Google's DolphinGemma AI Decodes Dolphin Language: Open Source Model to Aid Cetacean Communication Research

Google introduced DolphinGemma, an AI model designed to decode dolphin communication, developed with the Georgia Institute of Technology and the Wild Dolphin Project (WDP). DolphinGemma uses the technology behind Google's Gemma models, analyzing dolphin vocal sequences and predicting subsequent sounds, akin to autocomplete for dolphin clicks, whistles, and pulses. Central to the system is the SoundStream tokenizer, which divides dolphin sounds into acoustic units processed by a custom architecture to identify patterns. The model was trained using data collected by the WDP since 1985, studying Atlantic spotted dolphins (Stenella frontalis) in the Bahamas. This data revealed correlations between specific sounds and behaviors, such as signature whistles used by mothers and calves, intermittent 'squawks' associated with tension, and 'buzz' clicks observed during courtship. DolphinGemma recognizes and reproduces plausible vocal sequences. With 400 million parameters, it's optimized to run on Google Pixel devices used by researchers in the field. The WDP is exploring communication with dolphins via the CHAT (Cetacean Hearing Augmentation Telemetry) system, a underwater computer designed to introduce a shared vocabulary based on synthetic whistles linked to objects like sargassum or scarves. The goal is to establish symbolic interaction if dolphins imitate artificial whistles to request objects. CHAT, initially running on a Pixel 6, will be upgraded to Pixel 9, enabling simultaneous deep learning models and pattern matching algorithms. DolphinGemma can be used with CHAT to predict imitation attempts, enhancing interaction effectiveness. Google plans to release DolphinGemma as an open-source model in summer 2025, adaptable to other species like bottlenose dolphins. The aim is to accelerate scientific discovery, enabling researchers to analyze acoustic datasets and contribute to understanding marine mammals.