About — RFAI Monitor

What it does

RFAI is a six-layer intelligence stack sitting on top of a software-defined radio. Each layer adds understanding — from raw electromagnetic signal to searchable, queryable insight.

🔊

CAPTURE

SDR hardware samples RF spectrum at up to 20 Msps. VAD detects voice — squelch opens, recording begins. Saved as WAV automatically.

→

✍️

TRANSCRIBE

Faster-Whisper runs locally (tiny.en → small → medium). Streaming partial transcripts appear within seconds of speech. No data leaves your machine.

→

🧠

UNDERSTAND

Entity extraction pulls callsigns, frequencies, locations, Q-codes. Intent classification labels each transmission: convoy, emergency, check-in, navigation…

→

🔗

CORRELATE

Voice biometrics (MFCC + CTCSS) cluster transmissions by probable radio hardware. The same physical radio — same fingerprint — even across frequencies.

→

🚨

ALERT

Urgency scoring flags mayday, emergency, all-stations traffic. Alerts via desktop notification and Kokoro TTS spoken readback — hands-free monitoring.

→

🗄️

Frequency coverage

Any band the HackRF (or Airspy / RTL-SDR) can receive. Current presets cover Australian domestic bands — add any frequency with the custom VFO.

476–477 MHz

UHF CB

NFM · 40 channels

Trucking, tradies, convoys. Australia's busiest radio band.

118–137 MHz

Aviation VHF

AM · 25 kHz spacing

ATC, ATIS, pilot comms. 121.5 MHz international guard.

156–174 MHz

Marine VHF

FM · ITU channels

Coastal and harbour traffic. Ch16 = distress.

144–148 MHz

2m Amateur

FM · repeaters + simplex

Ham radio. 146.52 MHz national simplex calling.

88–108 MHz

FM Broadcast

WFM · 200 kHz spacing

Commercial FM radio. Wide-band demodulation.

26–28 MHz

HF CB (27 MHz)

AM · 40 channels

Long-range CB. Skip conditions during solar max.

162–163 MHz

NOAA / BOM

FM · continuous broadcast

Weather alerts, hazard warnings, storm data.

430–440 MHz

70cm Amateur

FM / digital

Ham radio. Digital modes, APRS, repeaters.

1 MHz – 6 GHz

Custom VFO

WFM / NFM / AM / SSB

Tune to any frequency. LNA/VGA gain configurable.

Technology stack

Hardware

HackRF OnePrimary SDR — 1 MHz to 6 GHz, 20 Msps, half-duplex TX/RX

Airspy MiniFuture: 24–1700 MHz, 12-bit ADC, true IQ — enables RF fingerprinting Tier 2b

RTL-SDR v3Budget option — 500 kHz–1.75 GHz, 8-bit, receive-only

Signal Processing

libhackrf (ctypes)Direct hardware access — no GNU Radio dependency

numpy / scipyFFT, FIR decimation, demodulation chains (WFM/NFM/AM)

SoapySDRHardware abstraction — same code drives HackRF, Airspy, RTL-SDR

sounddeviceLow-latency audio output, 400ms ring buffer, PortAudio backend

AI / ML

faster-whisperCTranslate2-optimised Whisper. tiny.en → small → medium. ROCm GPU supported.

Kokoro-82MTTS readback and transmit synthesis. Apache 2.0. am_michael voice.

Ollama (local LLMs)Intent classification. phi4:14b, qwen2.5-coder:7b, gemma2:2b. Falls back to keyword rules.

BERT NERdslim/bert-base-NER — entity extraction (locations, organisations)

HDBSCANVoice + CTCSS biometric clustering — identifies probable transmitter hardware

Backend

FastAPI + uvicornREST API + Server-Sent Events. Port 4310.

SQLite + FTS5All transmissions, full-text search, zero config. WAL mode.

watchdog (inotify)Transcribe daemon: inotify sentinel protocol watches audio dirs

httpx + SSE proxyDashboard proxies daemon SSE events to all browser clients

Frontend

Svelte 5 + SvelteKitReactive UI. SSE live updates. SPA build via adapter-static.

Server-Sent EventsLive feed: recording_started → partial_transcript → transmission_complete

Web Audio APIClick sounds, mic capture (hold-to-record PTT), waveform visualisation

MediaRecorder APIBrowser mic capture for transmit panel demo mode

Transmitter fingerprinting

RFAI can identify individual physical radios — not just voices — using a multi-tier biometric system. Each tier adds confidence.

Tier 0

Verified callsign

Whisper extracts a formal callsign (VK2ABC, etc.) from the transcript text. High confidence — the operator stated their identity on air.

VK2ABC · ✓

Tier 1

Spoken handle

Pattern matching extracts self-reported aliases ("this is Night Hawk"). Medium confidence — self-reported, not verified.

