Skip to main content

FAQ

Does RFWhisper send my audio anywhere?​

No. Core runs 100% locally. Models are fetched once (SHA-256 verified) and cached. No telemetry is ever sent without explicit, disclosed opt-in. This is a GPLv3 project — if that ever changes, fork it.

Will it decode encrypted transmissions?​

No. RFWhisper is a noise-reduction tool applied to already-demodulated audio. It does not decrypt, descramble, or otherwise defeat intentional obfuscation. Doing so would likely violate your country's regulations and the project's Code of Conduct.

Is this a replacement for my rig's NR?​

Usually yes — turn your rig's NR off when using RFWhisper, or they'll fight each other. The one exception is classical noise blanking (for keyclicks / impulsive transmitter splatter), which does something different; keep that on if it helps.

Does it work for digital modes (FT8, FT4, PSK31, RTTY)?​

Yes, with the non-regression gates guaranteeing the decoder sees at least as many valid decodes on denoised audio as on raw. Start with the ft8 profile and measure.

What about CW? Won't it soften my dits?​

No. This is the most protected signal type in the project. The cw profile uses shorter frames and lower NN aggressiveness; a CI gate (cw_transient_test) measures RMS in the first 5 ms of every keying onset and refuses to merge any model that drifts outside ±1 dB.

Can I use this on a rig that doesn't have a USB CODEC?​

Yes — any way you get audio into the computer works: line-out into a USB sound card, a SignaLink, a direct audio cable into mic-in. If your rig is SDR-based (ANAN, Flex, KiwiSDR) use the rig's native audio routing into a virtual cable and have RFWhisper read from that.

Does latency matter for FT8?​

Less than you'd think. FT8 is a 15-second mode with wide synchronization tolerance; a 100 ms denoiser latency is invisible to the decoder. For SSB ragchewing and CW where you're reacting to other operators, the under-100-ms budget matters more.

CPU-only or GPU?​

CPU-only is the default and works great on Apple Silicon, any Ryzen 5000+, Intel 8th gen+, and Pi 5. GPU (CUDA / DirectML / CoreML via ONNX Runtime) helps on weaker hardware or when you're running several RFWhisper pipelines in parallel. Select providers with rfwhisper info providers.

Can I run multiple RFWhisper pipelines in parallel?​

Yes — each instance is a separate process. Typical case: HF SSB denoiser on your main rig + a second instance on a 2m FM radio. Just route each to its own virtual cable.

Does it work with [specific SDR]?​

If the SDR has a SoapySDR driver, yes (v0.2+). See Hardware → SDRs for the current matrix. If your SDR isn't listed, please test and file a report.

Can I train my own model on my shack's noise?​

Yes (v0.5). See Fine-Tuning for Your QTH — the recipe is about 30 minutes of work plus training wall time.

What license are community models under?​

Whatever the contributor chooses, as long as it's GPLv3-compatible. Models in the core repo prefer MIT / Apache-2.0 / BSD so downstream users have the maximum freedom.

Can I use RFWhisper commercially?​

Yes — GPLv3 permits commercial use. You must ship source for any modifications you distribute, and you cannot impose additional restrictions downstream. If you sell a product that includes RFWhisper, the user has the right to inspect, modify, and redistribute the version you shipped.

Why GPLv3 and not MIT / Apache?​

Because this is ham-community code and we want it to stay ham-community code. GPLv3 guarantees that any improvements flow back into the public commons. See README § License for the full reasoning.

Will this work on Windows 10 / old Linux / older macOS?​

  • Windows 10 22H2 — officially supported.
  • macOS 13+ — officially supported.
  • Ubuntu 22.04+ / Debian 12 / Fedora 38+ — officially supported.
  • Older systems may work; we can't promise CI coverage.

How do I benchmark my hardware?​

rfwhisper bench report --profile ssb --out report.html

Upload report.html to Discussions → Benchmarks — your numbers help us keep the reference targets honest.

What if I find a bug?​

See Reporting Bugs in CONTRIBUTING.md. Please include rfwhisper doctor output and a short audio clip if the issue is audio-quality related.

How do I contribute?​

See CONTRIBUTING.md. Code, samples, hardware reports, docs fixes, and community support in Discussions are all first-class contributions.

Where do I send QSL cards?​

Via the bureau (or direct with SASE) to the project's honorary call, once we're assigned one. Until then, keep your QSLs to the hams you worked — that's the real exchange. 73