In a groundbreaking collaboration rooted in extensive oceanic research, Google has partnered with marine biologists and AI researchers to develop a large language model (LLM) named DolphinGemma. This model is designed not for interaction with humans, but to facilitate communication with dolphins. The initiative, in cooperation with the Wild Dolphin Project (WDP) and researchers at Georgia Tech, aims to decode how dolphins communicate among themselves and explore potential ways for humans to communicate with them. This marks a new milestone in a journey that has been ongoing for over 40 years.
Since 1985, the Wild Dolphin Project has conducted the world’s longest continuous underwater study of dolphins, focusing on a group of wild Atlantic spotted dolphins in the Bahamas. The team has non-invasively collected comprehensive underwater audio and video data on individual dolphins, capturing courtship sounds, aggressive calls during conflicts, and distinctive “signature whistles” used as identifiers among the dolphins. This extensive dataset has informed the training of DolphinGemma, an AI model designed to interpret dolphin vocalizations much like how ChatGPT processes human language.
DolphinGemma, built with approximately 400 million parameters on research similar to Google’s Gemini models, operates in an audio-in, audio-out fashion. It listens to dolphin sounds, predicting subsequent vocalizations to learn the structure of their communication.
The integration of artificial intelligence in understanding animal communication is accelerating the pace at which researchers can decipher these complex interactions. AI models are employed to analyze various animal sounds, including dog barks and bird whistles, using pattern recognition and contextual analysis to understand the underlying communication signals. Cetaceans, which include dolphins and whales, are particularly suitable for AI analysis due to their intricate and nuanced communication methods, which are easily recorded and analyzed.
Moreover, DolphinGemma has the capability to generate new dolphin-like sounds with appropriate acoustic patterns, potentially facilitating real-time, simplified exchanges between humans and dolphins. This interaction uses a system referred to as Cetacean Hearing Augmentation Telemetry, or CHAT, which is an underwater computer programmed to emit dolphin sounds associated with familiar objects, such as seagrass or researcher paraphernalia. This system aims to encourage dolphins to mimic these synthetic whistles to request specific items, potentially broadening the system’s vocabulary as more natural dolphin sounds are decoded.
CHAT technology is integrated into modified smartphones, allowing researchers to establish a basic shared vocabulary with dolphins. The system could enable researchers to respond to dolphins’ mimicked whistles by providing associated items, akin to a form of dolphin charades.
Future developments of CHAT intend to incorporate advanced processing and smarter algorithms for quicker responses and clearer interactions between dolphins and humans. However, this progress invites ethical considerations concerning interactions with dolphins in their natural habitats if communication technologies become more advanced.
Looking ahead, Google plans to release DolphinGemma as an open model this summer, facilitating research across different dolphin species, including bottlenose and spinner dolphins. This initiative may significantly enhance scientific understanding of these oceanic mammals. While the realization of a fully reciprocal dialogue with dolphins is not imminent, the project hints at the broader possibilities that AI models could unlock.
