Google’s DolphinGemma AI Translates Dolphin Sounds, Paving the Way for Interspecies Communication

Google has introduced a new artificial intelligence system known as DolphinGemma, designed to interpret dolphin vocal patterns with the long-term aim of enabling communication between species.

Scientists have been captivated for years by the detailed clicks, whistles, and pulses that resonate among dolphins beneath the waves. Their aspiration has been to understand and decode the structure behind these intricate sound signals.

Working in collaboration with engineers from the Georgia Institute of Technology and drawing on field observations from the Wild Dolphin Project (WDP), Google has presented DolphinGemma as part of efforts to meet this objective. Announced in line with National Dolphin Day, the pioneering AI model serves as an innovative tool in the study of cetacean vocalizations. Built to capture the underlying framework of dolphin sounds, the system is also capable of producing original audio sequences that mimic natural dolphin calls.

For more than three decades, the Wild Dolphin Project—active since 1985—has conducted the longest continuous underwater study of dolphins. Their work has focused on context-related sounds, including:

• Unique “whistles” that serve as identifiers much like personal names during encounters such as maternal reunions.
• Rapid burst-pulse “squawks” linked to situations of conflict or forceful interactions.
• Clicking “buzzes” recorded during mating displays or pursuits involving sharks.

WDP’s primary objective is to reveal the basic structure and possible significance of these natural sound series by examining grammatical patterns that might represent a language system. This continuous, detailed research has provided a solid base along with carefully labeled data that has been crucial in training advanced AI systems such as DolphinGemma.

Studying the vast volume and complexity of dolphin communications poses a significant challenge that artificial intelligence is well-equipped to tackle. DolphinGemma employs advanced audio processing technologies to address these challenges. The model incorporates the SoundStream tokenizer to capture dolphin sounds efficiently and channels this information through an architecture developed for handling elaborate sequences.

Drawing cues from the technology behind Google’s Gemma family of lightweight, open models—which share components with the Gemini series—DolphinGemma is designed as a system that both accepts and produces audio signals. Trained on sequences from WDP’s extensive natural sound database, the AI recognizes repeating patterns and structural features. It can also forecast the subsequent sounds in a series in a manner similar to human language models that predict the next term.

Comprising roughly 400 million parameters, DolphinGemma has been optimized for smooth and efficient operation, even when running on the Google Pixel smartphones that WDP uses for collecting acoustic data during fieldwork. As the Wild Dolphin Project begins deploying the model this season, expectations are high that research will progress at a much faster pace. The system automatically detects consistent sequences and patterns that once required extensive manual effort from scientists, thereby revealing previously unnoticed structures and meanings in dolphin communication.

DolphinGemma concentrates on interpreting natural vocalizations, and a parallel initiative is examining active two-way communication. The CHAT (Cetacean Hearing Augmentation Telemetry) system, developed in partnership between WDP and Georgia Tech, seeks to establish a straightforward, shared vocabulary rather than attempting a literal translation of dolphin sounds. This approach involves linking specially created synthetic whistles—distinct from natural calls—to specific items that appeal to dolphins, such as scarves or seaweed. In demonstrations, researchers connect a particular synthetic whistle with an object, anticipating that the innate curiosity of dolphins will prompt them to replicate the sound when they wish to request that item. As understanding of natural dolphin sounds expands through work with systems like DolphinGemma, it is possible that certain sound patterns may eventually be incorporated into the CHAT framework.

Both the analysis of natural vocalizations and the reciprocal CHAT system rely on sophisticated mobile technology. Google Pixel phones act as central processors for high-quality audio data, operating in real time amid the challenging conditions of the ocean. For example, the CHAT system depends on these devices to:

• Recognize a potential mimic amid ambient sounds.
• Pinpoint the specific synthetic whistle produced.
• Notify the researcher through underwater bone-conducting headphones about the dolphin’s expressed request.

This arrangement enables scientists to respond quickly with the appropriate item, reinforcing the association between the whistle and the object. An earlier version of the system operated using a Pixel 6, but plans are underway for an upgraded CHAT system projected for summer 2025 that will use a Pixel 9. The new setup will combine speaker and microphone capabilities and run deep learning models alongside template matching algorithms concurrently, resulting in improved performance.

Using smartphones such as the Pixel cuts down on the reliance on bulky, costly custom-built equipment. This approach simplifies system upkeep, reduces energy consumption, and results in a more compact design. The prediction capabilities of DolphinGemma integrated with the CHAT system may also allow quicker identification of mimicry, leading to smoother and more effective exchanges.

Recognizing that breakthroughs often arise from cooperative efforts, Google plans to release DolphinGemma as an open model later this summer. Although the system was trained using data from Atlantic spotted dolphins, its design may prove beneficial to scientists studying the vocal patterns of other cetacean species, even if modest adjustments are needed for varying acoustic profiles. The goal is to provide researchers worldwide with advanced tools for assessing their own audio collections, thereby accelerating the collective endeavor to better understand these intelligent marine creatures. The shift from passive recording to active interpretation of sound patterns brings the possibility of narrowing the communication gap between humans and dolphins ever closer.