Your backyard, but make it a multi-species data pipeline

What it does

BirdNET-Go listens to soundcards or RTSP audio streams around the clock, runs local AI classification on the audio, and serves results through a web dashboard. It targets birds by default, handles bats via ultrasonic-capable hardware, and can juggle multiple models against multiple audio sources simultaneously.

The interesting bit

The “multi-model gallery” is the unusual part: you install BirdNET, Google Perch, BattyBirdNET regional classifiers, and a location-aware geomodel from inside the app itself, then run them in parallel and let cross-model agreement vote on whether that really was a Eurasian blue tit. The project also ships with a browser-based terminal, TLS management, OIDC/SSO, and a “database doctor”—ops features you don’t typically find in hobbyist wildlife tooling.

Key highlights

• Runs on Raspberry Pi 4; single static Go binary with embedded TFLite model
• 6,500+ bird species via BirdNET v2.4; 14,795 species via Google Perch v2 (ONNX); 11 regional bat classifiers
• Cross-model consensus scoring to reduce false positives
• Alert rules engine routing to Discord, Slack, Telegram, MQTT/Home Assistant, webhooks, shell scripts, BirdWeather, and more
• Live spectrogram streaming and detection heatmaps rendered in-browser
• PWA installable, 15 UI languages, species names in 40+ languages
• Hot-reload settings and model installs without restart

Caveats

• Bat detection requires Linux plus ultrasonic-capable hardware; not all platforms get all models
• ONNX Runtime for Perch, BattyBirdNET, and the geomodel is optional and requires separate installation
• CC BY-NC-SA 4.0 license is non-commercial

Verdict

Worth a look if you want a self-hosted, offline-first bioacoustics station with actual production conveniences. Skip it if you need commercial use or just want a phone app to identify a single bird call.