Why RFWhisper?
The noise floor isn't what it used to be
If you've tried to work a POTA station on 40m from a townhouse, chased DX through a neighbour's plasma TV, or decoded FT8 next to an EV charger, you know the modern noise floor is unrecognizable compared to a decade ago. The culprits:
| Noise source | Character | Typical footprint |
|---|---|---|
| Switch-mode / LED drivers | Broadband hash, 20–150 kHz spurs | HF wide-band raised floor |
| Solar inverters (MPPT) | Rhythmic buzz | 40m–10m, drifts with sun |
| VDSL / PLC / powerline ethernet | Wideband raised floor, data combs | 1.8–30 MHz |
| Plasma TVs, touch lamps, grow lights | Narrowband carriers | Random on HF |
| Neighbor electric fences, thermostats | Regular click trains | Rural/suburban |
| Atmospheric QRN | Impulsive crashes, seasonal | LF/MF/low-HF |
Traditional DSP tools (analog NB/NR, classical Wiener filters, simple spectral subtraction) either:
- Miss modern complex impulsive + stationary mixtures.
- Wreck the signal — they smear CW keying transients, introduce "underwater" artifacts that confuse FT8/WSJT-X decoders, or chew the atmospheric hiss so aggressively that weak signals disappear under a gated floor.
The RFWhisper approach
Use a deep neural network designed for speech denoising (DeepFilterNet3) and fine-tune it on real amateur-radio noise — recorded powerline buzz, inverter hash, static crashes — so it learns to preserve SSB phonemes, CW keying transients, and FT8 tones while killing QRM.
Then wrap it in a pipeline that:
- Runs locally (CPU or GPU) — no cloud, no telemetry.
- Integrates with GNU Radio + SoapySDR so any SDR or rig's audio works.
- Emits to a virtual cable so WSJT-X / fldigi / JS8Call / your headphones see clean audio transparently.
- Stays under 100 ms p99 latency (target < 50 ms by v0.3).
- Is GPLv3 so nobody can take it closed.
The non-negotiables
Three things RFWhisper will never do, full stop:
- Damage CW keying transients. We run a CI gate that measures RMS energy in the first 5 ms of each dit. A model that drifts outside ±1 dB of raw does not ship.
- Reduce FT8/FT4 decode counts. We replay a 15-minute FT8 cycle against WSJT-X on every candidate model. If it loses even one decode or fabricates one false decode, it's rejected.
- Mandate an internet connection at runtime. Models are fetched once at install (SHA-256 verified), and the pipeline runs offline forever after.
Testable value: four concrete examples
These live in tests/audio/ and run in CI. See ROADMAP.md § Acceptance Criteria for the full list.
| # | Demo | Pass criterion |
|---|---|---|
| 1 | Weak 40m SSB under S7 powerline buzz | ≥ 4/5 blind listeners prefer denoised AND copy +1 word/sentence |
| 2 | 15-min FT8 cycle through WSJT-X | Denoised decodes ≥ raw decodes, zero false decodes |
| 3 | 25 WPM CW under atmospheric QRN | fldigi copy accuracy unchanged, crashes −6 dB audibly |
| 4 | End-to-end latency probe | < 100 ms p99 on i5-8xxx / M1 / RPi 5 |
Who this is for
- Contesters who want to keep decoding weak stations in a crowded band.
- POTA / SOTA operators running from townhouses, condos, or next to a solar farm.
- Weak-signal specialists on 630m / 2m EME / microwave where every dB counts.
- EmComm operators who need a tool that works when the internet doesn't.
- DSP and ML hackers who want a serious, well-documented sandbox for real-time audio neural nets.
- Any ham who has ever shrugged at their S9+10 noise floor and said, "there's nothing I can do about it."
There is something you can do about it. 73