Disclosure: this write-up was drafted by AI (Claude) in the interest of last-minute time constraints.

Inspiration

Most US airports have no control tower. Pilots there separate themselves by announcing their position on a shared frequency (CTAF) and looking out the window. Aircraft without ADS-B Out show up on no traffic display; their only trace is their radio calls. Nobody is listening and comparing.

What it does

  • Streams the airport's radio (LiveATC) and live ADS-B (adsb.lol).
  • Cuts each transmission at the squelch, transcribes it with ElevenLabs Scribe, and parses position reports ("left base runway 26").
  • Matches the speaker to an ADS-B aircraft, or creates a radio-only "ghost" track with an uncertainty circle when the aircraft isn't on ADS-B.
  • Predicts every aircraft along the traffic pattern and finds conflicts ahead of time.
  • Speaks short warnings with ElevenLabs TTS, sped up and filtered to sound like VHF radio. Warnings wait for a clear frequency and cut out if a pilot keys up, so real calls aren't stepped on.
  • Archive mode replays real recorded near-misses with a timeline: click a warning to jump to it, hover to freeze-frame the situation.

How we built it

Python (uv), FastAPI + websockets, PyAV for incremental MP3 decode, ElevenLabs Scribe v3 and Flash v2 TTS, Leaflet map, Web Audio for the radio effect. Evaluated on a week of real Butler, PA (KBTP) radio + ADS-B from the TartanAviation dataset.

Results (KBTP week, 35 h)

  • Flagged every real near-miss in a week of recorded traffic: 9 of 9 losses of separation (< 0.5 NM) found in the ADS-B data, at 6.6 alerts per hour.
  • Speaker identification: 82% on hand-labelled calls.
  • ~1.2 s from end of speech to spoken advisory.

Challenges

Matching misheard callsigns to aircraft; radio-only aircraft with no position but their own words; keeping the alert rate low enough to be useful; not talking over pilots.

What's next

Our live demo listens to the public LiveATC feed for Whiteman (KWHP), a community airport whose tower is part-time, so it is an uncontrolled CTAF field when the tower is closed. It is a demo only, not used by pilots or controllers. The same advisor could also serve as a controller in-ear aid while the tower is open. Next: Burbank, parsing tower phraseology, and a callsign recommender to reduce sound-alike N-numbers.

Built With

Share this project:

Updates

Submission history