Dolphins are considered some of the smartest creatures on the planet. They can collaborate, teach each other new skills, and recognize themselves in a mirror. For decades, scientists have been trying to understand the complex set of whistles and clicks that dolphins use to communicate. It looks like researchers could soon make significant progress with the help of a new open-source AI model from Google and Pixel smartphones.
In search of new areas for the application of generative AI, Google is interested in collaborating with the Wild Dolphin Project (WDP). This group of researchers has been studying the behavior and relationships of the Atlantic spotted dolphin community since 1985.
One of the main goals of the WDP is to analyze the influence of the sounds dolphins make on their social interactions. Understanding the structure and patterns of dolphin “vocal communication” is necessary to determine whether it reaches the level of a full-fledged language. “We don’t know if animals have words,” says Denise Herzing from the WDP.
Using decades of underwater recordings, the researchers have been able to associate certain basic behaviors with specific sounds. For example, they have identified distinctive whistles that appear to act as names, allowing two individuals to find each other over long distances. Dolphins also routinely emit sound patterns, dubbed “squawks,” during conflicts.
WDP’s ultimate goal is to speak dolphin language, if one really exists. That quest has resulted in a massive, heavily labeled dataset that is too labor-intensive for humans to parse, but Google says is ideal for generative AI.
A New Model for Underwater Language
DolphinGemma, a generative AI model based on Google’s open-source Gemma AI models, uses Google’s SoundStream audio technology. The model was trained using the Wild Dolphin Project’s acoustic archive. The team hopes that DolphinGemma will help uncover complex patterns that will lead to a shared vocabulary.
DolphinGemma works in the same way as human-based language models: it receives a query and predicts the next token. The scientists believe that these predicted tokens could be sound sequences that dolphins would understand.
Google developed DolphinGemma with the WDP research approach in mind. The scientists use Pixel phones in the field, which places additional demands on the model. Running AI models on a smartphone is a challenging task due to limited resources. The larger and more capable the model, the more RAM and computing power it requires. DolphinGemma is trained on about 400 million parameters, which is quite small compared to typical modern models, which have tens of billions of parameters.
For the past few years, WDP has been using a device built at the Georgia Institute of Technology called CHAT (Cetacean Hearing Augmentation Telemetry), which is based on a Pixel 6 smartphone. The team uses CHAT to record, listen to, and synthesize dolphin voices.
For the 2025 summer research season, Google will provide the WDP team with Pixel 9 smartphones, which will allow CHAT to simultaneously run deep learning models and template-matching algorithms. The scientists do not currently plan to feed DolphinGemma-generated sounds directly to CHAT, meaning that both approaches will be used in parallel.
Progress, Possibilities, and What’s Next
Of course, it is unlikely that DolphinGemma and the new CHAT will immediately allow people to easily communicate with dolphins, but the hope is that over time, the system will provide at least basic interaction. The use of generative AI in this context represents a significant step forward in decoding animal communication. Like other Gemma models, DolphinGemma is an open-source project. Google will make the model publicly available this summer, adds NIX Solutions. While DolphinGemma was trained on the sounds of Atlantic spotted dolphins, Google suggests it could be adapted for other cetacean species.
The ongoing collaboration between Google and the Wild Dolphin Project demonstrates how cutting-edge technology and long-term field research can come together to explore uncharted areas of biology. It also highlights the potential of compact, efficient AI models to deliver insights from massive datasets collected in natural environments. DolphinGemma, despite its relatively small size in the world of AI models, is built to run on consumer hardware while supporting sophisticated acoustic analysis.
With every summer season, the WDP collects more recordings and behavioral data, gradually expanding the dataset that fuels DolphinGemma. As the model evolves and adapts, researchers hope it will continue to generate meaningful correlations between sound patterns and social behaviors. While full understanding may still be years away, tools like DolphinGemma offer a promising path forward.
We’ll keep you updated as more integrations become available and as the model continues to mature. As more open-source resources are shared and refined, other researchers may join the effort to interpret marine mammal communication across species. The study of dolphin language is far from complete, but projects like this one bring us closer to meaningful interspecies dialogue.