DOllphingemma, the new extension of the Gemma of Google series-a range of open-source artificial intelligence models-represents a major tool in a long-standing scientific effort: understand how dolphins communicate with each other. The model was developed in collaboration with researchers from Georgia Tech and Wild Dolphin Project (WDP), a non -profit organization that has been studying the Tired Dolphins of the Atlantic and their behavior since 1985. Thanks to AI, the objective is today to go beyond observation, by orienting itself towards a real interaction.
The linguistic model has been drawn into a combination of human linguistic data and a large set of dolphin sound recordings. The data comes from both the Gemma series from Google and the WDP archives. Unlike traditional language models focused only on text, Dolphingemma is interested in sound patterns characteristic of dolphins, with the aim of recognizing them, analyzing them and even reproducing them in a synthetic way:
- Whistles used by mothers and little ones to meet
- Pulsed cries often heard during conflicts
- Clicks produced during the nuptial parade or the hunt for sharks
Field application
What makes Dolphingemma particularly useful in current research is that it is light enough and efficient to operate on a smartphone. This allows researchers to listen to the sounds issued by dolphins live, to compare them to the outputs of the model, and potentially send in return of “Dauphin type” sounds generated by AI.
This summer, the Wild Dolphin Project plans to use Google Pixel 9 smartphones to create a platform capable of generating synthetic vocalizations and capturing dolphins sounds in order to produce suitable responses. The Pixel 6 had been used for the first experiments, but the Pixel 9 will allow both IA models to be carried out and real -time patterns recognition algorithms.
A major advance for research
Dolphins have long been considered one of the most intelligent species on the planet, with complex social structures and an apparently organized communication system. However, their “language” is still largely misunderstood. AI models like Dolphingemma could make it possible to detect patterns or even decipher meanings in their sounds. It would be a revolutionary advance in the field.
This project illustrates the potential of open-source AI in specialized fields, ranging from text generation to exploration of animal language, on a common technological basis.
