August 7, 2025, marks a significant advancement in the scientific field as the Perch team releases a new model that assists conservationists in faster analysis of bioacoustic data to protect endangered species, from Hawaiian honeycreepers to coral reefs. Scientists utilize microphones (or underwater hydrophones) to collect vast amounts of audio dense with vocalizations from various species such as birds, frogs, insects, whales, and fish, revealing insights about the species present and the health of the ecosystem. However, processing such extensive data remains a monumental challenge.
Today, we unveil an update to Perch, our AI model designed to help conservationists analyze bioacoustic data. This new model exhibits superior off-the-shelf predictions for bird species compared to its predecessor, particularly adapting better to new environments like coral reefs. It has been trained on a broader range of animal data, including mammals, amphibians, and anthropogenic noise, nearly doubling the data volume from public sources like Xeno-Canto and iNaturalist. The model can disentangle complex acoustic scenes over thousands or even millions of hours of audio data, enabling it to answer various questions, from "how many babies are being born" to "how many individual animals are present in a given area."
To aid scientists in protecting our ecosystems, we are releasing this new version of Perch as an open model available on Kaggle. Perch is not limited to recognizing bird sounds; the new model has been trained on a wider range of animals, including mammals, amphibians, and anthropogenic noise.
Since its initial launch in 2023, the original version of Perch has been downloaded over 250,000 times, and its open solutions have been well-integrated into tools for working biologists. For instance, Perch's vector search library is now part of Cornell's widely-used BirdNet Analyzer. Additionally, Perch is assisting BirdLife Australia and the Australian Acoustic Observatory in building classifiers for several unique Australian species. Our tools enabled the discovery of a new population of the elusive Plains Wanderer, a remarkable finding that will aid in shaping the future of many endangered bird species.
According to Paul Roe from James Cook University, recent research has shown that the earlier version of Perch can identify individual birds and track their abundance, potentially reducing the need for catch-and-release studies to monitor populations. Biologists from the LOHE Bioacoustics Lab at the University of Hawaiʻi have utilized it to monitor and protect honeycreeper populations, which are culturally significant in Hawaiian mythology and face extinction due to avian malaria spread by non-native mosquitoes. Perch helped the LOHE Lab find honeycreeper sounds nearly 50 times faster than their usual methods, allowing them to monitor more honeycreeper species over larger areas.
While the Perch model predicts species present in recordings, that is only part of the equation: we also provide tools for scientists to quickly build new classifiers from a single example and monitor species with scarce training data or specific sounds like juvenile calls. Given one example sound, vector search with Perch surfaces the most similar sounds in a dataset. A local expert can then mark the search results as relevant or irrelevant to train a classifier. This combination of vector search and active learning with a strong embedding model is termed agile modeling. Our recent paper, "The Search for Squawk: Agile Modeling in Bioacoustics," demonstrates that this approach works across birds and coral reefs, allowing the creation of high-quality classifiers in under an hour.
Looking ahead, our models and methods are maximizing the impact of conservation efforts, freeing up more time and resources for meaningful on-the-ground work. From the forests of Hawaiʻi to the reefs of the ocean, the Perch project exemplifies the profound impact we can achieve when applying our technical expertise to the world's most pressing challenges. Every classifier built and every hour of data analyzed brings us closer to a world where the soundtrack of our planet is one of rich, thriving biodiversity.
Learn more, download the new Perch model from Kaggle, read our papers, and explore our GitHub repository.
This research was developed by the Perch team: Bart van Merriënboer, Jenny Hamer, Vincent Dumoulin, Lauren Harrell, Tom Denton, and Otilia Stretcu from Google Research. We also thank our collaborators Amanda Navine and Pat Hart at the University of Hawaiʻi, and Holger Klinck, Stefan Kahl, and the BirdNet team at the Cornell Lab of Ornithology. And all our friends and collaborators whom we would have written about in this blog post if only we had another thousand words.
Blogger's Review: The launch of this model undoubtedly propels the field of bioacoustics forward, leveraging AI technology to rapidly analyze ecological audio data, significantly enhancing conservation efficiency. The flexibility and openness of the new model will help scientists tackle the increasingly severe challenges in ecological environments. Looking forward to more applications and developments in the future!