DolphinGemma: What Google's Dolphin AI Does (and Doesn't)

DolphinGemma is an AI model from Google that learns the structure of dolphin sounds: you feed it a sequence of clicks and whistles, and it predicts what's likely to come next, or generates new dolphin-like sounds. Google built it with researchers at Georgia Tech and trained it on recordings from the Wild Dolphin Project, and announced it on National Dolphin Day, April 14, 2025. Two things up front, because they're what most people want to know: it doesn't translate dolphins into English, and you can't download it yet. Google DeepMind's page still lists it as in development. If you want to know what it actually does, and why it's still a big deal, stay with me.

Who's behind DolphinGemma

The announcement on Google's blog is signed by Dr. Denise Herzing, founder and research director of the Wild Dolphin Project (WDP), and Dr. Thad Starner, a research scientist at Google DeepMind and a professor at Georgia Tech. That pairing says a lot: the data comes from the WDP, the CHAT hardware from Georgia Tech and the AI from Google.

The data is the part nobody else has. Since 1985, the WDP has been studying one community of wild Atlantic spotted dolphins in the Bahamas, underwater, across generations. Its archive is decades of video and audio, matched to individual dolphins, their life histories and what they were doing at the time. Over the years, the team has tied some sound types to behaviors. Google's post gives three examples:

• Signature whistles, unique "names" that mothers and calves can use to find each other again.
• Burst-pulse "squawks", often heard during fights.
• Click "buzzes", often used during courtship or when chasing sharks.

What the model actually does

Fair warning: this gets a little technical. DolphinGemma is an audio-in, audio-out model. Google's SoundStream tokenizer turns dolphin sounds into tokens (small pieces the model can work with), and a model with around 400 million parameters learns which sounds tend to follow which. It's the same trick large language models use to predict the next word in a sentence, applied to whistles and burst pulses. It draws on Google's work on its Gemma open models, and it's small enough to run directly on the Pixel phones the WDP takes into the field.

What's the point? Spotting recurring sound patterns, clusters and reliable sequences in hours of recordings, a job that until now took huge amounts of human effort. In early testing, Google showed spectrograms (pictures of sound over time) of whistles and burst pulses the model generated itself.

CHAT and the Pixel phones

There's a second, parallel project that often gets mixed up with DolphinGemma: CHAT (Cetacean Hearing Augmentation Telemetry), an underwater computer from Georgia Tech. The WDP itself asks on its CHAT page whether it's "a dolphin translator or an interface," and its answer is that the only "translation" is from the computer's whistle to an English word in the diver's ear. CHAT doesn't try to decode the dolphins' natural language. Instead, it plays new, artificial whistles that stand for objects the dolphins like to play with, such as sargassum, seagrass or a scarf. Divers use the whistles with each other while passing the toys around, hoping the curious dolphins copy a whistle to ask for one.

For that to work underwater, the device has to hear the copy over ocean noise, recognize which whistle it was in real time and tell the diver through bone-conducting headphones, so the diver can hand over the right object fast. A Pixel 6 handled the real-time sound analysis. The next version is built around a Pixel 9, which Google said would also act as speaker and microphone and run deep learning models and template matching at the same time; Google slated that research for summer 2025.

This is where the two projects meet: according to Google, DolphinGemma's predictions could help CHAT recognize a possible copy earlier in the sequence, so the researchers can respond faster and the interaction flows better. Google says relying on Pixel phones cuts the need for custom hardware, lowers power use and makes the device smaller and cheaper, which matters a lot when your lab is a boat in the open ocean.

Can you try DolphinGemma?

Not yet. In the April 2025 announcement, Google said it was planning to share DolphinGemma as an open model that summer, so researchers studying other species, like bottlenose or spinner dolphins, could fine-tune it on their own recordings:

That release hasn't happened. Google DeepMind's DolphinGemma page says the model is currently in development and that it will be openly available on release, and there's no DolphinGemma model from Google on Hugging Face or Kaggle, where Google publishes its Gemma models. If you find a "Dolphin Gemma" download, check the publisher: the ones out there are unrelated chatbot models from other authors that just share the name.

Personally, I like that the people doing the work keep the claims modest. The WDP's long-term goal is to understand whether there's structure in dolphin sounds that could point to language, and DolphinGemma is a tool for sifting through the recordings faster. Whether it gets us any closer to a conversation is something we'll only learn after the open release, when other research teams can test it on their own dolphins. 🐬