Hyderabad:

"This approach in the quest for interspecies communication pushes the boundaries of AI and our potential connection with the marine world," Google said in a blog post. Google has introduced a new large language model, DolphinGemma, aimed at helping scientists decode dolphin language. In collaboration with researchers at Georgia Tech and the field research of the Wild Dolphin Project (WDP), Google announced progress on the foundational AI model, trained to learn the structure of dolphin vocalisations and generate novel dolphin-like sound sequences.

DolphinGemma makes use of WDP's vast, labelled dataset to analyse dolphins' natural and complex communication. Since 1985, WDP has conducted the world's longest-running underwater dolphin research project, which has accumulated a rich, unique dataset comprised of decades of underwater video and audio paired with individual dolphin identities, life histories, and observed behaviours.

For decades, WDP has correlated sound types with behavioural contexts, such as signature whistles that can be used by mothers and calves to reunite, burst-pulse "squawks" during fights, and click "buzzes" during courtship or chasing sharks.

In addition to analysing natural communication, WDP is also exploring a way for humans to communicate with dolphins. They developed the CHAT (Cetacean Hearing Augmentation Telemetry) system, in partnership with the Georgia Institute of Technology, to establish a simpler, shared vocabulary.

DolphinGemma uses a special tool called SoundStream to turn these sounds into data, which is then processed by a powerful AI system designed for complex patterns. Despite being advanced, the model is small enough to run directly on Pixel phones used in the field by WDP.

CHAT produces unique whistles that aren’t like natural dolphin sounds, but links them to objects dolphins enjoy, like seaweed or scarves. Researchers first show how it works between humans, hoping that dolphins—who are naturally curious—will learn to copy the whistles to ask for these items. Over time, as scientists decode more dolphin sounds, they can add them to the system too.

To enable two-way interaction, the CHAT system first needs to:

Hear the mimic accurately amid ocean noise.

Identify which whistle was mimicked in real-time.

Inform the researcher (via bone-conducting headphones that work underwater) which object the dolphin "requested."

Enable the researcher to respond quickly by offering the correct object, reinforcing the connection.

A Google Pixel 6 was used to analyse dolphin sounds instantly with high accuracy. The upcoming phase of research, planned for summer 2025, will use a Google Pixel 9, integrating speaker/microphone functions and using the phone's processing to run both deep learning models and a template matching algorithm simultaneously.

Read More :

IIT-M Launches PG Diploma Courses For Upskilling In Emerging Tech Domains