I fed a global bird model foreign audio. It was confidently wrong. Then I fixed it.

Description and use

BirdNET, a global bird sound classifier, incorrectly identified species not present in Chidambaram, India. A local pipeline was built to download the model weights, generate a regional species list from iNaturalist, and filter predictions, yielding accurate local identifications.

Problem

A global bird identification model confidently predicted species that do not exist in the local region, leading to misleading results.

Outcome

The pipeline achieved 45.3% of detections within the regional species pack, compared to 0% for the global model on local audio. Analysis time was under 4 s for 83.8 s of audio after a 1 s model load.

Workflow

• download-model.mjs – fetch BirdNET+ weights and labels
• fetch-pack.mjs – build local species list from iNaturalist within a radius
• convert.mjs – convert .m4a/.mp3/.ogg to 32 kHz mono WAV
• analyze.mjs – run ONNX inference and write detection CSV
• summarize.mjs – score predictions against the local pack and generate docs/overview.md
• case.mjs – generate a case study from detection data

Evidence

BirdNET is a genuinely impressive open-source project. Point it at a recording and it will confidently tell you what you heard.

Source

The model was not confused - it was sure, and wrong.

Source

chidambaram-birdnudge is that button. It runs the same open weights on your own machine, then filters every prediction through a species list built from real observations near you.

Source

The pipeline is five scripts: ...

Source