Google DeepMind Perch Update: Advancing Bioacoustics for Endangered Species

Google DeepMind has released an update to Perch, an AI model designed to accelerate the analysis of bioacoustic data to protect endangered species. This updated model improves state-of-the-art off-the-shelf bird species predictions and expands its capabilities to include mammals, amphibians, and anthropogenic noise, enabling conservationists to process millions of hours of audio data more efficiently.

Enhanced Model Capabilities and Training

The new version of Perch is trained on nearly twice as much data as the previous iteration, utilizing public sources such as Xeno-Canto and iNaturalist. This expanded dataset allows the model to better adapt to diverse environments, including underwater settings like coral reefs, and disentangle complex acoustic scenes.

Beyond simple species identification, the model is versatile enough to address specific ecological questions, such as estimating animal abundance or determining the number of individuals present in a given area.

Agile Modeling and Classifier Development

To support scientists working with scarce training data, Google DeepMind provides tools that implement "agile modeling." This approach combines vector search with active learning, allowing researchers to build high-quality classifiers in under an hour starting from a single sound example.

Using the Perch embedding model, a local expert can use vector search to find similar sounds within a dataset and mark them as relevant or irrelevant, rapidly training a classifier for specific sounds, such as juvenile calls, or for species with limited available recordings.

Field Impact and Success Stories

Since its 2023 launch, the initial version of Perch has been downloaded over 250,000 times and integrated into professional tools like Cornell's BirdNet Analyzer. Real-world applications include:

  • Species Discovery: In collaboration with BirdLife Australia and the Australian Acoustic Observatory, Perch helped discover a new population of the elusive Plains Wanderer.
  • Population Monitoring: Research has demonstrated that Perch can identify individual birds and track abundance, which may reduce the reliance on catch-and-release studies.
  • Conservation Efficiency: Biologists from the LOHE Bioacoustics Lab at the University of Hawaii found honeycreeper sounds nearly 50x faster than previous methods, accelerating the monitoring of species threatened by avian malaria.

Availability

Google DeepMind has released the updated Perch model as an open model available on Kaggle.

Sources