Ever wanted to communicate with dolphins? Google DeepMind unveils DolphinGemma: an AI model that analyzes and translates dolphin language.
Dolphins have a complex communication system. Every click or whistle the animal makes has a specific meaning. Dolphin researchers dream of being able to decipher this language, and with Google DeepMind’s latest AI model, that dream might become a reality. In a blog post, DeepMind showcases DolphinGemma: an AI model trained to understand dolphin language.
DolphinGemma is the result of a collaboration between Google and dolphin scientists. The model builds upon Google’s Gemma model and is specifically designed for audio analysis. It has been trained on a unique dataset containing thousands of hours of underwater audio, linked to specific behaviors and individual dolphins.
Whistles and Clicks
The researchers have spent years documenting various types of dolphin sounds, such as whistles and clicks, and the behaviors associated with those sounds. By feeding this data to the AI model, researchers can detect recurring sound structures more quickly.
Thanks to its ‘modest’ size (400 million parameters), the Dolphin-Gemma model can run locally on smartphones used by researchers in the field. The latest version runs on the Pixel 9 and can simultaneously execute AI models and pattern recognition. Special equipment was developed to capture and analyze dolphin sounds in real-time.
Talking to Dolphins
The goal is to use this model to gain a better understanding of possible meanings behind dolphins’ natural communication. But it can also work in the reverse direction: researchers can use AI to develop synthesized dolphin sounds to create a shared vocabulary between humans and animals.
Using an underwater computer, researchers can create whistle signals linked to objects that the animals find interesting, such as seaweed or toys. The computer listens to which whistle signal the animal mimics and links it to the correct object, so the dolphin gets what it ‘asks’ for. By better understanding structural patterns in dolphin sounds, meaningful interaction between humans and animals comes closer.
Google plans to release DolphinGemma as open source this summer, allowing other researchers to use it as well. Although the model is trained on the sounds made by a specific dolphin species, it can be adapted to other species through additional training.