The pipeline that works
Four stages, each replaceable:
| Stage | What it does | Tools we use |
|---|---|---|
| Capture | Tune + demodulate RF to audio | HackRF One / Airspy Mini via SDR software |
| Transcribe | Speech → text, locally | Whisper (open source), no cloud |
| Understand | Classify intent/topics, alert on keywords | Small local classifier models |
| Archive | Searchable history + dashboards | SQLite + live dashboard |
Transmitter fingerprinting (identifying devices by signal characteristics) adds who-said-it on top of what-was-said.
Hardware reality table
| Setup | Cost (AUD) | What you get |
|---|---|---|
| RTL-SDR dongle | ~$40 | Receive-only, narrow bands — enough to start |
| Airspy Mini | ~$180 | Continuous 24/7 monitoring, better dynamic range |
| HackRF One | ~$450 | Wideband, transmit-capable for licensed use |
Any modern PC runs the AI side; a GPU matters only for 24/7 multi-channel transcription.
Legal and ethical lines (read this part)
Working reference implementation
We built RFAI (v0.8.0) to do exactly this: SDR capture → Whisper transcription → intent classification → fingerprinting → searchable archive, as a desktop app with no cloud calls.
The RFAI dashboard as deployed: spectrum view, live transcription and classified signals.
Verify it yourself (don't trust us)
The live demo runs the actual interface in your browser. Technical specifics so you can compare against your own build:
| Component | What we run |
|---|---|
| Speech-to-text | Whisper (open weights), local inference — no API keys anywhere in the pipeline |
| Architecture | Shared DSP library + SvelteKit/Tauri 2 desktop front end |
| Capture | Airspy Mini and HackRF One, continuous duty cycle |
| Archive | SQLite; text-only storage keeps a year of a busy band under ~1GB |
| Version | v0.8.0, in daily use at the studio since mid-2026 |
When NOT to build this
If you just need to transcribe a podcast or a meeting, use plain Whisper on the audio file — no radio hardware needed. If you need phone calls or encrypted traffic: stop, that's interception territory.
FAQ
Can Whisper handle noisy radio audio? Yes, with practice — it's robust to compression and moderate noise; pre-filtering audio helps most.
How much storage does archiving need? Text is tiny — a year of monitoring a busy band is a few hundred MB of SQLite. Keep audio only if you must.