Causal neural context yielded higher speech spectrogram scores in three SEEG participants

We tested whether linear reconstruction of a speech spectrogram from stereoelectroencephalography (SEEG) high-gamma activity requires neural frames after the corresponding sound. Such frames can contain auditory feedback and are unavailable to an online speech brain-computer interface. Using the SingleWordProductionDutch-iBIDS dataset, we compared a symmetric neural context (−200 to +200 ms) with a causal context (−400 to 0 ms) in three participants. The pipeline used a 70–170 Hz Hilbert envelope, 50 ms windows shifted by 10 ms, 23 log-mel audio bins, PCA with 50 components, least-squares regression, 10 contiguous folds, and mean Pearson correlation across bins and folds. The causal window scored higher in all three participants: 0.802 versus 0.776 for sub-04, 0.568 versus 0.537 for sub-05, and 0.727 versus 0.705 for sub-08. In sub-05, the 0 to +400 ms window scored 0.414, the −450 to −50 ms window scored 0.558, and the 0 ms frame alone scored 0.370. These descriptive results suggest that, in this small sample and with this linear decoder, most decodable information was present before or at the sound rather than after it. There was no group significance test (n=3). Audio was pitch-shifted in the dataset, so the target was its shifted log-mel spectrogram. Chance shifts, a single-lag scan, and per-contact cross-correlations were included in the code but did not finish in this run.

speech production · stereoelectroencephalography · high-gamma activity · neural decoding · speech spectrogram · brain-computer interface · causal context · log-mel spectrogram