Google AI May Be Close to “Speaking Dolphin” with New DolphinGemma Model

This is the type of artificial intelligence application that really catches attention. Researchers have developed DolphinGemma, the first large language model (LLM) designed to understand dolphin communication. This breakthrough could significantly speed up our ability to translate what these amazing creatures are saying, something that until now has taken decades of manual research.

The Vision of Speaking Dolphin

According to Dr. Denise Herzing, founder of The Wild Dolphin Project (WDP), “The goal is to one day be able to speak dolphin.” Her organization exclusively studies a group of Atlantic spotted dolphins living off the coast of the Bahamas. She has been recording and organizing their vocalizations for over 40 years and has collaborated with Dr. Thad Starner, a researcher at Google DeepMind.

Together, they have trained an AI with a vast library of dolphin sounds. The model can be continuously expanded with new data and refined to more accurately represent what these sounds might mean. “Feeding dolphin sounds into an AI like DolphinGemma will give us a better understanding of patterns and subtleties that humans might miss,” explained Herzing.

Dolphin Communication Methods

These marine mammals communicate through a variety of sounds, including whistles (some of which are names), echolocation clicks for hunting, and burst-pulse sounds for social interactions.

Since the 1980s, scientists have used underwater microphones (hydrophones) to record these sounds, analyzing them with spectrograms. Then, they play the sounds back to observe the dolphins’ reactions. This manual process is very labor-intensive, making the use of AI a game-changer.

The team’s new language model, built on the same technology behind Google’s Gemini models and its popular chatbot, represents dolphin sounds as “tokens,” or interpretive sound units.The team trained it using the WDP’s acoustic database, and it can identify patterns and structures within the dolphins’ vocalizations, even predicting what the next sound might be — just like Google’s search suggestions or predictive text while typing emails.

Taking it a step further: Trying to communicate with dolphins

But the researchers don’t just want to observe dolphins talking to each other — they aim to communicate directly with them. Since the 1990s, Herzing’s team has been experimenting with rudimentary ways of “talking” to dolphins. The first attempt involved a large, dolphin-sized keyboard mounted on the side of a boat, with each key corresponding to an artificial whistle linked to a toy. The idea was for the dolphins to point at the keys to request toys.

Over time, this evolved into an underwater portable keyboard. While it didn’t lead to true conversations, it showed that dolphins were attentive and willing to engage in learning. In 2010, the team began collaborating with the Georgia Institute of Technology, where Starner is a professor, to develop more sophisticated communication technologies.

This led to the creation of the CHAT (Cetacean Hearing and Telemetry) system — a wearable device with hydrophones, speakers, and an interface for divers. It detected specific dolphin whistles and associated them with objects or concepts. The researchers could then trigger artificial whistles to “speak” to the dolphins.

Integration of Smartphones for Improved Communication

More recently, the team has incorporated smartphones (Google Pixel) into the system, making it smaller, more efficient, and easier to use. The idea is to demonstrate an action, like handing over a toy, and play a sound so the dolphin can associate it with the object. If the dolphin mimics the sound, the system informs the researcher, who then responds by providing the requested object, reinforcing the connection.

With AI that predicts responses and identifies patterns more quickly, scientists can interact more naturally with the dolphins. The new version of CHAT uses less power and can be maintained more easily than its predecessor.

The WDP team will deploy DolphinGemma in the upcoming field season. Google has stated it will make the model publicly available by mid-year, allowing researchers worldwide to use it. The hope is that with some adjustments, the model can also be adapted to other cetacean species, like bottlenose or spinner dolphins.

If dolphins have a language, then they probably also have culture, noted Starner. Understanding what they talk about, what their priorities are, and how their societies operate could give us a whole new perspective on communication and intelligence in the animal kingdom.

Read the original article on: New Atlas

Read more: Dolphins in Gulf Found Contaminated with Fentanyl