We present a real-time RF-to-speech pipeline thatdecodes inner speech from RF-inferred neural surrogates usinga word-state HMM with GPT-style priors. Starting from 1.5ms FFT triage (0.754 AUROC), we map spectral confidenceto link quality q ∈ [0, 1], which predicts command success(73.9% → 100%) and p95 latency (2.6 s → 375 ms). ABayesian decoder with language priors … Continue reading End-to-End RF-Inferred Inner Speech Decoding: FFT Triage to Bayesian Command Reconstruction