Skip to main content

Training from Scratch

Target release

v0.5 ships the rfwhisper train CLI end-to-end. The steps below describe the intended recipe; pieces marked (v0.5) are not yet stable. Help shape them — see Contributing.

Most contributors should fine-tune an existing model on their own noise. Training from scratch is only worth it if you're:

  • Researching architectural variants
  • Targeting a very different mode family (e.g. digital voice, RTTY-only)
  • Building a quantized tiny model for Pi Zero 2 W

Dataset composition​

A good ham denoiser dataset has three components:

  1. Clean speech — studio-quality, license-clean (LibriSpeech + VCTK + CC-BY community ham-speech contributions).
  2. Isolated ham noise — recorded off-air with a dummy load or during band silence:
    • Powerline / corona buzz
    • Switch-mode + inverter hash
    • PLC / VDSL combs
    • Plasma / LED drivers
    • Atmospheric QRN (curated from seasonal recordings)
    • Birdies / narrow carriers
  3. Room / rig impulse responses — so the synthesis reflects a typical mic + shack chain.

See CONTRIBUTING.md § Submitting RF/Audio Samples for the metadata schema we use.

Synthesis​

# (v0.5) mix clean speech with ham noise at controlled SNRs
rfwhisper train build-dataset \
  --clean  data/clean/ \
  --noise  data/noise/ \
  --out    data/train/ \
  --snr-range -10 20 \
  --samples 200000 \
  --rate 48000

Per-sample metadata is written alongside so we can stratify evaluation by noise class.

Architecture​

RFWhisper uses the stock DeepFilterNet3 architecture with minor modifications:

  • Input: 48 kHz mono, 10 ms frames, log-power spectrogram + phase for deep-filter stage
  • Stage 1 (ERB): coarse magnitude enhancement over ERB-scaled bands
  • Stage 2 (Deep Filter): fine-grained complex-valued deep filtering for phase-aware refinement
  • Output: denoised 48 kHz mono audio

Reference: DeepFilterNet3 paper.

Train​

# (v0.5)
rfwhisper train fit \
  --arch deepfilternet3 \
  --data data/train/ \
  --val  data/val/ \
  --epochs 200 \
  --batch 32 \
  --lr 5e-4 \
  --amp bf16 \
  --out runs/dfn3-ham-v1/

Checkpoints, TensorBoard logs, and model cards are written under runs/.

Evaluate​

rfwhisper train evaluate \
  --ckpt runs/dfn3-ham-v1/best.pt \
  --eval data/eval/

This runs:

  • PESQ / STOI / SI-SDR on the eval set
  • Ham-specific effective SNR gain
  • CW transient gate
  • FT8 decode regression gate
  • RTF + latency on the reference hardware of the runner

Any gate failure here is a stop condition.

Export​

rfwhisper train export \
  --ckpt runs/dfn3-ham-v1/best.pt \
  --out  models/dfn3-ham-v2/ \
  --variants fp32 int8 \
  --opset 17 \
  --simplify

Produces:

  • model.fp32.onnx + sha256
  • model.int8.onnx + sha256 (QDQ calibration using the eval set)
  • parity.json — PyTorch vs ORT diff (RMS ≤ 1e-3, max ≤ 1e-2)
  • card.md — auto-filled model card (you'll still want to hand-edit the "known failure modes" section)

Publish​

Open a PR with:

  • The exported .onnx and .json pinned via Git LFS
  • An entry in rfwhisper/models/registry.yaml
  • The model card under docs/models/model-cards/
  • Eval numbers in the PR description

See Model Cards for the template.