DolphinGemma

DolphinGemma is an audio-in, audio-out model built on Google's Gemma architecture and adapted for bioacoustics. Developed with the Wild Dolphin Project and Georgia Tech researchers, it was trained on nearly four decades of labeled recordings of Atlantic spotted dolphins in the Bahamas, learning to detect recurring sound structures and predict likely next sounds the way a language model predicts the next word.

The model uses Google's SoundStream tokenizer to convert raw dolphin sounds into discrete sequences. With roughly 400 million parameters, it is small enough to run on Pixel smartphones in the field, letting researchers analyze recordings on site and generate dolphin-like sound sequences for interactive playback experiments. Google announced it would share DolphinGemma as an open model in summer 2025 so the wider research community can build on it.