New AI Google model allows people to understand dolphin language

The Google company has once again surprised the world of technology by introducing its new development called DolphinGemma. This large language model is designed to help scientists understand the complex sound signals of dolphins and possibly understand their "conversations".

Neowin writes about it.

The project is being implemented jointly with researchers from the Georgia Institute of Technology and the Wild Dolphin Project (WDP) team, led by Dr Denise Herzing. The organisation has long been studying the natural environment and behaviour of dolphins, especially Atlantic spotted dolphins, recording their social structures, habits, and sound signals. The WDP's many years of experience have made it possible to correlate certain types of sounds with specific animal behaviour:

• Individual whistles — serve as a kind of "names" and help mothers and cubs find each other.
• "Squawks" — often occur during aggressive fights.
• Buzzes — usually heard during courtship or when dolphins are chasing sharks.

How the new model works

DolphinGemma is based on the audio data collected by WDP. Using its proprietary SoundStream tokeniser, it "decomposes" complex dolphin signals into smaller fragments and analyses them in a manner similar to how language models work for human speech. Instead of words or images, the algorithm processes purely audio and predicts what sound should be heard next.

Importantly, DolphinGemma is not a cumbersome project: approximately 400 million parameters allow the model to be run even directly on smartphones. The WDP researchers use a Google Pixel, so they don't need specialised equipment that is difficult to maintain at sea.

Prior to DolphinGemma, the focus of the researchers was on the CHAT (Cetacean Hearing Augmentation Telemetry) system. It tried to establish a simple "common vocabulary" with dolphins through synthetic whistles associated with certain objects, from algae to various toys. Google's new model allows for a deeper exploration of possible patterns in the "language" of marine mammals.

Other researchers will be able to use DolphinGemma this summer to apply the model to analyse acoustic databases and accelerate the study of intelligent dolphins.

As a reminder, Gemini Live has new features. Google's AI can now "see" the user's smartphone screen and provide answers in real time.

We also wrote that AI poses a threat to popular email services. Experts warn that phishing attacks based on large language models are on the horizon.