From a sound overhead to a name in the log.
HeliTracker is four things working in sequence: a microphone that never stops, a model that knows what a rotor sounds like, a lookup that finds the flight, and a record that anyone can read.
Listen
A weatherproof microphone feeds audio to a small onboard computer, sampled at 16 kHz, in short overlapping windows. Nothing is written to disk. The audio exists for about a second and is gone.
Recognize
Each window runs through an audio classification model trained on a large library of labeled environmental sound. It scores the window for "helicopter" against hundreds of other classes: traffic, wind, aircraft engines, birds.
Match
When the helicopter score stays high across consecutive windows, the device queries live flight tracking for rotorcraft near its coordinates. Tail number, operator, altitude, and heading are attached to the event.
Record
The event is written to a shared database and pushed to the phone app. Total time from the pass to your phone: usually under a minute.
The device
There is deliberately nothing exotic in it. Every part is off the shelf, replaceable, and cheap enough that a community group can afford several.
- Computer
- Raspberry Pi single-board computer running a hardened Linux install
- Microphone
- USB microphone in a weather shield, mounted outdoors or at an open window
- Sound model
- YAMNet audio event classifier, run locally on the device, tuned for rotorcraft
- Flight data
- Live ADS-B and transponder data, queried per detection
- Storage
- Cloud database of detection events only. No audio files, no video
- Alerts
- Push notification to the iPhone app within about a minute of each pass
- Reliability
- Hardware watchdog auto-restarts a frozen unit; persistent logs survive power loss
- Power
- Standard wall outlet, roughly the draw of a phone charger
- Network
- Wi-Fi or Ethernet
What it does not do
People reasonably ask whether a microphone on the block is listening to them. It is not, and the design makes that a property of the system rather than a promise.
Audio is processed in memory and discarded within seconds. The classifier outputs a score, not a transcript. The only thing that leaves the device is a short structured event: time, duration, confidence, and the matched flight. There is no recording function to switch on, and no stored audio to subpoena, leak, or lose.
The device also has no camera. It does not track people, cars, or anything on the ground. It tracks one thing: helicopters.
Everything in a single log entry
One helicopter pass produces one row. These are the fields in it, the same columns you get in the spreadsheet export.
- Timestamp
- Local date and time the pass was heard
- Registration
- The aircraft's tail number
- Callsign
- The identifier the flight was using on the radio, when different from the tail number
- Operator
- The company or owner the aircraft is registered to
- Aircraft type
- Make and model, such as Sikorsky S-76 or Bell 407
- Altitude
- Height in feet at the moment it was matched
- Heading
- Direction of travel as a compass bearing
- Loudness
- Average and peak sound level during the pass, in decibels
- Anonymous flag
- Set when the aircraft was flying without broadcasting its identity
- Position
- The aircraft's latitude and longitude at the match, so it can be placed on a map
Why the flight match matters
A count of loud moments is easy to dismiss. A count of named aircraft, with altitudes and operators, is something an operator, an airport manager, or a town attorney has to answer.
Altitude patterns
Over time the log shows whether aircraft are honoring voluntary altitude agreements or flying well under them.
Operator concentration
Most corridors are dominated by a handful of operators. The data shows which ones, and how often.
Time of day and week
Friday afternoons and Sunday evenings tell a different story than a monthly average. The log keeps every pass so the pattern is visible.