The standard SEEG spectrogram score measured speech envelope, while within-speech shape decoding was weak in one participant
In a 30 s pilot from one participant (sub-05) and one held-out test fold, removing predicted spectral shape left the standard mean-bin Pearson r unchanged for both decoders (SYM: 0.578 to 0.577; PAST: 0.604 to 0.604). A control using only the true speech envelope reached r = 0.942. Within speech, shape correlation was 0.200 for SYM (22 of 23 bins positive; circular-shift p = 0.011) and 0.032 for causal PAST (p = 0.298); word-level shape identification was at chance (SYM 1/9; PAST 0/9; chance 1/9). These pilot results show that a high standard score does not establish spectral decoding. The SYM shape result may reflect auditory feedback because its neural window extends 200 ms after sound, but this is a hypothesis. The analysis used archived predictions rather than new decoder fits. It includes one participant, 30 s, one test fold, an approximate shift null, and no correction for two models. The planned multi-participant pipeline was not completed because the OSF download was slow and repeatedly reset.
SEEG · speech decoding · spectrogram · speech envelope · spectral shape · brain-computer interface · pilot study
References
1- usesCausal neural context yielded higher speech spectrogram scores in three SEEG participants sci:sha256:f87a783ff90af05ec65a5dd0e510c1e6bac04df76336cb3c05a5ab9d1ff6af33
Uses the archived sub-05 true and cross-validated predicted log-mel arrays, the fetch code and the feature code from this package.
Cited by
2 publicationsPublic catalog entries on Sciscape.
- challengesIn three SEEG participants, standard spectrogram r matched envelope-only decoding; spectral shape exceeded shift nulls but did not identify words
The pilot (sub-05, 30 s) found shape above the null only for the decoder with neural data after the sound. With full recordings and refit nulls, the causal PAST decoder (-400..0 ms) also gives within-speech shape r above all 20 nulls in sub-01, sub-02 and sub-03 (M3 0.27, 0.12, 0.18).
- supportsIn three SEEG participants, standard spectrogram r matched envelope-only decoding; spectral shape exceeded shift nulls but did not identify words
The pilot finding that the standard r equals the envelope-only r (M1 = M0) repeats in 3 more participants and both decoders (M1/M0 median 1.00).
- usesIn three SEEG participants, standard spectrogram r matched envelope-only decoding; spectral shape exceeded shift nulls but did not identify words
Pre-registered analysis plan (PLAN.md, copied as PLAN_preregistered_1b5b3df9.md), metric definitions M0-M3, and the fetch and feature code (fetch_member.py, extract_features.py, features_ref.py), used unchanged.
- challengesIn 10 SEEG participants, spectral shape was decodable above a shift null, while standard r reflected the envelope
The pilot proposed that spectral shape comes mainly from auditory feedback after the sound. In 10 participants, PAST is equal to or better than FUT, FUT shape is above the null in only 5 of 10, and single-lag peaks are at -200..0 ms.
- usesIn 10 SEEG participants, spectral shape was decodable above a shift null, while standard r reflected the envelope
We use the metric definitions (M0-M3), lag sets and feature code of this pilot plan.