Weak-Signal Work
EME, 630m, 2200m, 6m sporadic E, microwave — the modes where you live 3 dB above the noise floor and every photon matters. RFWhisper isn't a miracle, but a ham-tuned DNN that doesn't damage signal structure is a meaningful addition to a weak-signal operator's toolkit.
Why weak-signal benefits most
Classical NR is built around the assumption that noise is stationary and the signal is loud. Weak-signal modes break both assumptions:
- The signal is barely above the noise. Any gating threshold swallows it.
- The noise is often non-stationary — ionospheric propagation flutter, man-made QRM drift, ambient hash.
- Waveforms are sacred — Q65, JT9, JT65, WSPR decoders rely on tone phase/amplitude continuity. A denoiser that artifacts is worse than no denoiser.
RFWhisper's hard gates (CW transient, FT8 decode counts) exist specifically because we designed it against the temptation to over-process. On JT-family modes we default to gentler blend factors and longer frames.
Recommended profiles
| Mode | Profile | Notes |
|---|---|---|
| JT65 (HF / VHF / EME) | ft8 (close enough for v0.1) | v0.3 adds a dedicated jt65 profile |
| Q65 (EME / MS / tropo) | ft8 → jt65 (v0.3) | Same NN, tuned aggressiveness |
| JT9 | ft8 | Narrow, same approach |
| WSPR | ft8 | Extremely narrow; verify decoder via v0.1 regression gate |
| 630m / 2200m CW | cw | Crash reduction helps a lot |
| 2m SSB EME (audio post-demod) | vhf-ssb (v0.3) | — |
EME / meteor scatter
- Feed RFWhisper the rig audio output after demod (not raw IQ).
- Keep the rig's own NR off — RFWhisper replaces it, doesn't augment it.
- WSJT-X sees the virtual cable as a standard input. The non-regression gates guarantee decoder parity.
- Burst modes (MSK144, Q65-F) benefit from longer pre-filtering (use the default profile; don't override frame size manually unless you know what you're doing).
630m / 2200m (LF / MF)
Noise is different here — seasonal atmospheric QRN dominates, not man-made QRM. A CW-tuned profile with aggressive crash reduction helps a lot without harming keying transients.
rfwhisper denoise-live --model deepfilternet3 --profile cw \
--telemetry ./630m-session.jsonl
Microwave / SHF
At microwave frequencies you're typically working narrow modes (CW, SSB, JT4, JT65). Signal chains are long and picky; insert RFWhisper after the demodulator, before the virtual cable feeding your decoder. Same setup as HF.
6m sporadic E
- Band opens fast, closes fast. You don't want to be fiddling with aggressive denoising when a mult rolls by.
- Use the
ssborft8profile, A/B toggle on space bar, and focus on the radio.
Weak-signal contributions we'd love
- EME recordings before/after RFWhisper with known-good reference decodes.
- 630m winter recordings of the seasonal atmospheric QRN — this is hard to synthesize and invaluable for training.
- WSPR long-term spots before/after with sequential WSPRnet logs.
See CONTRIBUTING.md § Submitting RF/Audio Samples.
The operator's honest take
A +3 dB denoiser on a weak-signal station is not going to change what's physically copyable in theory. It changes:
- Operator fatigue on long sessions.
- Decoder margin when the signal is right at the edge.
- Time-to-decide whether a signal is real or imagined.
Those are real, measurable gains. 73