Melissa Donovan
Google, in collaboration with researchers from Georgia Tech and the Wild Dolphin Project (WDP), has introduced DolphinGemma, a foundational AI model designed to analyze and generate dolphin-like vocal sequences. The announcement made on Tuesday via Google’s blog coincides with National Dolphin Day and highlights ongoing efforts to study non-human animal communication using artificial intelligence.
DolphinGemma has been trained to identify structural patterns in the vocalizations of Atlantic spotted dolphins (Stenella frontalis). The model is built on data collected through long-term underwater field studies and integrates audio-processing technologies developed at Google.
Decades of Field Research on Dolphin Communication
The Wild Dolphin Project, established in 1985, conducts longitudinal research on a specific community of wild Atlantic spotted dolphins in the Bahamas. Its methodology emphasizes non-invasive, in-situ observation, which has resulted in a substantial dataset of audio-visual recordings. These recordings are paired with individual dolphin identifications, life histories, and documented behaviors.
Researchers at WDP focus on analyzing natural dolphin vocalizations in context. Common types of sounds include:
Signature whistles, which are often used by individuals such as mothers and calves to reunite
Burst-pulse sounds, typically associated with social conflict or aggression
Click trains and buzzes, frequently observed during courtship or in response to predators
By linking these vocalizations with specific behavioral observations, researchers aim to detect patterns that could indicate structure within the dolphins’ communication system.
Overview of the DolphinGemma Model
DolphinGemma utilizes a ~400 million parameter model and incorporates Google’s SoundStream audio tokenizer. This model architecture is designed for sequence prediction and pattern recognition, similar to those used in natural language processing. It processes dolphin vocalizations to identify structural elements and predict subsequent sound patterns in a sequence.
The model is trained on WDP’s extensive archive of labeled acoustic data. It can run on mobile hardware, such as Google Pixel smartphones, which are currently used by WDP in field environments.
Applications in Field Research and the CHAT System
In addition to analyzing natural vocalizations, DolphinGemma is being tested alongside the Cetacean Hearing Augmentation Telemetry (CHAT) system, a tool developed by Georgia Tech and WDP. CHAT is designed to explore two-way interaction by associating synthetic whistles with specific objects dolphins interact with, such as seagrass or scarves.
The system aims to determine whether dolphins will mimic these synthetic signals to request items. CHAT functions by:
Capturing and identifying dolphin vocalizations in real time
Transmitting data to researchers via bone-conduction headphones
Supporting rapid object delivery in response to mimicry
Earlier versions of the system were powered by the Pixel 6. The current iteration uses the Pixel 9, enabling on-device processing for real-time AI model execution, acoustic analysis, and signal matching. This setup reduces the need for custom equipment, enhances portability, and supports real-time responsiveness in open-ocean research.
Model Access and Broader Research Implications
Google plans to release DolphinGemma as an open-source model in mid-2025. Although it was trained on Atlantic spotted dolphin vocalizations, the architecture allows for adaptation to other species through fine-tuning. Researchers studying other cetaceans, such as bottlenose or spinner dolphins, may apply the model to their datasets.
The goal of this release is to support broader research efforts in marine bioacoustics by providing a tool for pattern recognition and analysis of animal communication.
DolphinGemma represents a collaborative effort between technology developers and field researchers to enhance the study of dolphin vocal behavior. The combination of AI modeling and long-term biological observation aims to provide new methods for analyzing complex sound patterns in marine species.
More information about the Wild Dolphin Project can be found at www.wilddolphinproject.org.
Comments