Listen to the world.
A live bird recognition experiment. Audio stays on your device.
1. Prepare your device
Install about 70 MB once, then load the model. Use Safari on iPhone with this page open.
Ready to install. No microphone is active.
2. Use your region
Location and time of year help filter plausible species. Your approximate location stays on this device and is included in diagnostic exports. You can check location before installing the bird model.
Location not enabled. Listening is unfiltered.
Location access details
No request made yet.
3. Listen for birds
Keep your phone still. Each analysis covers three seconds. Stop listening before locking your phone.
No audio processed.
Keep this page foregrounded for the 20-minute stability check. After installation, test an offline restart separately.
What we hear
No bird suggestions yet.
Pilot 0.5: BirdNET browser display cutoff ≥ 0.15. No repeat required. Suggestions may be wrong; scores are not probabilities. Regional cutoff ≥ 0.05 when location is enabled; unusual visitors may be excluded.
4. Run validation tests
Start listening, then run this test indoors with headphones disconnected and moderate speaker volume. Select the original six calls, the robin/pigeon comparison, the negative controls, or the complete set. Each 12-second clip has a six-second processing gap. Your microphone hears them normally.
Start listening to enable the test. You can also play individual clips below, including from another device.
These are playback smoke tests, not a measured accuracy benchmark. Wind/traffic/speech controls are synthetic, not field recordings. The original weak robin and pigeon clips are retained. Recordings may overlap model training data and may contain other birds. Same-phone speaker/microphone routing needs iPhone validation. Test indoors to avoid broadcasting calls to wild birds.
Reference recordings, controls & credits
Test details & diagnostics
Export contains device details, species scores and timings. It contains approximate location (rounded to 0.01°) when enabled, and no audio. Detailed logs retain up to 200 events; aggregate timing and stability metrics cover the full session.
Unfiltered comparison
Top raw audio candidates from the same window, before regional filtering. Regional scores are separate from audio scores.