.svg){kind=spacer}
Details
- Google DeepMind released Perch 2.0 in August 2025, a bioacoustics foundation model trained primarily on terrestrial animals including birds, mammals, amphibians, and insects across 14,597 species.
- Despite containing almost no marine mammal audio in its training data, Perch 2.0 demonstrates strong performance on underwater acoustic tasks including whale vocalization classification and killer whale subpopulation identification.
- The model uses transfer learning to adapt to marine environments, leveraging pre-trained sound understanding capabilities while learning new parameters for underwater species identification through few-shot learning.
- Perch 2.0 outperforms or matches competing bioacoustics models including AVES-bird, AVES-bio, BirdNet v2.3, and Perch 1.0 across multiple marine validation datasets: NOAA PIPAN, ReefSet, and DCLDE 2026.
- Google Research and Google DeepMind created an end-to-end demo with the NOAA Passive Acoustic Archive dataset on Google Cloud to support marine bioacoustics research and enable rapid classifier development from labeled examples.
Impact
Perch 2.0's successful transfer from terrestrial to marine bioacoustics represents a significant advance in scalable species monitoring for ocean conservation. The model's ability to generalize across acoustic domains—learning detailed features from bird classification that translate to whale vocalizations—demonstrates that foundation model scaling and diverse training data can overcome domain-specific challenges without explicit retraining on target audio. This directly addresses a critical bottleneck in marine conservation: the high cost and complexity of underwater data collection and species identification at scale. By enabling rapid classifier development from minimal labeled examples, Perch 2.0 reduces the time and resources needed to deploy species identification systems across diverse underwater environments. The research positions transfer learning as a viable strategy for bioacoustics monitoring where labeled data is sparse, lowering barriers to adoption for conservation teams with limited resources. Integration with Google Cloud infrastructure further democratizes access to these tools for the broader cetacean acoustics and marine research community.