Night Hawk

Tier 2

Radio hardware cluster

18-dimensional feature vector: MFCC voice coefficients, fundamental pitch, spectral centroid, CTCSS sub-audible squelch tone, duration patterns. HDBSCAN clusters transmissions by probable physical radio.

CTCSS is the strongest discriminator — it is programmed into the radio itself, so the same hardware always emits the same sub-audible tone regardless of speaker.

~C1042 · ≈

Tier 2b

IQ RF fingerprinting Coming — needs Airspy

With raw IQ data from the Airspy Mini: carrier frequency offset, phase noise signature, transmitter turn-on transient. True hardware fingerprinting that works even without voice or CTCSS.

Pages in this app

📡

Dashboard

Live monitor. Real-time transmission feed, band activity chart, signal level, partial transcript streaming, operator tracking. MIC DEMO button lets you inject audio without SDR hardware.

📋

Log Room

Full transmission history. Filter by band, intent, urgency, date range. Full-text search across all transcripts. Flag transmissions for long-term retention.

💬

Chat Room

IRC-style view. Transmissions are "messages", bands are "channels". Identity tiers shown visually. Live bubble streams partial transcripts. Transmit panel: type text or hold to record voice.

📻

Scanner

Spectrum waterfall + VFO controls. SDR hardware configuration. Band presets, gain settings, squelch threshold, scan mode. Requires connected SDR hardware.

📊

Analysis

Stats over time. Activity sparkline by hour. Intent breakdown chart. Multi-model comparison bakeoff: Claude vs phi4 vs qwen vs keyword. Live model arena.

🔬

Transmitter ID

Speaker/radio cluster viewer. Shows all identified probable-same-hardware groups. Cluster quality score, CTCSS tone, dominant band. Manual label tools.

ℹ️

About

This page. Full system architecture, technology stack, frequency coverage, roadmap, demo mode guide.

Demo mode — no SDR required

RFAI runs without any radio hardware. Three ways to inject audio:

1

🎙 MIC DEMO button

Click 🎙 MIC DEMO in the Dashboard Live Status panel or Chat Room. Records from your computer microphone on the server side. Writes a WAV file to /tmp/rfai-demo/ which the transcribe daemon picks up automatically via the sentinel protocol. Your voice appears in the live feed as a real transmission.

POST /api/demo/mic-record · POST /api/demo/mic-stop

2

🎙 Hold-to-record PTT (Chat Room)

In the Chat Room transmit panel, hold the 🎙 HOLD button. Uses the browser's MediaRecorder API to capture from your microphone. On release, the audio blob is sent to POST /api/transmit/audio. Played through the soundbar and logged to the DB as source='transmitted'.

3

✍️ Text transmit (Chat Room)

Type a message in the Chat Room and press ▶ TRANSMIT. Kokoro TTS synthesises it as the am_michael radio voice, plays through your soundbar, and the transmission is logged to the DB.

POST /api/transmit/text · body: { text, band, frequency_hz }

Roadmap

✅ DoneP0 — Scanner Foundation

HackRF One — live IQ capture via ctypes
Spectrum + waterfall display (PyQt5)
WFM / NFM / AM demodulation with FIR decimation chains
RMS VAD + adaptive squelch gate
Auto-record WAV on transmission detect
Band presets: UHF CB, Aviation, Marine, Amateur, FM

✅ DoneP1–P5 — Intelligence Core

faster-whisper STT — streaming partial transcripts via SSE
Kokoro-82M TTS — spoken readback and transmit synthesis
Entity extraction — callsigns, Q-codes, locations, phonetic alphabet
Intent classification — 8 categories + urgency score (0–1)
SQLite + FTS5 — all transmissions, full-text search
7-model intent bakeoff (Claude, phi4, qwen, gemma, keyword)

✅ DoneP6 — Web UI Showcase

Svelte 5 SPA — Dashboard, Log Room, Chat Room, Scanner, Analysis, TX ID
Live SSE feed — recording_started → partial transcript → complete
MIC DEMO mode — showcase without SDR hardware
Transmit panel — TTS synthesis + soundbar playback + DB logging
Transmitter fingerprinting UI — cluster browser, identity tiers
FastAPI backend + uvicorn + CORS on ports 4310/4311/4312

🔧 In ProgressP7 — UI Polish

Light mode + alternate dark themes (CSS variable swap)
Audio playback — inline player for archived transmissions
Compact message density — reduce row padding
Frequency filter pills + colour-coded by freq or ID
CB/Marine/Aviation channel lookup table (477.125 → "UHF CB Ch40")
Flag SSE broadcast — live sync across browser tabs
Svelte source committed to git (currently only build is tracked)

