DolphinGemma
2025-04-14 · Publisher : Google · Domain : Artificial Intelligence
Google, with Georgia Tech and the Wild Dolphin Project, presents DolphinGemma, a model of about 400 million parameters trained on dolphin sounds and light enough to run on Pixel smartphones.
What changes
- Model based on Gemma research, trained on the Wild Dolphin Project's acoustic database of Atlantic spotted dolphins (monitored since 1985).
- Uses the SoundStream audio tokenizer to represent sounds and predict the continuation of a sound sequence.
- About 400 million parameters, sized to run directly on Pixel phones used in the field.
- Integrated into the CHAT (Cetacean Hearing Augmentation Telemetry) system to anticipate dolphin sounds; research planned with a Pixel 9 in summer 2025.
### Specifications - Type: research audio model, running on a phone. - Release as an open model announced for “this summer” (2025); actual availability has not been verified. ### Pricing Not applicable (research). ### Benchmarks No numerical score published. ### Key takeaways An example of a frugal specialized model for field scientific research, with no direct link to enterprise uses.
Links
Tags : LLM, Google, Gemma, Audio, Recherche