A fixed SEEG reader gave no positive outputs from 258 inputs with no annotated sound and 94 at quiet decision times in one later listening test

We tested a fixed speech reader on continuous public SEEG data from two listening records. The planned W3 test reproduced 84 correct decisions from 84 selected windows. The same reader gave zero positive outputs from 258 half-second inputs with no annotated sound at 0.1 s steps. It gave negative outputs from 46 of 649 full speech windows. A post hoc audit found 94 positive outputs at 467 decision times outside annotated sound. All 94 inputs still overlapped sound. In W2, a threshold from separate earlier continuous silence reduced final silent positives from 6/104 to 3/104. All 17 final sentences still had a positive output by 0.75 s. The fraction of positive speech windows fell from 71.8% to 64.1%. These results describe short clinical records during passive listening. They do not establish intended speech, artifact removal, or safe stimulation control. The package preserves source hashes, retrieval code, spectra, models, scores, figures, and independent checks.

stereo-EEG · speech decoding · continuous evaluation · brain-computer interface · silent intervals · decision timing