Google DeepMind's Perch 2.0 bioacoustics foundation model, trained primarily on bird vocalizations, demonstrates surprising effectiveness in classifying underwater whale sounds despite never hearing marine audio during training. The model achieves killer performance on marine acoustic tasks, transferring knowledge from terrestrial bioacoustics to underwater ecosystems. This breakthrough addresses the challenge of monitoring mysterious ocean sounds like the recently identified 'biotwang' attributed to Bryde's whales by NOAA.

Google's approach to AI for bioacoustics is evolving to enable faster connections from new discoveries to scientific insights at scale. The company has a history of collaborating with scientists on whale monitoring, including humpback detection models and multi-species whale classifiers released in 2024. Perch 2.0's success stems from its ability to learn detailed acoustic features from similar bird calls, which transfers effectively to distinguishing whale vocalizations and subpopulations. The model outperforms other pre-trained systems like AVES-bird and AVES-bio on most underwater classification tasks.

The research team evaluated Perch 2.0 using a few-shot linear probe approach on three marine datasets: NOAA PIPAN recordings, ReefSet coral reef sounds, and DCLDE killer whale data. Results show consistent top or second-best performance across different sample sizes and classification challenges. Google is sharing an end-to-end tutorial in Google Colab demonstrating how researchers can use Perch 2.0 to create custom whale vocalization classifiers through the NOAA NCEI Passive Acoustic Data Archive on Google Cloud.