machine-learning A field note by Vikrant Sharma
Security cameras that identify birds by sound
Someone repurposed their home security setup to log every bird species that flies past. The microphone was already there.
Jason Tucker turned his security cameras into an automatic bird identification system using BirdNET-Go. The interesting bit is not that bird identification models exist. It is that he already had the hardware sitting on his house. Security cameras ship with microphones. Most people never think about the audio feed. Tucker routed the RTSP audio stream through BirdNET-Go, which runs inference locally on a Raspberry Pi. The model listens for bird calls, identifies the species, and logs timestamps. No cloud API. No subscription. The setup runs 24/7. He has a database of every bird that showed up in his yard, sorted by time of day and season. The model handles overlapping calls and background noise without manual cleanup. This is the kind of hardware repurposing that makes sense once you see it. Cameras already capture audio for motion detection alerts. That audio feed is a structured data source if you point a model at it. BirdNET-Go is open source and trained on thousands of species, so the accuracy is good enough for casual logging. The practical angle is using existing infrastructure differently. Most IoT devices have sensors that only serve one narrow feature. A camera microphone listens for glass breaking or voices. A doorbell mic listens for the button press. But the raw audio stream is still there, and it can feed other models without replacing the hardware. I would add a notification filter. Logging every sparrow is fine for a database, but I would want alerts only for rare species or migration events. The model outputs confidence scores, so thresholding by species rarity is straightforward. Still, this is a better use of a security camera than watching an empty driveway.
Source: I turned my security cameras into an automatic bird identification system
