DolphinGemma
DolphinGemma is a ~400M-parameter Gemma-derived audio model trained on Wild Dolphin Project recordings to model dolphin vocalizations, and it runs on Pixel phones in the field.
Uses the SoundStream tokenizer to represent dolphin sounds and a sequence model to analyze them and generate dolphin-like sound sequences, with Georgia Tech and the Wild Dolphin Project. Google planned an open release for summer 2025.
- Date
- Monday, 14 April 2025
- Lab
- Kind
- model
- Access
- research preview
Niche science application of the Gemma recipe; no claim of "translation" is made.
Sources
This record was checked against its sources on 6 October 2026. How we check