Photo credit: www.popsci.com
Breakthrough in Dolphin Communication: New AI Model Set to Bridge the Gap
A revolutionary large language model AI system is on the verge of enabling human-dolphin conversations, with its debut slated for the upcoming months. Researchers are preparing to evaluate whether DolphinGemma and its complementary Cetacean Hearing Augmentation Telemetry (CHAT) system can adequately translate and mimic some of the intricate vocalizations of dolphins. If the trials are successful, this endeavor could mark a significant milestone after more than 40 years of research, documentation, and conservation initiatives.
Dolphins are among the most intelligent and communicative creatures on the planet. Their social structures are so multifaceted that researchers at the Wild Dolphin Project (WDP) have devoted four decades to understanding their communication. Through extensive underwater recordings and observations, the WDP has built a substantial archive of interactions involving a particular group of Atlantic spotted dolphins in the Bahamas. This long-term study has allowed researchers to associate specific sounds with behaviors such as mating rituals, unique vocal identifiers, and social disputes among dolphins.
While the hypothesis of interspecies communication has long intrigued experts, the lack of sophisticated technology capable of deciphering the dolphins’ complex auditory signals—comprising whistles, clicks, and burst pulses—has hindered earlier attempts. However, the advent of large language models (LLMs) has prompted researchers to explore the possibility of applying similar technologies to dolphin communication. To advance this idea, the WDP recently collaborated with Google and the Georgia Institute of Technology, providing engineers with a vast, labeled dataset of dolphin sounds for the purpose of training the LLM.
The outcome of this collaboration is DolphinGemma, an AI model developed using the same foundational technology as Google’s Gemini systems. It operates on approximately 400 million parameters, functioning in a manner similar to predictive LLMs like ChatGPT, but tailored specifically for dolphin communication.
(Left) Early DolphinGemma spectrograms of clicks and burst pulses. (Right) Spectrogram illustrating a mother dolphin’s unique call for her calf. Credit: Google
DolphinGemma processes and interprets audio inputs, subsequently forecasting likely subsequent sounds for reproduction. This system is paired with CHAT, which is installed on modified Google Pixel smartphones. Rather than providing a complete translation of dolphin language, CHAT aims to assist humans in establishing a simplified, shared vocabulary with dolphins. The strategy includes teaching the dolphins synthetic sounds associated with their favorite objects, such as seagrass, sargassum, and even researchers’ scarves. Over time, the goal is for dolphins to learn how to request specific items when they wish to engage in play.
Although significant challenges remain before achieving effective communication between humans and dolphins, the innovative application of LLMs signals progress towards understanding these aquatic mammals better.
Source
www.popsci.com