in 2024. Analyzing a large dataset of codas recorded from the Eastern Caribbean clan, CETI researchers showed that sperm whale clicks combine into a structured repertoire that behaves like a phonetic alphabet. The whales build their calls from a set of rhythm and tempo building blocks, the way human languages build words from phonemes. The team catalogued 156 distinct codas and their basic components in this work.

The study also found two features that startled linguists. The first, called rubato, is when a whale stretches or compresses the overall duration of a coda while keeping its internal rhythm intact. The second, ornamentation, is the addition or slight modification of an extra click at the end of a coda. Crucially, these variations carry information in context rather than acting as noise, and they mean the number of distinct messages whales can construct is vastly larger than a fixed catalogue of call types. A combinatorial system with sub-second, deliberate adjustments is one of the hallmarks biologists look for when distinguishing a communication system from a simple set of fixed signals.

Why AI Was the Missing Instrument

Whale researchers before CETI faced an impossible data problem. A single hydrophone array can capture hundreds of thousands of clicks per day. A human analyst working by ear and spectrogram can label perhaps a few hundred codas in a day, and cannot perceive millisecond-scale timing differences at all. Machine learning changes the scale of the question. Models trained on labeled recordings can process years of audio in hours, detect patterns invisible to human perception, and test whether variations in clicks correlate with behavior, context, or the identity of the caller.

The work also draws directly on techniques built for human language. CETI collaborates with experts in natural language processing, and the team treats whale codas the way a linguist might treat an unknown human language: collect a massive corpus, find the units, measure how they combine, and test whether combinations follow rules rather than chance. The phonetic alphabet result came from exactly this approach, and it is the first time such a system has been documented in a non-human species at this level of detail.

From Understanding to Protection

CETI's vision, laid out when the project was selected as a TED Audacious Project, is ambitious: to gather enough behavioral and acoustic data, on the order of four billion clicks, to test whether meaningful two-way communication with sperm whales is possible. The team is careful to frame this as a long scientific road rather than a promised translation app, and external scientists have raised fair cautions about overinterpreting structure as language. But the conservation payoff is already arriving. Dominica has moved to protect sperm whale habitat in its waters, and researchers credit the growing scientific and public attention driven by CETI's work. The project also shares its recording and analysis tools, and its methods are being applied to whale populations in the Pacific, where evidence of social learning across vocal clans has been documented using similar acoustic analysis.

There is a deeper argument that CETI's founders make. If humans could one day demonstrate that sperm whales communicate structured information about their world, the moral case for protecting them and the ocean they live in becomes harder to ignore. Knowing what we are losing changes what we are willing to do about it.

The Widest Lens We Have Ever Pointed at Another Mind

Every previous attempt to understand animal communication was limited by the bandwidth of human attention. Machine learning removes that ceiling. A network of hydrophones, a fleet of robotic listeners, and models that never tire of listening to clicks are, together, the widest lens our species has ever pointed at another mind. Whether or not full translation ever comes, we are already seeing structure no one knew was there: an alphabet in the deep, spelled out one click at a time, by animals that have been talking for millions of years while we were only starting to learn how to listen.