Google’s New AI Model Helps Humans Communicate with Dolphins

Google, in collaboration with Georgia Tech and the Wild Dolphin Project (WDP), has launched DolphinGemma, an AI model developed to analyse and generate dolphin vocalisations. The announcement, made on National Dolphin Day, marks a new step in the effort to understand and potentially engage in two-way communication with dolphins.

DolphinGemma is trained on decades of underwater video and audio data from WDP’s long-term study of Atlantic spotted dolphins (Stenella frontalis) in the Bahamas. The model identifies patterns in dolphin vocal sequences and generates realistic dolphin-like sounds. 

“By identifying recurring sound patterns, clusters and reliable sequences, the model can help researchers uncover hidden structures and potential meanings,” Google said.

The AI system is based on Google’s lightweight Gemma models and leverages SoundStream for audio representation. At roughly 400 million parameters, the model is small enough to run on Pixel phones used in field research. It operates as an audio-in, audio-out system, designed to predict subsequent dolphin sounds much like a language model predicts the next word in human text.

“Understanding any species requires deep context,” the WDP team said. Their work since 1985 has documented individual dolphins’ life histories and linked vocalisations to specific behaviours. Signature whistles are used to reunite mother and calf pairs, while burst-pulse squawks and click buzzes are associated with fighting and courtship.

In parallel, the WDP and Georgia Tech have also developed the CHAT (Cetacean Hearing Augmentation Telemetry) system. CHAT uses synthetic whistles linked to objects like sargassum or scarves. If dolphins mimic these whistles, the system can recognise them and inform the researcher via bone-conducting headphones, enabling real-time interaction.

“A Google Pixel 6 handled the high-fidelity analysis of dolphin sounds in real time,” researchers explained. The next iteration, built around a Pixel 9, will integrate deep learning and template matching directly into the device, reducing the need for specialised hardware.

DolphinGemma will be released as an open model this summer. Although currently trained on Atlantic spotted dolphins, it is expected to benefit research into other cetaceans. Fine-tuning will be needed for species with different vocal patterns.

“We’re not just listening anymore. We’re beginning to understand the patterns within the sounds,” Google stated. The goal of the research is to bridge the gap between humans and dolphins through data-driven insights and shared interaction systems.