A conservation foundation from the end of the world, a Google Research collaboration, and 49,383 acoustic detections later, this is what happened.The goal is the first complete acoustic inventory of a Chilean Patagonian nature sanctuary using a 14,795-taxon foundation model.We are Fundación Kreen. We operate from Coyhaique, in Chilean Patagonia 45°34′16″S 72°04′07″O / -45.5712, -72.0685 a city of 60,000 people surrounded by mountains, rivers, and fjords that most of the world has never heard of. We manage the Meullín-Puye Nature Sanctuary, 100,000 hectares (247105 acres) in remote Patagonian fjord of one of the least-disturbed coastal ecosystems on the planet (-45.1386, -72.9697).We are not a tech company. We are seven people in a NGO who work in conservation at the end of the world.But this April, we built something that, honestly, we did not expect to build.The Problem Nobody Was TrackingThe fjords of Chilean Patagonia look pristine. And in many ways, they are. But underneath that postcard image, something is happening that nobody has a system to monitor.Chile is the second largest salmon producer in the world. The industry operates hundreds of farming centers directly inside the fjords, the same fjords where six threatened and near-threatened seabird species nest, feed, and migrate. These birds, the Flightless Steamer Duck (Tachyeres pteneres, Vulnerable), the Black-browed Albatross, the Pink-footed Shearwater, the Imperial Cormorant,transit between the industrial aquaculture zones and pristine protected areas like our sanctuary.The salmon industry releases over 350 metric tons of antibiotics into Patagonian waters every year. The birds eat in those waters. Then they fly to the sanctuary. Nobody was measuring what they carry with them.And then, on April 2026, two days before we submitted our grant application, Chile’s SAG (Servicio Agrícola y Ganadero) declared a sanitary emergency in the Biobío Region after confirming H5N1 avian influenza in a wild bird in Arauco. Nine Chilean regions now have confirmed cases. Aysén has no active bio surveillance system for wild bird H5N1 vectors. That is the problem we are trying to solve.We call the project Silent VectorsThe central hypothesis is this: migratory seabirds transiting between salmon aquaculture centers and pristine peatlands are acting as biological vectors, silently transporting antimicrobial resistant bacteria, nutrient subsidies, and potentially H5N1 across ecosystem boundaries. They are the connective tissue between an industrial zone and a protected area. And nobody is listening to them.Until now, Perch v2 by Google DeepMindWe had been running three Bird Weather PUC acoustic sensors at the sanctuary since November 2024. These little devices record ambient audio continuously and process it through BirdNET, generating validated species detections. Over 14 months, they produced 1,701 confirmed detections across 21 species, real field data, georeferenced, timestamped, stored in our Google Drive.But BirdNET, as good as it is, is a black box. We needed something open-source, trainable, and scientifically publishable.That is when we found Perch v2.Perch v2 is Google DeepMind's bioacoustics foundation model 14,795 species, 1,536-dimensional embeddings, trained on a massive global corpus. Unlike Perch 1 (which was bird-vocalization-classifier/4 on TF Hub and had significant Southern Hemisphere gaps), Perch v2 uses iNaturalist taxonomy and actually covers the species we care about.We loaded it via perch-hoplite:from perchhoplite.zoo import modelconfigsmodel = modelconfigs.loadmodelbyname('perch_v2')outputs = model.embed(waveform)# outputs.embeddings: shape (1, 1, 1536)# outputs.logits['label']: shape (1, 14795)Then we built a geofencing layer, a curated list of 35 species ecologically relevant to Aysén, mapped by scientific name directly to Perch v2's iNaturalist taxonomy indices. Species like Tachyeres pteneres at index 13,328. Thalassarche melanophris at index 13,594. Falco peregrinus at index 5,088.==TARGET_SPECIES = {===='Tachyeres pteneres': ('Quetro No Volador', 'CENTINELA', 'VU'),===='Thalassarche melanophris': ('Albatros Ceja Negra', 'CENTINELA', 'NT'),===='Ardenna creatopus': ('Fardela de Collar', 'CENTINELA', 'VU'),====# … 32 more species across 6 habitat categories====}==Calibrated thresholds by habitat type,40% for sentinel species (more sensitive), 60% for resident forest species (more strict), and ran the full analysis on 4,998 FLAC recordings from the sanctuary.The ResultsWe were not expecting what came out.49,383 acoustic detections. 35 species. All 6 sentinel species detected.