Problems found in studies that read speech from brain signals
Researchers found that muscle signals may be mistaken for brain activity in speech decoding studies.
Studies using scalp electroencephalography (EEG) to decode imagined speech often fail to report how large an effect they could detect. These studies often use overt speech as a positive control to validate their pipelines.
Researchers audited decoding of overt speech, articulator preparation, and imagined speech. They used datasets named ds003626, KaraOne, and Track 3 of the 2020 International Brain-Computer Interface Competition. They injected class-dependent effects of known amplitude into real recordings to measure the detection floor of five decoder families.
Scalp electroencephalography (EEG) is a method used to decode speech. A positive control is a way to validate a pipeline. A detection floor is a measure of the smallest effect a decoder can find.
In ds003626, two lip electrodes decoded overt speech at 0.7430. The eight external channels reached 0.7068, while the 128-channel scalp montage reached 0.4231. No decodable signal survived in the 8-30 Hz band.
In the other two corpora, accuracy rose into the range dominated by muscle activity. The overt-speech positive control is dominated by myogenic signal. This signal cannot validate a decoding pipeline.
The ds003626 condition codes are the reverse of the intuitive reading. The class encoding changed between public releases. Condition is a deterministic function of time within each session.
Key facts
- Overt-speech positive controls are dominated by myogenic signal.
- In ds003626, no decodable signal survived in the 8-30 Hz band.
- Accuracy in two corpora rose into the range dominated by muscle activity.