The wearable computer behind the effort to communicate with dolphins

One of the tools behind dolphin-communication research is a computer that a person has to wear underwater. That is the part of the problem Georgia Tech research scientist Scott Gilliland helps solve: not simply recognising a sound, but building electronics and a physical form that researchers can use in the sea.[1]

In an August 2025 profile by Georgia Tech’s Julian Hills, Gilliland described his preference for making early prototypes: “I’m good at making the first, second, and third prototypes.” His work combines circuit boards, embedded software and physical fabrication—skills that sit between an experimental idea and a usable instrument.[1]

A question that needed a new instrument

Gilliland and wearable-computing researcher Thad Starner joined the Wild Dolphin Project collaboration in the early 2010s. Denise Herzing needed a way to play sounds underwater and record high-frequency responses. The resulting CHAT system—Cetacean Hearing Augmentation Telemetry—was developed for attempted two-way interaction, with Gilliland refining its hardware and wearability.[1]

The engineering challenge can be understood without imagining an already translated conversation. A researcher needs to hear the relevant event, know that the device detected it and respond while the interaction is still happening. A recording that becomes interpretable only much later can answer some research questions, but it cannot support the same immediate exchange.

The deliberately simpler experiment

Google’s April 2025 account describes CHAT using synthetic whistles associated with objects such as a scarf or seaweed. The aim is to create recognisable shared signals during interaction. The device can detect a target sound and notify the researcher through bone-conduction audio, helping them respond with the associated object.[2]

That arrangement starts with a meaning supplied by the experimenters. It is different from taking a naturally occurring dolphin sound and discovering its meaning. Keeping the two jobs separate makes the research more intelligible: one asks whether a shared signalling convention can develop; the other investigates the animals’ existing communication.

A hypothetical session record illustrates what the instrument needs to preserve. It would link a played signal, the object available, a sound detected afterwards and the researcher’s response. The order matters. A sound produced before an object appears and the same sound produced afterwards are different observations, even when an audio classifier gives them the same label.

Where DolphinGemma fits

DolphinGemma is a separate audio model described in the same 2025 research account. It uses roughly 400 million parameters and learns from the Wild Dolphin Project’s recordings of Atlantic spotted dolphins, with contextual information built through its long-running fieldwork. Its task is to model patterns in dolphin audio, not to supply a verified dictionary of dolphin-to-English meanings.[2]

The two approaches can be complementary. Audio modelling can identify regularities worth investigating. An interaction device can let researchers investigate selected signals while preserving the surrounding event. Better prediction and better experimental access contribute different pieces of the work.

The chronology here is the published 2025 account, not an announcement of a newly available consumer translator. What makes the story compelling is the collaboration: field biologists define a question, audio researchers build a model, and someone has to make the electronics function on a human body in the ocean.

Gilliland’s role makes that last contribution visible. Before a research programme can collect the interaction it wants to study, its instrument has to let the people doing the work take part in it.

Report an error or suggest a correction