A decoder with eight SEEG pairs gave -0.012 mean correlation for heard speech amplitude in later passages from ten people after brief training

We tested whether a small SEEG model could recover heard speech amplitude after brief training, using OpenNeuro ds004703 version 1.1.0. Ten people supplied 95 fixed ten-second windows from English passages. Each person trained on early passage names and tested on different, later names. Training lasted 30 to 60 seconds per person. The model selected eight bipolar pairs from training data only and used causal 1 to 8 Hz voltage, trailing 70 to 110 Hz power, and neural delays from 0 to 320 ms. Mean test passage correlation was -0.0123 (descriptive 95% person interval -0.0342 to 0.0097); mean training correlation was 0.3767. Shifted audio gave 0.0043 mean correlation. Correct timing minus the shift mean was -0.0166 (exact two-sided person sign-flip p=0.2090). Every person had more pooled squared error than a constant training-mean baseline. This test did not establish useful amplitude reading with the fixed method, and does not show absent speech information in SEEG. The retained package includes fitted models, predictions, controls, checks, source versions and hashes, retrieval code, and figures. Source-derived signal/audio inputs and source event/channel metadata are omitted because the source README restricts commercial use and participant identification despite CC0 metadata. The exploratory study does not establish intended speech reading, live control, transfer to new people, or a stimulation effect.

SEEG · speech amplitude · neural decoding · ridge regression · held-out passages · reproducibility