DolphinGemma: Google’s AI could let us talk to dolphins

An interspecies communication breakthrough: Google develops an AI to decode dolphin vocalizations

A voice from the deep: how dolphins might speak through AI

DolphinGemma, Google’s ambitious new project, may become a turning point in how we understand animal communication. Created in collaboration with the Wild Dolphin Project (WDP), active since 1985, and the Georgia Institute of Technology, this innovation aims to decode and even generate dolphin vocalizations. The goal? To translate underwater sound into meaning and perhaps spark a real dialogue between humans and dolphins.

The project focuses on a specific community: the wild Atlantic spotted dolphins of the Bahamas. Years of underwater field research have produced a large archive of audio and video recordings linked to individual dolphins, revealing patterns like the signature whistles used by mothers and calves to reunite. Now, with the help of artificial intelligence, scientists are setting their sights even higher.

Inside DolphinGemma: high-performance audio AI

At the heart of DolphinGemma lies a powerful audio AI model, based on Google’s SoundStream tokenizer, designed to process and understand complex sound sequences. The model, which includes around 400 million parameters, has been trained on WDP’s vast acoustic database, one of the world’s richest resources on cetacean communication.

This AI technology can analyze the structure of dolphin vocalizations, detect repeating patterns, predict what sounds are likely to come next, and even generate new sounds that mimic real dolphin speech. It works similarly to language models for human communication, which anticipate words in a sentence—but this time, the aim is to build a shared vocabulary between species.

Interactive conversations: the challenge of the Chat system

Beyond decoding natural dolphin communication, the Wild Dolphin Project is also testing a bidirectional communication system for use in the ocean. This is where Chat (Cetacean Hearing Augmentation Telemetry)comes in—a submersible computer developed with Georgia Tech that emits synthetic whistles tied to objects dolphins enjoy, like sargassum or seagrass.

The hope is that dolphins will associate these sounds with specific objects and begin to imitate the whistles to “ask” for them, opening the door to meaningful interaction between species. If successful, this experiment could lead to the first real semantic exchange between humans and marine mammals.

An open source future for marine science

Google plans to release DolphinGemma as open source software this summer, allowing scientists worldwide to access the model and apply it to other dolphin species, such as bottlenose dolphins or spinner dolphins. This move could significantly accelerate research in the field of interspecies communication.

Ultimately, DolphinGemma is not just a technical feat—it’s a step toward empathy, connection, and mutual understanding between humans and the animal world. With its help, we might finally begin to understand the voices that have echoed through our oceans for millennia.