RESEARCH2026-05-02chien.digital

How we build a case

From raw VHF audio to a published case study — the methodology behind hungATC's research archive.

Introduction

A case study on hungATC begins with a single VHF transmission and ends with a footnoted document anyone can audit. Between the two sit roughly forty hours of work spread across capture, transcription, verification, and publication. This is the path a recording takes through the system.

Audio capture

We tap the standard ICAO frequencies for RCSS and RCTP via a Direct Feed receiver running on the GX10 server in Taipei. The stream is segmented into rolling thirty-second chunks, each tagged with channel, frequency, RSSI, and a UTC timestamp from a disciplined NTP source. Nothing is discarded — even silence is logged as a positive measurement that the receiver was online.

Two-stage transcription

The first pass runs hungASR, a Whisper Large V3 fine-tune that produces a draft transcript with word-level timestamps. The second pass sends the same audio through Gemini with the hungASR draft as a hint, which catches numerals, callsigns, and the kind of tower jargon Whisper alone misreads. Disagreements between the two passes flag the segment for human review.

Human verification

A reviewer listens to the original audio against the transcript and scores it on three axes: phonetic accuracy, callsign correctness, and instructional intent. Any score below the publication threshold sends the segment back to the queue. We log every reviewer touch so the verification chain is auditable later.

Publication

A case study renders from the verified transcript into MDX and ships through the same static export as the rest of the site. The audio file, the transcript JSON, and the rendered article all share a case ID so an external researcher can reproduce the chain end to end.

Limitations

VHF capture is line-of-sight and lossy, so distant aircraft drop in and out. We publish coverage maps with each case so readers know what was missed. Transcription is good but not perfect — every published case carries a confidence band and a public correction policy.