DolphinGemma

A large language model that uses dolphin audio to help scientists study how dolphins communicate — and hopefully find out what they're saying, too.

Developed in collaboration with researchers at the Wild Dolphin Project and the Georgia Institute of Technology, DolphinGemma could help us better understand dolphins and, one day, hopefully even help us communicate with them.


Capabilities

We’re training DolphinGemma to learn the structure of dolphin vocalizations and generate novel dolphin-like sound sequences.

edit_audio

Audio-in, audio-out

Processes sequences of natural dolphin sounds to identify patterns, structure, and ultimately predict the next sounds in a sequence.

waves

Uncovering potential meaning

Identifies recurring sound patterns, clusters, and reliable sequences, to help uncover hidden structures and, potentially, meanings behind dolphin sounds.

code

Establishing a shared vocabulary

With the help and expertise of the Wild Dolphin Project (WDP), along with its library of natural dolphin vocalizations, DolphinGemma may help establish two-way communication with dolphins one day.


Training DolphinGemma

The WDP has spent decades studying a community of wild Atlantic spotted dolphins. This research has produced a unique dataset of underwater audio and video that is cataloged by individual dolphins and the behaviors of each individual. DolphinGemma was trained on that audio dataset, allowing the model to identify recurring patterns and structures in dolphin vocalizations.

Eventually, we hope the model will be able to predict the next sounds in a given sequence of dolphin noises – much like how LLM models can predict the next word in human languages.



Partners


Building for the research community

DolphinGemma is currently in development. On release, it will be openly available – to give researchers a powerful tool to collectively deepen our understanding of intelligent marine life.

Until then, all the other open models in our Gemma collection are ready for you to explore, adapt, and build on.