Google has introduced an AI model named DolphinGemma, aimed at understanding dolphin communication and potentially facilitating interactions between species. The diverse clicks, whistles, and pulses that characterize dolphin vocalizations have intrigued scientists for years, prompting a quest to decode their intricate patterns.
In collaboration with engineers from the Georgia Institute of Technology and utilizing the extensive research conducted by the Wild Dolphin Project (WDP), DolphinGemma was announced around National Dolphin Day. This foundational AI model serves as a pivotal tool in advancing our comprehension of cetacean communication.
It has been trained to recognize and replicate the structure of dolphin sounds, even generating novel audio sequences reminiscent of dolphin calls. The Wild Dolphin Project, established in 1985, has conducted the longest continuous underwater study of dolphins, focusing on context-specific sounds, including signature whistles that act as identifiers and burst-pulse squawks commonly linked to aggression.
Their main objective is to uncover the underlying framework and meanings within these natural sound sequences, in a quest to identify grammatical patterns resembling language. DolphinGemma employs specialized audio technologies to handle the vast complexity of dolphin sounds, utilizing a SoundStream tokeniser to represent these sounds effectively.
It learns to identify recurring patterns in the sounds while predicting likely subsequent sequences, similar to how human language models function. With around 400 million parameters, the model is designed for efficient operation, even on devices like Google Pixel smartphones used by WDP during field studies.
Meanwhile, the CHAT (Cetacean Hearing Augmentation Telemetry) system, also developed by WDP in collaboration with Georgia Tech, aims to facilitate two-way interaction with dolphins. By teaching dolphins a simpler vocabulary using synthetic whistles associated with objects, researchers hope to encourage mimicry and communication.
Both the analysis of dolphin sounds and the interactive CHAT system rely on advanced mobile technology provided by Google Pixel phones. These devices enable real-time processing of audio data in challenging ocean environments, enhancing researchers’ ability to engage with dolphins effectively.
As Google plans to release DolphinGemma as an open model, it is expected to benefit researchers studying various cetacean species and accelerate global efforts to understand these intelligent marine mammals, inching us closer to bridging the communication divide between humans and dolphins.