📋 NextP8 — RF Fingerprinting Tier 2b

Airspy Mini integration — 12-bit IQ, true carrier phase data
Carrier frequency offset extraction — per-radio systematic offset
Transmitter turn-on transient — unique hardware signature
Phase noise floor — distinguishes radio model + unit
Combined IQ + voice + CTCSS fingerprint — high-confidence ID without callsign

📋 NextP9 — Multi-band Correlation

Cross-band operator tracking — same voice heard on UHF + Marine
Channel-hop detection — "go to 5" → auto-follow to ch5
Convoy/group session stitching — link related transmissions
Temporal pattern analysis — who transmits when, at what cadence

🔮 FutureP10 — qalarc.com Public Demo

Deploy on qalarc.com — live demo of the full UI and stack
Demo data: curated synthetic transmissions covering all intent types
Read-only public view — no live SDR in cloud (legal/practical)
Replay mode — watch a stored session play back in real time
Shareable permalinks — link to a specific transmission or time range

Architecture

┌─────────────────────────────────────────────────────────────────────┐
│  HARDWARE LAYER                                                      │
│  HackRF One ──────────────────── USB ──────────────── Desktop PC   │
│  (1MHz–6GHz · 20Msps · half-duplex)      AMD ROCm iGPU available   │
└─────────────────────┬───────────────────────────────────────────────┘
                      │ IQ samples (int8 → complex64)
                      ▼
┌─────────────────────────────────────────────────────────────────────┐
│  SCANNER (hackrf_scanner_v9.py · PyQt5 GUI)                         │
│  Engine thread: FFT → spectrum/waterfall → demodulate (WFM/NFM/AM) │
│  AudioOut: 400ms ring buffer → sounddevice → speakers               │
│  VAD + squelch → auto-record WAV → sentinel file created           │
└─────────────────────┬───────────────────────────────────────────────┘
                      │ WAV + .inprogress sentinel
                      ▼
┌─────────────────────────────────────────────────────────────────────┐
│  TRANSCRIBE DAEMON (port 4311)                                       │
│  inotify watches /tmp/rfai-test-audio/ + /tmp/rfai-demo/            │
│  sentinel created → emit recording_started SSE                      │
│  sentinel deleted → queue WAV for processing                        │
│  faster-whisper tiny.en (streaming) → partial_transcript SSE       │
│  faster-whisper small (final) → transcription_complete SSE         │
│  entities.py → callsigns, Q-codes, locations                       │
│  intent_llm.py → Ollama/keyword → intent + urgency                 │
│  → INSERT INTO transmissions (SQLite · WAL mode)                    │
└─────────────────────┬───────────────────────────────────────────────┘
                      │ SSE proxy + DB queries
                      ▼
┌─────────────────────────────────────────────────────────────────────┐
│  WEB DASHBOARD (FastAPI · port 4310)                                 │
│  GET /api/transmissions  GET /api/stats  GET /api/operators         │
│  GET /api/events  ← SSE proxy from daemon + DB poll                │
│  GET /api/clusters  ← fingerprint.py clusters                      │
│  POST /api/demo/mic-record  POST /api/transmit/text                │
│  POST /api/transmit/audio   GET /api/daemon/health                 │
│  Serves ui/build/ SPA at / (catch-all → index.html)                │
└─────────────────────┬───────────────────────────────────────────────┘
                      │ HTTP/SSE · CORS
                      ▼
┌─────────────────────────────────────────────────────────────────────┐
│  BROWSER UI (SvelteKit SPA · port 4312 / 4310)                      │
│  /dashboard  — live feed · stats · band chart · MIC DEMO            │
│  /log        — searchable table · flag · notes · export             │
│  /chat       — IRC-style · live bubble · transmit panel             │
│  /scanner    — waterfall · VFO · hardware config                    │
│  /analysis   — stats · model bakeoff · intent arena                 │
│  /fingerprint — cluster browser · identity tiers · label            │
│  /about      — this page                                            │
└─────────────────────────────────────────────────────────────────────┘

Legal & compliance

✅

Receiving is legal in Australia. The Radiocommunications Act 1992 permits reception of public radio communications (CB, amateur, aviation, marine) for personal use.

✅

All AI runs locally. No audio, transcripts or metadata leave your machine. Whisper, Kokoro, and Ollama are all local models.

⚠️

Transmitting requires a licence. Australian UHF CB (477 MHz) allows licence-free operation ≤5W on designated channels. Amateur bands require an ACMA amateur licence.

⚠️

Private communications — do not intercept. Cellular, DECT, encrypted and private services are not targeted. Reception of private communications is prohibited.

This system is designed for lawful monitoring of public radio traffic for personal, research and educational purposes only. Reception only. Personal use. Australian Radiocommunications Act 1992.

Every transmission. Understood.