









For decades, decoding dolphin clicks, whistles, and bursts has been a scientific challenge. Now, Google, in collaboration with Georgia Tech and the Wild Dolphin Project (WDP), is introducing DolphinGemma — an AI model trained to learn the structure of dolphin communication and generate dolphin-like sounds. The goal: to move closer to real interspecies dialogue.
A Foundation Built on Decades of Research
Since 1985, WDP has led the world’s longest-running underwater dolphin study, documenting generations of wild Atlantic spotted dolphins in the Bahamas. Their non-invasive fieldwork — capturing underwater audio, video, and behavior data — has created a unique dataset that links sounds to specific social contexts like reunion calls, fights, and courtship.
Training DolphinGemma
Using this rich dataset, DolphinGemma analyzes natural dolphin sounds to recognize patterns and predict the next sounds in a sequence, similar to how language models predict the next word. Built with Google’s SoundStream and Gemma model technologies, this 400M-parameter system can even run on Pixel phones used by researchers in the field.
Toward Two-Way Communication
By uncovering recurring sound patterns and potential meanings, DolphinGemma can help identify the underlying “grammar” of dolphin communication. This research complements WDP’s CHAT (Cetacean Hearing Augmentation Telemetry) system — an underwater interface that teaches dolphins to associate synthetic whistles with specific objects. Together, DolphinGemma and CHAT could pave the way for two-way conversations between humans and dolphins.
Reference: Google Blog: DolphinGemma: How Google AI is helping decode dolphin communication.