No consistent pitch-shifted voice leak was detected in four of ten Dutch SEEG participants

Speech brain-computer interface decoders could use acoustically leaked voice instead of neural activity. The public SingleWordProductionDutch-iBIDS recordings use audio pitch-shifted by 1–3 semitones, so the diagonal contamination test used in the dataset paper may miss a leak. We tested for correlations along audio-to-SEEG frequency ridges from −4 to +4 semitones, using 200 permutations of the SEEG frequency axis and the maximum across shifts. We processed 351 contacts from four of ten participants. The diagonal result reproduced the dataset paper: zero or one contact per participant had p < 0.01. The shift search found more contacts at p < 0.05 than expected under 5% in sub-04 (35/115; binomial p = 2×10⁻¹⁸) and sub-05 (16/60; p = 3×10⁻⁸), but not sub-08 (2/54; p = 0.76) or sub-10 (8/122; p = 0.27). Only 23% and 50% of the flagged contacts in sub-04 and sub-05, respectively, peaked in the expected 1–3 semitone band. Maximum ridge contrasts were small (0.015–0.020 correlation above the matrix mean). The results therefore show no consistent evidence of a pitch-shifted voice leak, although smooth spectral structure may inflate the shift-search statistic. Only four participants were processed; 200 permutations give a p-value floor of 0.005. The frequency-permutation null may be anti-conservative, and the analysis used whole-session correlation without a lag search or a test of the effect on decoding.

acoustic contamination · speech brain-computer interfaces · SEEG · stereo-electroencephalography · pitch shifting · signal leakage · spectral correlation · Dutch word production