Hello everyone,

I am a hobbyist software developer and nature enthusiast, and while I don’t have a background in ornithology or academic biology, I wanted to share an open-source engineering project I’ve been building over the last year. I'm hoping to get some feedback from the acoustic and conservation experts here!

The project is called Global Birdsong Radio (currently live at wilddetection.net).

The Concept:
It acts as a "Planetary Ear." It is a distributed network that actively listens to 50+ live nature cameras and microphones around the world (YouTube, IP Cams, Icecast). When a bird sings, it pushes a real-time notification to the user and logs it on a live interactive map.

The Architecture:
Because running 50+ live audio streams through a cloud server is prohibitively expensive and triggers massive IP bans, I built this using a Hybrid Edge/Cloud architecture:

The Edge (Local): A Python-based engine running locally on my hardware. It handles stream ingestion (via yt-dlp and headless browsers), audio normalization, and runs the BirdNET AI model locally.

The Hydra Network: To bypass YouTube rate limits, the local engine binds specific listening processes to different physical network adapters (SIM cards/Wi-Fi interfaces) to distribute the traffic.

The Cloud: Only lightweight metadata (timestamp, SNR, species, location) and small MP3 clips of verified detections are beamed to a central DigitalOcean PostgreSQL database and Flask API, which drives the public web map.

Handling the "Firehose" of Data:
My biggest challenge was false positives and alert fatigue (e.g., getting many alerts a day for Pigeons). To solve this, I built a few intelligence layers:

Bayesian Cooldowns: The system keeps a 14-day rolling history of every stream. It dynamically learns if a bird is a "Resident" (mutes alerts for hours) or a "Rare Visitor" (alerts immediately).

Adaptive SNR Distance: The engine measures the RMS volume of the detection against the background noise floor to establish a "Golden Anchor" (loudest known volume) for that specific species on that specific stream, allowing it to tag alerts as "Very Near" or "Far."

DSP Bioacoustic Vetoes: I added custom DSP math (checking spectral flatness and frequency isolation) to detect non-birds (like African Cicadas, Howler Monkeys or Tink Frogs), while using BirdNET's weather classifications (wind/rain) to veto false positives.

My Ask to this Community:
As an engineer, the code is solid, but I would love the perspective of actual researchers and bioacoustics experts:

Does a tool like this have value for citizen science or migration tracking?

Are there better open-source models I should look at beyond BirdNET and YAMNet for the edge?

Any feedback on the Web Map UI (wilddetection.net) would be deeply appreciated.

Thank you for your time, and I look forward to learning from this community!