Off the coast of Dominica, in the deep water east of the island where sperm whales gather in multigenerational family units, researchers float on the surface and lower hydrophones into the dark below. The water is warm. The clicking starts almost immediately — short, rhythmic sequences called codas, each lasting under a second, exchanged between animals that can weigh up to 60 tons. It has the quality of conversation. Listening to the recordings, even without any training, you get that sense — something is being communicated, passed back and forth, acknowledged. But what exactly has remained one of biology’s most persistent open questions.
Machine learning is now the tool that researchers believe might finally provide answers, or at least a framework for beginning to find them. Project CETI — the Cetacean Translation Initiative — has been building what it calls the Whale Acoustics Model, or WhAM, which converts sperm whale clicks and codas into structured mathematical representations that algorithms can analyze for patterns.
The work is conceptually similar to what large language models do with human text: look for recurring structures, identify combinatorial rules, map out what functions like grammar. The difference is that the dataset is entirely acoustic, the subjects don’t know they’re being studied, and nobody started with a translation dictionary.
The hardware involved is more interesting than it might sound in a grant proposal. Harvard researchers designed suction-cup bio-loggers — non-invasive tags deployed via drone — that attach to individual whales and record not just the sounds they make but the sounds they hear, allowing researchers to track who is speaking to whom within a pod.
Autonomous underwater vehicles follow whale groups over longer distances, capturing behavioral data that gets paired with the audio. The resulting datasets are multi-modal: sound, position, movement, social context, all timestamped and aligned. That combination is what makes the machine learning analysis genuinely more powerful than what was possible before.
What the models have found is striking without being conclusive. Sperm whale codas have a combinatorial structure — whales appear to build complex signals from smaller units in ways that look mathematically similar to how syllables combine into words in human speech.
Individual whales seem to have something like names, or at least consistent identifying patterns in their vocalizations. Regional dialects exist across different populations. Whales take conversational turns. These findings suggest sophistication that’s difficult to explain as simple signaling, but they don’t yet tell us what any specific exchange actually means.
