Google Employs AI to Analyze and Possibly Reply to Dolphin Communication

Dolphins captivate our imagination, and as we discover more about them, our ability to appreciate their astonishing nature increases. Famous for their remarkable intelligence and intricate social interactions, these marine animals express themselves through distinctive signature clicks and whistles, with some findings indicating they might even identify their own “names.” Now, due to advancements in artificial intelligence, we may soon be able to converse with dolphins in their unique language. Google is developing a large language model (LLM) specifically for dolphin sounds, named DolphinGemma.

This innovative project is a collaboration with the Georgia Institute of Technology and the nonprofit Wild Dolphin Project (WDP). Over the past four decades, the WDP team has conducted extensive research on Atlantic spotted dolphins (Stenella frontalis) and amassed extensive acoustic data from the species, which was then utilized to train the LLM. Following this, Georgia Tech and Google instructed the model to generate “dolphin-like” sound patterns. They successfully imitated the rapid clicks dolphins emit during dynamic social interactions, such as during confrontations or close encounters with one another.

Currently, the team’s objective is to explore how AI can assist in completing vocalization sequences—similar to how Google autocompletes your sentences while typing. WDP founder Denise Herzing mentions that without AI, “sifting through all the data and extracting those patterns would take nearly 150 years.”

Employing AI analysis speeds up the process and may help uncover patterns that could be overlooked by researchers. If the AI reliably produces consistent outcomes, it might indicate a recognizable pattern. Researchers can then examine WDP’s video data to observe the dolphins’ behaviors when they emit certain sounds, such as playing, fighting, or signaling threats to their pod.

The team also aims to investigate how dolphins react upon hearing new, AI-generated “words” that resemble dolphin sounds. They intend to use a wearable technology called CHAT (cetacean hearing augmented telemetry), devised by Georgia Tech, which enables researchers to listen and “speak” simultaneously. A device attached to a diver’s chest captures sounds, while another unit secured to their forearm plays them back.

Two divers equipped with the CHAT device will swim alongside a pod of dolphins, utilizing the invented sounds to “request” an object and to exchange it between themselves. If a dolphin replicates the sound for something like seagrass, the researcher will reward it by providing the item. “By showcasing this system among humans, the researchers anticipate that the inquisitive dolphins will learn to mimic the whistles to request these objects,” states Google. “Eventually, as we gain a clearer understanding of more of the dolphins’ natural sounds, those can also be integrated into the system.”

Google aims to release DolphinGemma as an open model this summer. The team expresses, “We aspire to equip researchers globally with the tools to explore their own acoustic datasets, speed up the identification of patterns, and collectively enhance our comprehension of these intelligent marine creatures.” Although we may not achieve a “dialogue” with dolphins in the near future, any advancement in deciphering their communication can increase our empathy for these extraordinary beings and contribute to their conservation.

Discover more about DolphinGemma on the Google blog.