Quick Start

From git clone to a before/after report in a few minutes. Everything here is reproducible from a seed — no sample pack, no model download.

Alpha status — read this first

RFWhisper is early alpha. The default deepfilternet3 model falls back to a classical spectral-subtraction stub (~+1 dB) unless you build with the dfn feature (below). The stub is for wiring and CI; the real neural denoiser is opt-in while it stabilises. See the project status.

1. Build

git clone https://github.com/JakenHerman/RFWhisper.git
cd RFWhisper
cargo build --release          # single binary at target/release/rfwhisper

2. Make a noisy signal

No downloads: the fixtures are generated from a seed, so these commands produce byte-identical audio on any machine.

rfwhisper samples list          # signals, noise types, and named presets

# An S3 SSB signal buried under an S7 powerline buzz (-24 dB SNR).
# --out is the noisy mix; --clean-out is the reference it was mixed from.
rfwhisper samples synth --kind mix --preset ssb_powerline_s3_s7 \
  --out ssb.wav --clean-out ssb.clean.wav

3. Denoise it, with a before/after report

rfwhisper denoise \
  --input       ssb.wav \
  --output      cleaned.wav \
  --reference   ssb.clean.wav \
  --model       spectral_stub \
  --report      report.json \
  --spectrogram report.html

report.json is the machine-readable summary — these are the real fields:

{
  "model": "spectral_stub",
  "input": "ssb.wav",
  "output": "cleaned.wav",
  "sr": 48000,
  "duration_s": 5.0,
  "inference_time_ms": 7.03,
  "rtf": 0.0014,
  "snr_gain_db": 0.41,
  "spectrogram_path": "report.html"
}

report.html is a self-contained before/after page — SNR tiles, side-by-side spectrograms on a shared scale, and a median-spectrum chart. Open it in any browser, offline. Listen to ssb.clean.wav, ssb.wav, and cleaned.wav back to back too.

Why the gain is small here

spectral_stub is a placeholder that measures around +0.4 dB on this preset. That is expected — it exists to exercise the pipeline, not to clean up your audio. Switch on the real model next.

4. Switch on the real DeepFilterNet3 model

The neural backend is opt-in because it pulls a larger dependency tree:

cargo build --release --features dfn      # ~4 min the first time (pulls tract)

rfwhisper denoise -i ssb.wav -o cleaned_dfn.wav \
  --model deepfilternet3 --spectrogram report_dfn.html

The DeepFilterNet3 model is bundled in the binary — no download. It runs at about RTF 0.017 (≈ 60× faster than real time) on a modern laptop CPU.

DeepFilterNet3 is a speech model

It denoises real speech well, but it correctly removes the synthetic tone-stack signal above as non-speech — so on a samples synth clip its measured "gain" is negative. To see it shine, feed it a real voice recording (see samples/). This is why the A1 SNR-gain gate needs real audio, not synthetic fixtures.

5. Real-time from rig → virtual cable (implemented, not yet gate-verified)

rfwhisper audio list            # find your device indices

# Pipe input device 3 → denoiser → output device 5 (a virtual cable)
rfwhisper denoise-live --in 3 --out 5 --model spectral_stub --blocksize 480

Point WSJT-X / fldigi / JS8Call at the virtual cable as their input. See the virtual-cable setup for VB-Cable / BlackHole / loopback.

What exists today

Command	Status
samples synth / samples list	✅ works
denoise (offline WAV → WAV)	✅ works
denoise --spectrogram (before/after HTML)	✅ works
--model deepfilternet3 (with --features dfn)	✅ works
denoise-live (realtime)	⚙️ implemented, not yet gate-verified
gui	🚧 prints a roadmap pointer; native GUI is v0.4

Next: prove it properly

Want to measure it rather than eyeball it? See the v0.1 Test Guide for the acceptance gates and how to run them.