How AI is helping advance the science of bioacoustics to save endangered species

Source: DeepMind Date: 2025-10-24 URL: https://deepmind.google/blog/how-ai-is-helping-advance-the-science-of-bioacoustics-to-save-endangered-species/

Summary

Google DeepMind updated Perch, its bioacoustics ML model, with expanded training data covering mammals, amphibians, and anthropogenic noise — nearly 2x the original dataset. Results: 250,000+ downloads since 2023, integrated into Cornell’s BirdNet Analyzer, discovered a new Plains Wanderer population in Australia, and enabled University of Hawaii LOHE Lab to find honeycreeper sounds 50x faster than manual methods. New “agile modeling” capability builds high-quality species classifiers in under an hour from minimal training examples via vector search and active learning.

Implications

50x faster species identification is the conservation field productivity number that matters. LOHE Lab’s honeycreeper work represents a typical conservation bottleneck: researchers have thousands of hours of passive acoustic monitoring data and insufficient analyst capacity to process it. 50x faster means the same team can survey 50x more habitats in the same time, or process years of backlogged recordings in weeks. That’s a real resource multiplier, not an incremental improvement.

Agile modeling — high-quality classifiers in under an hour from minimal examples — is the long-tail species unlock. The original Perch was powerful for well-represented species. The barrier for rare species was labeled training data. Active learning from minimal examples means a conservation team with 10 verified recordings can build a classifier and screen millions of recordings for that call. That’s the capability that matters for species with no prior acoustic research infrastructure.

250,000 downloads and BirdNet integration signals that Perch is field infrastructure, not research software. BirdNet is the dominant citizen science bird identification app. Perch integration means its methodology is embedded in a tool used by millions of birders globally — generating labeled training data at scale while serving users. That’s a data flywheel no academic lab can replicate.

Watch:

• Agile modeling performance on genuinely rare species with fewer than 50 labeled examples — the one-hour claim needs adversarial testing on minimal-data scenarios
• Expansion beyond birds and mammals to insects and marine invertebrates, where acoustic monitoring is far less developed
• Whether the Plains Wanderer population discovery triggers formal habitat protection — the policy downstream of the AI finding is the conservation outcome that matters