| Species | IUCN | Detections | Max Confidence ||----|----|----|----|| Thalassarche melanophris | NT | 2,739 | 99.6% || Spheniscus magellanicus | NT | 2,494 | 98.5% || Tachyeres pteneres | VU | 1,343 | 74.3% || Leucocarbo atriceps | LC | 883 | 90.1% || Ardenna creatopus | VU | 415 | 99.6% || Leucophaeus scoresbii | NT | 180 | 97.9% |The Flightless Steamer Duck,Tachyeres pteneres,the most critical sentinel species for our project, the one that was completely absent from Perch 1's training corpus, appeared 1,343 times. At 74.3% maximum confidence. In recordings made at coordinates where we have confirmed field sightings.The Peregrine Falcon appeared in 4,998 detections at 99.4% confidence. A known trans-hemispheric migrant. Exactly where it should be.The Google Research ConnectionHere is something we did not plan for but that changed everything.While working on this, we noticed that Perch 1 was missing several Southern Hemisphere species from its training data. We filed Github issue in the google-research/perch repository, reporting the gap and framing it as a One Health biosurveillance problem,not just a birding inconvenience.Tom Denton, a researcher from Google Research who works on Perch, responded within hours. He confirmed that all 11 missing species are actually covered in Perch v2, pointed us to the correct model variant, and offered a free technical support for us. His bio says "in it for the birbs."He also provided the URL to download the Perch v2 labels file directly from Google Cloud Storage,which solved a technical problem we had been fighting for two days:[https://storage.googleapis.com/chirp-public bucket/models/perchv2/assets/labels.csv ](https://storage.googleapis.com/chirp-public bucket/models/perchv2/assets/labels.csv)That one line of information was worth more than five days of debugging.What We Learned About Running Perch v2 in the Real WorldA few things that are not obvious from the documentation:bird-vocalization-classifier/4 is Perch 1, not Perch 2. Despite the version number. Perch v2 is loaded via perch-hoplite and uses a completely different API,model.embed() instead of model.infer_tf().Perch v2 from Kaggle requires TensorFlow 2.20+. The XlaCallModuleOp version 10 error you will hit on Colab is a TF version mismatch. Update TF before loading the model.Perch v2 uses iNaturalist taxonomy with scientific names in lowercase. Not eBird codes. speciestoidx['tachyeres pteneres'] not speciestoidx['fltduc1']. This tripped us up for an entire afternoon.The labels CSV is not bundled with the Kaggle model. Download it separately from the URL Tom provided.Geofencing dramatically improves signal quality. Without it, Perch detects European and North American species in Patagonian soundscapes,not because they are there, but because the training corpus is biased toward the Northern Hemisphere. Filtering to locally relevant species cuts false positives by roughly 98%.What Comes NextWe have 53,818 FLAC recordings sitting in Google Drive. We are running the full analysis now,all of them, no geofencing, every taxon Perch v2 knows about at ≥70% confidence. The goal is the first complete acoustic inventory of a Chilean Patagonian nature sanctuary using a 14,795-taxon foundation model.We are also applying to the Google.org Impact Challenge: AI for Science with this work. The project,Silent Vectors— proposes combining Perch v2 with AlphaEarth Foundations (for peatland carbon dynamics) and field biosurveillance protocols with Universidad de Concepción to build the first One Health automated monitoring system for H5N1 and AMR in Patagonian seabirds.All code is open source. The full acoustic dataset will be published on Google Earth Engine under CC-BY 4.0.A Note on Working From the End of the WorldPeople sometimes ask us why a small conservation foundation in Coyhaique is building AI pipelines instead of just doing fieldwork.The honest answer is that fieldwork alone is no longer enough.The H5N1 emergency in Chile is real and moving south. The salmon industry is expanding. The fjords are changing faster than anyone can survey them manually. The only way to generate the baseline data that conservation decisions will need for the next 30 years is to automate detection at scale,and to do it with models that are open, reproducible, and honest about what they know and what they do not.Perch v2 is not perfect. may It needs fine-tuning for Southern Hemisphere species. It requires calibration for the specific soundscape of Patagonian fjords. We know that.But 49,383 detections including all 6 sentinel species,including a Vulnerable endemic duck that most acoustic models have never heard of,is a pretty good start.\… in it for the birbs!!\