Amateur radio · open source
RFWhisper pulls weak voice and data modes out of modern QRM and QRN with ham-tuned deep learning — on your machine, in real time, under a testable latency budget. No cloud, no “send my audio to a server,” just better copy for SSB, CW, FT8, and whatever the bands throw at you.
Not a product chart — a sketch. Push the broad mess down without smearing the parts of the passband you care about. Real metrics live in the test guide.
No cloud inference. No accounts. Audio stays on the machine in front of you — laptop, field box, or Raspberry Pi.
DeepFilterNet3 as the primary; RNNoise when you need something tiny. Trained and fine-tuned on real QRM, not just speech corpora.
Roadmap criteria you can run: FT8 non-regressions, CW transient checks, end-to-end latency. If it does not help on-air, it does not ship.
Universal hardware via SoapySDR; v0.2 brings flowgraphs and gr-dnn ONNX. Bring whatever rig or dongle the bands give you.