A low-cost sensor node on a Raspberry Pi Zero 2 W that measures noise level (dB(A)), recognises the type of noise on the device, and records temperature, humidity, pressure, air quality and power. Everything is uploaded to the Sunbird noise API. No audio ever leaves the device.
| Measured | How | Rate |
|---|---|---|
| Noise level, dB(A) | I²S MEMS mic (INMP441), A-weighted LAeq | every 60 s |
| Noise type (19 classes, e.g. boda-boda, generator, crowd) | YAMNet + TRILL model, runs on the Pi | every 60 s |
| Temperature, humidity, pressure, gas / air quality | BME680 (optional PMS5003 for PM2.5) | every 60 s |
| Battery voltage, RAM, CPU temperature | INA219, system | every 60 s |
Looking for the Arduino Mega / GSM Gen-2 sensor? That is a different design: SunbirdAI/urban-noise-sensor. Both report to the same API.
On a Pi running Raspberry Pi OS Bullseye 32-bit, logged in as pi, with internet and the hardware wired (wiring):
curl -fsSL https://raw.githubusercontent.com/SunbirdAI/seas/main/install.sh | sudo bash -s -- --device-id SEAS-4
Use your own unit's name for --device-id. It finishes by offering to reboot. It:
- installs the system and Python packages (TFLite, BME680, INA219 …);
- enables I²C / I²S / UART and configures the microphone;
- copies the scripts and the model to
/home/pi/environmentsensor/; - sets 512 MB of swap (the model needs about 200 MB of RAM);
- installs and starts two services that restart on failure and start at boot.
It is safe to run again to update a unit. Add --with-battery to also enable the (experimental) battery monitor service, or --reboot to skip the prompt.
Prefer to do it by hand, or something failed? See docs/manual-install.md.
seas-status # services, last uploads, memory, temperature
journalctl -u seas-env-logger -u seas-noise-inference -f # live logs
You should see 201 after each upload, one environmental and one classification per minute.
| Service | Script | Sends |
|---|---|---|
seas-env-logger |
pi/env_logger.py | temperature, humidity, pressure, air quality, RAM, CPU temperature, battery voltage, dB(A) |
seas-noise-inference |
pi/noise_inference.py | predicted noise class + probability |
seas-battery-monitor |
pi/battery_monitor.py | optional, off by default |
Per-device settings (device ID, calibration offset) live in /etc/seas/seas.env. After editing: sudo systemctl restart seas-env-logger seas-noise-inference.
install.sh one-command installer
pi/ scripts that run on the sensor
models/ noise classifier (v3 = YAMNet + TRILL); older versions in models/archive/
hardware/ PCB, schematic, enclosure
docs/ guides
rain-sensor/ rain detection from audio (separate pipeline)
legacy/ earlier scripts, kept for reference
- The dB(A) value uses a provisional calibration offset (from the microphone datasheet) until each unit is calibrated. See docs/calibration.md.
power_usageis currently the battery voltage, not watts.- Model accuracy is about 67 % on the test set; confidence is low in quiet places.
Different parts use different licenses, as is usual for open hardware.
| What | Where | License | File |
|---|---|---|---|
| Firmware and scripts | pi/, install.sh, rain-sensor/ |
GNU GPL v3.0 | LICENSE |
| Hardware design files (PCB, schematic, enclosure) | hardware/ |
CERN-OHL-W-2.0 | LICENSE-HARDWARE |
| Documentation and trained models | README.md, docs/, models/, rain-sensor/*.tflite |
CC BY 4.0 | LICENSE-DOCS |
Copyright (c) Sunbird AI Ltd.