DolphinGemma
Lab & Provider: Google DeepMind
Overview & Technical Scope
Compact bioacoustic sequence model (~400M parameters) developed by Google DeepMind in collaboration with the Wild Dolphin Project and Georgia Tech. Applies the SoundStream neural audio codec tokenizer with Gemma-derived sequence modeling to predict acoustic token transitions in recordings of wild Atlantic spotted dolphins.
Gemma Architecture (~400M parameters) + SoundStream Neural Audio Codec
Evaluated on acoustic token transition perplexity and sequence continuation using longitudinal hydrophone recordings of wild Atlantic spotted dolphins. Demonstrates next-token predictive pattern detection across discrete acoustic token vocabularies without semantic translation.
Computational acoustic sequence prediction model strictly limited to predicting acoustic token transitions. Does not perform, claim, or support semantic translation of dolphin vocalizations into human language.
Note: Computational acoustic pattern modeling and next-token sequence predictions must not be conflated with semantic translation of non-human vocalizations into human language.