BCIwiki (bciwiki.com) — According to a Journal of Neural Engineering record dated 2026-10-20, researchers affiliated with Araya Inc. and Sincxpress Education SRL examined neural contributions to EEG-based speech decoding. The work addresses a difficulty in non-invasive speech brain-computer interfaces (BCIs): EEG recorded during speaking can contain both speech-related neural activity and myoelectric artifacts generated by articulation.
The researchers used explainable AI to examine a convolutional neural network predicting spoken words from ultra-high-density EEG. They assessed which electrodes and time periods contributed to predictions and compared these contributions with EMG signals. Electrode-wise decoding contributions could not be explained by their mutual information with EMG, a measure of statistical association. Periods contributing to EEG decoding also differed from those contributing to decoding based solely on EMG.
Contributing periods partially overlapped between overt and minimally overt speech. Across all speech conditions, denoising increased decoding contributions from electrodes in speech-related brain areas. The reported temporal overlap concerns minimally overt speech; it does not establish that overt and covert speech share the same decoding mechanisms.
The authors interpret the findings as supporting the use of appropriately preprocessed EEG to decode spoken words from neural activity. This interpretation concerns signal origins and decoding mechanisms, rather than demonstrating complete removal of myoelectric artifacts or clinical benefit for people with speech impairments.