Google develops new AI model to decode dolphin sounds
15 Apr 2025

Google has created a new artificial intelligence (AI) model, called DolphinGemma, to analyze dolphin vocalizations.

The new tool is part of Google's effort to integrate generative AI into its various projects.

The tech giant's partnership with the Wild Dolphin Project (WDP), a group that has been studying dolphins since 1985, serves as the foundation for this groundbreaking project.

Wild Dolphin Project's role in dolphin communication research
Research collaboration

The WDP has been employing a non-invasive technique to observe a particular pod of Atlantic spotted dolphins. They record videos and sounds of the animals, along with meticulous notes on their behavior.

One of the primary goals of the project is to understand how dolphin calls affect their social behavior.

Over decades, researchers have associated certain basic movements with particular sounds made by the dolphins.

Understanding dolphin vocalizations and social interactions
Vocalization analysis

The WDP's research has shown that Atlantic spotted dolphins use unique whistles, like names, for identification and recognition.

They also produce consistent "squawk" sound patterns during conflicts.

WDP researchers believe understanding the structure and patterns of these vocalizations is key to determining if their communication can be classified as a language.

"We do not know if animals have words," says Denise Herzing from WDP.

DolphinGemma: A breakthrough in dolphin communication
AI model

DolphinGemma, built on Google's Gemma open AI models, aims to predict patterns in dolphin communications.

The model leverages a Google-developed audio tech called SoundStream to tokenize dolphin vocalizations. This way, the sounds can be processed as they're recorded.

Google trained the model on WDP's acoustic archive and it works just like human-centric language models by predicting subsequent tokens from an initial dolphin vocalization input.

DolphinGemma: A smartphone-compatible AI model
Compatibility

DolphinGemma was specifically designed with WDP's research methodology in mind.

Since the team uses Pixel phones in their fieldwork, an efficient model was necessary due to limited resources on smartphones.

Despite being compact (just about 400 million parameters), it still proves to be a powerful tool for studying dolphin communications.

The team has been using a device called CHAT based on Pixel 6 for creating synthetic dolphin vocalizations and responding to them.