Google built an on-device AI to decode dolphin chatter

Dolphins have spent millions of years talking to each other using whistles and clicks. Humans, meanwhile, have spent the last four decades holding hydrophones in the ocean, trying to figure out what any of it means.

On National Dolphin Day, Google unveiled DolphinGemma. Developed in collaboration with Georgia Tech and the Wild Dolphin Project (WDP), it is a foundational AI model designed to decode the structure of dolphin vocalizations and generate novel, dolphin-like audio sequences.

Predicting dolphin clicks like LLM tokens

Since 1985, the Wild Dolphin Project has studied a specific community of wild Atlantic spotted dolphins in the Bahamas. Over forty years of non-invasive underwater research, they gathered decades of synchronized audio and video matched to individual dolphin identities and social behaviors.

To make sense of that massive acoustic library, Google adapted the tech behind its lightweight Gemma models. DolphinGemma sits at around 400 million parameters and uses Google’s SoundStream audio tokenizer to turn marine acoustics into tokens.

It works as an audio-in, audio-out system. Just as an LLM predicts the next word in a sentence, DolphinGemma analyzes natural dolphin vocalization sequences to forecast what sound is likely to come next. By mapping these sound clusters, scientists can surface recurring grammatical patterns that were previously impossible to spot by hand.

Running the underwater translator on a Pixel 9

Deciphering natural chatter is one half of the project; the other is two-way interaction. Georgia Tech and WDP previously built the Cetacean Hearing Augmentation Telemetry (CHAT) system, a wearable underwater computer designed to create a shared vocabulary between humans and dolphins.

Instead of trying to speak full dolphin, CHAT associates synthetic, artificial whistles with objects dolphins like to play with, like sargassum, sea grass, or scarves. When dolphins hear humans use the whistle to trade an object, they naturally mimic the sound to ask for the toy themselves.

Underwater researchers wear bone-conducting headphones. The heavy lifting runs entirely on a commercial Google Pixel smartphone tucked inside the CHAT housing. The upcoming summer 2025 field season upgrades to a Pixel 9, which runs both template matching and DolphinGemma on-device.

Because DolphinGemma can predict the likely continuation of a sound sequence, CHAT can detect when a dolphin is mimicking a whistle before the animal even finishes the vocalization. That shaves critical seconds off human response time, reinforcing the reward loop before the dolphin loses interest.

Google plans to open-source DolphinGemma in summer 2025 so external researchers can fine-tune it for other cetacean species like bottlenose or spinner dolphins.

Why it matters

For marine biologists and bioacoustics researchers, running a 400M audio model directly on an off-the-shelf smartphone is a major upgrade. Fieldwork on open water is brutal on electronics; replacing bulky, power-hungry custom rigs with a waterproofed Pixel significantly cuts equipment cost and power draw. It turns edge computing into a practical field asset rather than a lab novelty.

In our view, the cleverest part of this initiative is how it sidesteps the trap of pretending to "solve" animal language overnight. Building a synthetic, two-way bridge around shared toys is far more grounded than training an unsupervised black box and claiming it understands cetacean philosophy. It gives researchers an empirical feedback loop to test whether meaning is actually being transferred.

That said, the hurdles ahead are steep. Marine environments are notoriously noisy, and running real-time acoustic recognition against ambient waves and engine hum pushes mobile silicon to its limits. More importantly, getting a dolphin to parrot an artificial tone for a scarf proves operant conditioning, not full interspecies communication. Bridging the gap from functional mimicry to genuine semantic understanding remains a massive scientific leap.