Can Nova-3 understand air traffic control?

The problem

ATC radio is close to a worst case for speech-to-text: 8 kHz AM audio, squelch noise, clipped phraseology at speed, and callsigns no general model has seen enough of. That makes it a sharp benchmark for what Deepgram’s keyterm prompting actually buys you.

What I built

A pipeline that takes a LiveATC stream, converts it with ffmpeg, and feeds identical audio bytes to two Deepgram Nova-3 websockets — one baseline, one with keyterm prompting. A Next.js UI shows both live transcripts with word-level confidence coloring and diff highlights, plus an audio relay of exactly what the model hears, so you can verify by ear.

Why the judgments hold

I hold a commercial pilot certificate and CFI, so the keyterm lists and the accuracy calls are grounded in how these radio calls are actually spoken — the demo measures a real domain, not a synthetic benchmark.