Google introduces an AI program to interpret shark talk that runs on Pixel phones.

When it might be simpler to converse with animals, why try to understand Gen Z jargon?

Google unveiled DolphinGemma today, an open-source AI model that can learn animal communication by analyzing their keystrokes, whistles, and burst pulse. The announcement was made in time for National Dolphin Day.

The model, developed in collaboration with Georgia Tech and the Wild Dolphin Project ( WDP ), teaches how to create dolphin-like sound sequences by learning the vocalizations of dolphins.

The discovery may help identify whether shark communication reaches the same level as language.

DolphinGemma, the longest-running underground dolphin research project in the world, draws on decades of properly analyzed audio and video data that WDP has collected since 1985.

The project used a non-invasive method known as” In Their World, on Their Terms” to study Atlantic Spotted Dolphins in the Bahamas for generations.

Google said in its statement that the unit can aid researchers in discovering hidden buildings and potential meanings within the dolphins ‘ natural communication by identifying repeating audio patterns, clusters, and reliable sequences.

The AI design, which contains around 400 million parameters, can run on Pixel phones used by researchers in the field. It uses Google’s SoundStream tokenizer to process whale sounds and make later sounds predictions in sequence, similar to how sentences are made by human language experts.

DolphinGemma doesn’t work independently. It is integrated with the CETACET ( Cetacean Hearing Augmentation Telemetry ) system, which associates synthetic whistles with particular items that dolphins enjoy, such as scarves, seagrass, and sargassum, potentially fostering a common vocabulary for interaction.

According to Google, “in the future, these habits, combined with synthetic noises created by the experts to refer to things with which the dolphins like to play, may create a shared vocabulary with the dolphins for engaging communication.”

For real-time study of whale sounds, industry analysts now use Pixel 6 devices.

For the summer 2025 study season, the team intends to switch to Pixel 9 tools, which will run both deep learning models and model matching techniques together.

The switch to smartphone systems significantly lessens the need for specialty equipment, which is a significant plus for marine fieldwork. Making relations more liquid, DolphinGemma’s predicted features enable analysts to predict and identify potential mimics earlier in sounds sequences.

What cannot be understood is what needs to be understood

DolphinGemma joins a number of other AI activities aimed at cracking the human language’s script.

An audio language model that can identify animal species, approximate age, and whether noises indicate problems or play was recently developed by the Earth Species Project ( ESP), a nonprofit business. It is not really language, but it can still be used to establish some basic communication.

The concept, which is based on a combination of human speech, environmental sounds, and pet vocalizations, has produced encouraging results even with new species it hasn’t encountered before.

Another significant effort is made in this area by Project CETI.

It is led by researchers from Imperial College London, including Michael Bronstein, who is among the researchers studying sperm whale communication, and examines their intricate patterns of clicks over long distances.

The team is currently using deep neural networks and natural language processing to study 143 click combinations that might make up a type of phonetic alphabet.

Researchers at New York University have taken inspiration from baby development for AI learning, despite the focus of these projects on decoding animal sounds.

Through footage from a head-mounted camera worn by an infant between the ages of 6 months and 2 years old, their Child’s View for Contrastive Learning model ( CVCL ) used to teach language.

The NYU team found that their AI, unlike traditional AI models that require tens of thousands of words to train, could learn from naturalistic data effectively like human infants do.

This summer, Google intends to release an updated version of DolphinGemma, which might extend its use beyond Atlantic spotted dolphins. The vocalizations of various species may still need tweaking the model.

WDP has spent a lot of time establishing connections between dolphin sounds and specific behaviors, including signature whistles used by mothers and calves to reunite, burst-pulse” squawks” during conflicts, and click “buzzes” used during courtship or when chasing sharks.

We’re not just listening anymore, Google declared. We’re beginning to understand the patterns in the sounds, which could lead to a time when the gap between dolphin and human communication might just narrow slightly.

edited by Josh Quittner and Sebastian Sinclair