How AI trained on birds is surfacing underwater mysteries
Google Research
Google DeepMind released Perch 2.0, a bioacoustics foundation model trained primarily on bird and terrestrial animal sounds, which unexpectedly performed well at classifying marine mammal vocalizations despite containing no underwater audio in its training data. The model achieved top or second-best performance across three marine datasets (NOAA PIPAN, ReefSet, and DCLDE) when evaluated with few-shot learning using 4 to 32 labeled examples per class. The findings enable researchers to rapidly create custom whale vocalization classifiers using transfer learning, reducing computational requirements and supporting faster analysis of newly discovered marine sounds through a publicly available end-to-end tutorial on Google Colab.
Why it matters
Climate & Sustainability
