The Effects of Classification Method and Electrode Configuration on EEG-based Silent Speech Classification.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Speech-based Brain-Computer Interfaces (BCIs) require effective classification of imagined and intended speech.
• Distinguishing between these two speech types is crucial for developing advanced assistive communication technologies.
• Cortical electroencephalography (EEG) offers a non-invasive method for capturing brain activity related to speech production.

Purpose of the Study:

• To effectively classify imagined speech versus intended speech using scalp-recorded EEG signals.
• To evaluate and compare different EEG feature classification methods for speech decoding.
• To explore the potential for enhancing silent speech interface capabilities.

Main Methods:

• EEG signals were recorded from eleven subjects producing Mandarin-Chinese monosyllables in both imagined and intended speech conditions.
• EEG features were extracted and classified using Common Spatial Pattern (CSP), time-domain, frequency-domain, and Riemannian manifold-based methods.
• Classification accuracy was assessed using both whole-brain and left-hemisphere electrode configurations.

Main Results:

• The Riemannian manifold-based method achieved the highest classification accuracy, reaching 85.9%.
• Classification performance with a left-hemisphere electrode configuration was comparable to that of the whole-brain configuration.
• This indicates that focused electrode placement can yield significant results.

Conclusions:

• The Riemannian manifold method demonstrates superior performance for classifying imagined versus intended speech from EEG.
• Effective classification is achievable even with a reduced set of electrodes (left-hemisphere only).
• These findings hold promise for expanding the command capabilities of silent speech interfaces and BCIs.