An early 20th amplitude percentile label reader had lower mean error than ridge regression on 100 later SEEG speech records from nine people at 400 ms

We tested whether amplitude labels for coefficient fitting improved a sound amplitude reader beyond low amplitude labels. The source was public OpenNeuro ds004703, version 1.1.0. The test used 100 later speech records from nine people. Both readers used the same eight bipolar SEEG pairs and 120 inputs. The primary output delay was 400 ms. The low state used each person's early 20th amplitude percentile. Mean MSE skill was 13.848% for ridge amplitude regression and 14.865% for the low label reader. The mean difference was -1.017 percentage points. A descriptive 95% person interval was -5.165 to 3.505 points. This does not establish equivalence or population superiority. At the early 10th percentile and 400 ms, the sign reversed. The amplitude reader had 6.363 points more mean skill. Thus, the state threshold matters. Mean amplitude correlation fell from 0.343 across all target times to 0.158 within louder intervals. Residual tracking remained after state means were removed. The low label reader used binary labels for coefficient fitting. Contact selection, penalty selection, and output calibration still used early amplitude targets. The test records had been scored before. This is an exploratory evaluation control, not an independent replication or a test of speech intent. The package preserves 154 feature records, 48 sound target files, all predictions, four figures, code, hashes, and replay instructions. Fresh pinned source ranges reconstructed every input record and sound target. An offline replay and an independent optimizer check passed.

SEEG · speech amplitude · brain-computer interface · ridge regression · low amplitude labels