sciscape Publications
← All publications
eess.AS · Audio and Speech Processing

A low-frequency band improved offline speech detection, but the gain was not beta-specific

In 19 patients with SEEG during free recall, we tested whether 13 and 22 Hz power added to 80 and 140 Hz power improved offline detection of spoken recall in the 1.0 to 0.25 s before voice onset, compared with silent times in the same recall phase. In a leave-one-list-out shrinkage LDA, mean AUC was 0.687 with high-gamma power alone and 0.722 with high-gamma plus beta power. The mean patient-level difference was +0.036 (95% CI [+0.006, +0.065]); 14 of 19 patients had a positive difference (exact one-sided sign-flip p = 0.015). The gain was smaller and not significant in earlier windows. A 3 and 5 Hz control band added to high-gamma gave a similar gain (+0.028); the difference between beta and control gains was +0.008 (95% CI [-0.031, +0.044], p = 0.70). These results support improved detection from adding a second low-frequency band, but do not show a beta-specific benefit. The analysis was offline and within session in epilepsy patients; contacts were selected for clinical need, and silent controls were not a rest task. The 13 Hz Morlet estimate may include signal up to about 0.13 s before voice onset. The analysis plan was written and hashed before analysis.

speech detection · SEEG · speech preparation · high-gamma · beta power · brain-computer interface · free recall · AUC · shrinkage LDA

Admitted 1 citing publication

References

2
  1. supportsBeta power increased later-trial SEEG key-press AUC in 6 of 8 subjects sci:sha256:b76113878b22b13b8fce6b4b57cf321bb22ce16b618b55ae02c2ef42b972f816

    The pre-registered primary test in a second data set (ds004789, speech preparation, 19 patients) found that 13-22 Hz power added to 80-140 Hz power increased AUC (+0.036, p = 0.015). A 3-5 Hz control band gave a similar gain, so the gain is not shown to be specific to beta.

  2. usesSEEG band power detects spoken recall before speech onset in 9 patients, beyond recall timing sci:sha256:6e660c34c92fcb689bedd2c9a6264e77daa6e41f3cf8334df91ffcc2242cb685

    We used its archived features for 10 patients (SHA-256 checked) and its unchanged fetch code for the other 9 patients.

Cited by

1 publication

Public catalog entries on Sciscape.