More than a decade ago, Google had played an April Fool’s prank on users — it had demonstrated a technology that would enable users to talk to animals. The technology had been made up, but it had anyway elicited wonder and amazement from users until they’d been told that it was all a joke. In 2025, Google has actually released a technology that could help us talk to animals.
Google has unveiled DolphinGemma, an AI model designed to analyze and generate dolphin vocalizations, potentially enabling humans to “converse” with these intelligent marine mammals. Announced on National Dolphin Day in collaboration with Georgia Tech and the Wild Dolphin Project (WDP), DolphinGemma leverages decades of field research to decode the clicks, whistles, and burst pulses that form dolphin communication. This innovation, detailed in a recent Google blog, could reshape marine biology, AI applications, and our understanding of animal intelligence.
DolphinGemma is built on Google’s Gemma framework, a lightweight, open-source model derived from the same technology powering its Gemini AI systems. With approximately 400 million parameters, it’s optimized to run on Google Pixel phones used by WDP researchers in the Bahamas, where they study wild Atlantic spotted dolphins. Trained on WDP’s extensive dataset—decades of underwater audio and video linked to individual dolphin behaviors—the model processes natural dolphin sounds to identify patterns and predict subsequent vocalizations, akin to how language models predict human speech.
This research is already making an impact in the real world. WDP, which has conducted the world’s longest-running underwater dolphin study since 1985, can now use DolphinGemma to uncover hidden structures in dolphin communication that previously demanded exhaustive human analysis. For instance, the model can cluster recurring sound sequences—like signature whistles used by mothers to call calves or burst-pulse squawks signaling conflict—potentially revealing a grammar-like system. This could accelerate discoveries about dolphin social dynamics and cognition, offering businesses and tech firms insights into bio-inspired AI systems.
Beyond analysis, DolphinGemma supports WDP’s experimental two-way communication efforts through the CHAT (Cetacean Hearing Augmentation Telemetry) system, developed with Georgia Tech. CHAT associates synthetic whistles with objects dolphins enjoy, such as seagrass or scarves, aiming to build a shared vocabulary. Researchers use Pixel phones—soon upgrading to Pixel 9—to detect dolphin mimics in real-time amidst ocean noise, enabling rapid responses that reinforce learning. DolphinGemma’s predictive capabilities enhance CHAT’s efficiency, making interactions smoother and more scalable.
Google plans to release DolphinGemma as an open model this summer, inviting global researchers to adapt it for other cetaceans, like bottlenose or spinner dolphins.
Interestingly, the very first LLMs were built as translation tools — the seminal transformer paper, also published by Google, had envisioned how transformers could be useful for translation between languages. Transformers ended up creating the AI revolution, and were the cornerstone of modern LLMs which can write code, diagnose diseases, and write IPO documentation and legal briefs. But it now turns out that these models can be extended beyond human languages, and help us understand the languages of other creatures on our planet.