66 species over 68 days, identified by BirdNET-Go in York.
BirdNET-Go uses a local AI model running on a Raspberry Pi that uses about the same energy as a standard lightbulb.
Every one of 7,266 bird calls, drawn as points of light. Scroll downwards to see the season progress, one morning at a time.
The first call of each morning, set against that day's sunrise. The robin is usually first to sing, 44 minutes before the sun rises.
Three bird songs from 22 September, each fifteen seconds long. Lower frequency background noise has been filtered out, allowing the bird song to be more clearly heard. Each picture is a spectrogram: time runs left to right, pitch rises up the page, and every mark is a note.
Two-thirds of all bird song happened in the four hours around sunrise.
Each ridge in front is one species' day, from three in the morning to ten at night. Most rise steeply at dawn and fall away by mid-morning. Woodpigeons and magpies come back in the evening, coal tits keep going all morning, and the crows never quite stop.
One line for each species heard five times or more, one circle for every day on which it was heard. The bigger the circle, the more that species sang that day.
Still mornings brought more than five times as much bird song as windy ones.
Calm mornings* are loud ones. In the hours around sunrise, a still morning averaged 26 calls an hour. Once the wind passed 20 km/h, that fell to under 5. Some of that is the birds, which sing less in wind, and some is the microphone, which hears less through the noise. Rain barely mattered: it fell in only five of the morning hours all season.
* Morning means half an hour before sunrise to three and a half hours after. Local hourly wind speed data from Open-Meteo.
Two species of owl, with fourteen calls. Tawny owls were heard on four nights, the first on 15 August. Barn owls were heard in September, on three nights, and were almost always detected with high confidence: five of six calls at 100%. Both the tawny owl and barn owl were detected on the night of 21–22 September.
The small moons on the y-axis show the moon's phase on that night. Most owl nights were dark ones: on four of the six nights, the moon was about a third lit or less.
34 species heard three times or fewer. Some were likely passing over. Others, like a night heron in York, are more likely BirdNET-Go mishearing something more mundane; the less certain it was, the fainter the species' name.
All the birds detected in the visualisations on this page were picked out by a Raspberry Pi 5 listening to a microphone at the front of our house, 24/7. BirdNET-Go chops the sound into three-second slices, converts each one into a spectrogram, and runs it through a small AI model trained to recognise bird songs. All the AI runs locally on the Raspberry Pi, and no audio data is sent anywhere; no data centres involved!
The AI model used is BirdNET, which began as a research project at the Cornell Lab of Ornithology's K. Lisa Yang Center for Conservation Bioacoustics, together with Chemnitz University of Technology. Trained on recordings from around the world, it now knows the songs and calls of more than 6,000 species. BirdNET-Go is an independent, open-source project that wraps it up and runs it on a tiny computer that uses the same energy as a standard lightbulb and gives you a nice web interface to view the data.
66 species.
7,266 songs.
68 days.
Recorded and identified with BirdNET-Go. BirdNET model detections only, exported from InfluxDB; 6,819 Perch model detections were excluded because the model did not run consistently. Hooded and Carrion Crows are counted together as Carrion Crow, and 36 rows labelled only Coloeus monedula are counted as Eurasian Jackdaw. Sunrise and twilight times are calculated for York.