FT8 Clear — AI Denoising for Weak Signals

FT8 Clear is an experimental audio preprocessing layer that sits between the radio and WSJT-CB. The FT8 protocol and decoder remain untouched: audio from the radio is denoised in real time, routed through a virtual audio device and handed to the existing software. Success is measured where it matters—more correct decodes, or fewer false decodes, at the same signal-to-noise ratio.

The model adapts to each station without retraining. It estimates a live noise profile from roughly two seconds of radio silence at the end of each slot and uses that profile to condition a compact FiLM U-Net. The network predicts a time-frequency mask, preserves the original phase and blends part of the untouched signal back into the result. That blend is central to the design: it softens subtle neural artefacts that can confuse the decoder.

The target hardware is a Raspberry Pi 4. To keep deployment conservative, the architecture stays below 500,000 parameters, uses only well-supported convolutional operations and keeps STFT processing outside the model graph. A second model of roughly 150,000 parameters is trained as a fallback. Models are exported to ONNX, quantised to int8 and processed once per 15-second FT8 slot rather than as a continuous stream.

Training data comes primarily from a synthetic transmitter-to-channel-to-receiver simulation, including weak signals, fading, multipath, interference, filter variations and timing drift. A local research cockpit on the M1 Mac combines recording, live waterfalls, sample management, training and A/B testing. The first measured v3 result at −18 dB matched the unprocessed signal’s 70% decode rate while reducing false decodes from 15% to 10% at a blend setting of α 0.7. Raspberry Pi field validation is the next major step.