End-to-End RF-Inferred Inner Speech Decoding: FFT Triage to Bayesian Command Reconstruction

We present a real-time RF-to-speech pipeline that
decodes inner speech from RF-inferred neural surrogates using
a word-state HMM with GPT-style priors. Starting from 1.5
ms FFT triage (0.754 AUROC), we map spectral confidence
to link quality q ∈ [0, 1], which predicts command success
(73.9% → 100%) and p95 latency (2.6 s → 375 ms). A
Bayesian decoder with language priors reduces WER from 2.8%
to 1.1% at 10 dB SNR (60.7% relative reduction), with posterior
concentration on correct word spans. The system integrates with
tactical control systems for hands-free command execution. Full
end-to-end reproducibility: make all generates IQ → WER.