Listening to birds with AI: a biodiversity experiment
Running Apple Watch field recordings through Google DeepMind's Perch model to see whether consumer devices can feed research-grade biodiversity data.
By Writam Bugga · · 2 min read
Carbon is one lens on climate. Biodiversity is another, and it is even harder to measure. I wanted to know whether casual recordings could produce useful ecological data, so I tried it.
The setup
I had fifteen field recordings from an Apple Watch, from eight seconds to more than thirteen minutes long. I ran them through Perch, Google DeepMind's open bioacoustics model, in a Colab notebook. I used one AI assistant to design the pipeline and another to run the code, while I set the goals and checked the work.
The blocker that taught me the most
The first run failed with a tensor shape mismatch. Perch expects exactly five seconds of audio at 32 kHz, which is 160,000 samples. My 16-second test file had about 512,000.
The fix was a sliding window: five-second segments with 2.5 seconds of overlap, so a bird call on a boundary is not lost. All six segments of the test file then processed, returning class scores and 1,280-dimensional embeddings.
What is still missing
- Mapping class indices to species names, using a taxonomy such as eBird's.
- A proper biodiversity report with habitat context and timing patterns.
- Batch processing and visualization.
Why it matters
- Consumer hardware can be good enough. That opens citizen science and cheap monitoring.
- Model constraints drive architecture. A fixed input size shaped the whole pipeline.
- Nature-based solutions need measurement too. Restoration and conservation claims face the same credibility question as product carbon claims.
The experiment lives at birdconservationdata.netlify.app. It is a learning project, not a monitoring tool, and the Cornell Lab of Ornithology remains the place for serious bird science.