The same basic prediction task behind a chatbot can be applied to dolphin sounds. Google introduced DolphinGemma in April 2025 to learn patterns in recordings and generate sequences resembling dolphin vocalizations.
That does not mean the model translates dolphin conversations. Predicting which sound follows another establishes an acoustic pattern. Connecting that pattern to meaning requires evidence from the animals' behavior.
The project combines Google's audio modeling with Georgia Tech and the Wild Dolphin Project. The field researchers have studied Atlantic spotted dolphins in the Bahamas since 1985, linking recordings with observations of identifiable animals.
That context is essential. An audio file alone does not explain what was happening when a sound occurred. Associated video and behavioral records give researchers something against which to test a proposed interpretation.
DolphinGemma processes sound as both input and output. Google describes a model with about 400 million parameters, the adjustable values learned during training. Its size was chosen to run on Pixel phones used for field research.
The announcement also discusses CHAT, a separate underwater interaction system involving artificial whistles and objects. That controlled interaction work should not be confused with decoding the dolphins' natural communication.
The useful result is a tool for finding and testing patterns in a large archive. The 2025 announcement offers no dolphin-to-English dictionary. The sounds still need context.
