Contesting
Contest weekends are where QRM piles up fast: dozens of stations packed into a few kHz, overloaded front-ends, exchanges barely above the noise floor. RFWhisper is tuned to preserve decode quality under exactly this condition — never at the cost of busting a call.
Hard requirements for contest use​
- Zero decode regressions. If a raw pass decodes
K3XYZand the denoised pass decodesKJXYZ, that's a bug in the model. Run the FT8 regression gate on your chosen model before the contest. - Low latency. No operator wants to hit the TX button two beats behind the other station. Stay under 100 ms p99.
- Bypass toggle on one key. You'll want to A/B instantly when a station comes up.
- Predictable level. Soft limiter catches overshoots so your AGC / noise blanker upstream stays happy.
Recommended profiles​
| Contest type | Profile | Notes |
|---|---|---|
| SSB phone contest (CQWW SSB, Sweepstakes Phone) | ssb | Full aggression OK — speech-centric |
| CW contest (CQWW CW, ARRL DX CW) | cw | Transient preservation is non-negotiable |
| RTTY (RTTY Roundup) | rtty (v0.3) | Careful — don't filter the 170 Hz shift |
| FT contests (FT Roundup) | ft8 / ft4 | Gentler; let the decoder do its job |
Running vs S&P​
- Running (CQ): keep the denoiser on. You're looking for the faintest caller at the bottom of the pile.
- S&P: consider running the denoiser for searching, but when you hear a clear station you want to work, the raw signal already has what you need — the denoiser is break-even to mildly positive here.
Sample contest scenario: 40m SSB at 0200 UTC​
A typical late-night 40m SSB contest:
- Band crowded: 7.150–7.200 wall-to-wall.
- Your neighbour's inverter adds an S5 rhythmic buzz.
- You're trying to copy an exchange at S3.
rfwhisper denoise-live \
--in "USB Audio CODEC" --out "CABLE Input" \
--model deepfilternet3 --profile ssb \
--telemetry ./contest-log.jsonl
contest-log.jsonl records rolling SNR gain, RTF, and latency per minute so you can analyze the session later.
Spotting weak multipliers​
Anecdotally: the biggest single-operator benefit is hearing the call sign prefix under splatter from an adjacent strong station. The NN preserves consonant structure that classical NR tends to smear.
What RFWhisper is not​
- A noise blanker — we don't cancel impulsive transmitter artifacts (keyclicks, splatter). Use your rig's NB for those.
- A passband filter — use your rig's PBT / DSP bandwidth to set the crop; RFWhisper works inside whatever you hand it.
- A replacement for a good antenna. +3–6 dB of SNR gain on a compromise antenna is still less than what a real antenna change does.
After-action reports welcome​
Post your contest results with RFWhisper in the Contesting Discussion. Include:
- Contest + your category
- Band + mode
- Model used + profile + hardware
- Rate improvements / subjective observations
- Any decodes you're confident the raw audio would not have produced
Your real-world data is how we decide what to improve next.