A new large language–model artificial intelligence could soon help people understand dolphins — and help dolphins understand people.
Developers at Google and researchers at several U.S. institutions plan to debut the system in the coming months. The team will conduct tests to see if the DolphinGemma model and its companion, Cetacean Hearing Augmentation Telemetry (CHAT), can translate and imitate some of the complex vocalizations of dolphins. If successful, the work could mark the culmination of more than four decades of research, Popular Science reported.
What is known about the new development?
Dolphins are some of the smartest and most communicative animals on Earth. The social interactions of these mammals are so complex that researchers with the Wild Dolphin Project (WDP) have spent the last 40 years trying to decipher them. Over that time, scientists have amassed a large collection of underwater audio and video recordings. Those recordings, in particular, document the behavior and vocal signals of an Atlantic spotted dolphin community living near the Bahamas. Researchers have correlated the dolphins’ sounds with behaviors such as courtship and disputes.
Scientists have long suspected humans might be able to communicate with cetaceans, but they lacked technology advanced enough to analyze and imitate dolphins’ underwater whistles and other sounds.
With the emergence of large language models (LLMs), researchers wondered whether the same principles could be applied to dolphin communications. To explore that idea, WDP scientists teamed up with Google and the Georgia Institute of Technology. The researchers provided engineers with a vast dataset of audio — whistles, clicks, and pulses from dolphins — to train the LLM.
The result is the AI model DolphinGemma, built using the same technology principles as Google’s Gemini systems. DolphinGemma was developed with roughly 400 million parameters to function much like language models such as ChatGPT, but for dolphins.
DolphinGemma first receives and interprets audio input, then predicts likely subsequent sounds. The model then interacts with a system called Cetacean Hearing Augmentation Telemetry (CHAT), installed on modified Google Pixel smartphones. CHAT isn’t designed to produce complete translations of dolphin ‘language,’ but it can help humans convey a simplified, generalized vocabulary.
Experts hope that, in time, dolphins might even learn to ask for toys when they want to play.
There’s still a lot to do before humans and dolphins can bridge the gap in interspecies communication. But with this new language model, those conversations have taken a step closer.
