Relative voice pitch did not improve held-out SEEG high-gamma prediction beyond envelope and voicing in six listeners

We tested whether relative voice pitch predicts bipolar SEEG high-gamma amplitude (70–150 Hz) during passive listening to natural speech beyond speech envelope and voicing. The analysis used 8 subjects and the first 12 passages per subject, with a plan written before fitting models. A YIN-type tracker estimated F0 at 32 Hz; relative pitch was log2 F0, z-scored within each passage and represented by five passage-quintile bins. Leave-one-passage-out ridge TRF models compared envelope plus voicing (BASE), envelope plus pitch bins (PITCH), and a shuffled-pitch control (SHUF), using lags of 0–375 ms and channel-specific regularization selected by 3-fold inner cross-validation. Six subjects had responsive channels (r(BASE) ≥ 0.10; 39 channels total); two subjects had none and were excluded from the primary endpoint. Mean r(PITCH)−r(BASE) was −0.0052 across the six subjects (all six negative; exact one-sided sign-flip p = 1.000; two-sided t 95% CI −0.0085 to −0.0019). Mean r(BASE) was 0.209 and mean r(PITCH) was 0.204; PITCH was higher on 6 of 39 responsive channels. PITCH exceeded SHUF in 4 of 6 subjects (mean +0.0018, p = 0.094; 95% CI −0.0007 to +0.0043), which was not significant. None of the 39 responsive channels had p < 0.05 for PITCH > BASE. Among 1,015 non-responsive channels, 79 (7.8%) had p < 0.05; 63 of these had negative r(BASE), and this excess was interpreted exploratorily as ridge behavior on channels that do not track speech. On synthetic harmonic F0 glides, the tracker’s median error was 1.3–2.1 cents with 0% octave errors. The results show that this relative-pitch code added no held-out prediction beyond envelope and voicing in these data; they do not show that SEEG has no pitch information. Limits include 8 subjects (6 with responsive channels), passive listening, no electrode anatomy, selection by BASE, one feature code and no correction for multiple secondary tests. The upper confidence limit for PITCH−SHUF was +0.004 r, about 2% of mean r(BASE). External recordings, audio, event tables and source-derived input caches are not included; source files and terms are described in the main text.

voice pitch · SEEG · high gamma · speech prosody · ridge TRF · passive listening