Google Launches DolphinGemma AI to Decode and Generate Dolphin Language

• DolphinGemma is a foundational AI model developed by Google in partnership with The Wild Dolphin Project (WDP) and Georgia Tech. It is trained on over 40 years of underwater audio-visual data from a single pod of Atlantic spotted dolphins (Stenella frontalis) observed in the Bahamas.
• WDP’s long-term, non-invasive research is the world’s longest-running underwater dolphin study, capturing individual life histories, social interactions, and associated sound patterns since 1985.
• The model is based on Google’s Gemma AI family and includes approximately 400 million parameters, making it small and efficient enough to run directly on Pixel 6 and upcoming Pixel 9 smartphones, enabling real-time deployment in the field.
• DolphinGemma uses SoundStream, a proprietary audio tokenizer developed by Google, to convert dolphin vocalizations — including signature whistles, burst-pulse squawks, and click buzzes — into tokens for sequence modeling.
• Like a language model for human text, DolphinGemma functions as an audio-in, audio-out model, predicting the next sound in a sequence to identify recurring structures and potential meanings in dolphin vocalizations.
• The model also generates synthetic dolphin-like sound sequences, aiding researchers in creating an artificial but learnable vocabulary for two-way interaction.
• These capabilities build upon WDP’s earlier innovations like CHAT (Cetacean Hearing Augmentation Telemetry), originally a dolphin-sized keyboard in the 1990s and now a wearable underwater device integrated with Pixel smartphones.
• CHAT’s interactive workflow:
  • Artificial whistles are linked to objects (e.g., scarves, sargassum).
  • Dolphins mimic these sounds in real time, detected by hydrophones.
  • The system identifies the sound and notifies researchers through underwater bone-conduction headphones.
  • Researchers respond by offering the object, reinforcing associative learning.
• The new Pixel 9-powered CHAT (slated for the 2025 summer research season) will allow simultaneous operation of deep learning and template-matching algorithms, improving interaction speed and fluidity.
• While DolphinGemma is not directly connected to the CHAT speaker system, it enhances recognition and prediction of vocal patterns, accelerating how quickly humans can respond to dolphin behaviors.
• The open-access nature of the model (release planned for mid-2025) means it can be fine-tuned for other cetaceans, including bottlenose and spinner dolphins, fostering collaboration across the marine biology research community.
• Researchers caution that this is not a translation system, but rather a powerful analytical tool that could help detect the building blocks of a dolphin communication system — potentially language.
• The project aims to answer fundamental questions in ethology and animal cognition, such as whether dolphins possess symbolic communication, social transmission of knowledge, or cultural behavior.
• Public interest surged following its announcement on National Dolphin Day (April 14, 2025), due to the novelty of using AI to interpret non-human intelligence, sparking conversations around ethical implications, AI applications in wildlife research, and interspecies understanding.

Why This Matters:

DolphinGemma bridges marine biology and artificial intelligence to explore one of science’s most compelling questions: can we understand and one day “speak” with another intelligent species? By modeling complex sound patterns and enabling low-latency field interaction, this initiative expands the frontier of interspecies communication and cognitive science. It also sets a precedent for how compact, energy-efficient AI models can be deployed for high-impact environmental research, moving beyond human-centric applications of generative AI.