Speech mode classification from electrocorticography: transfer between electrodes and participants

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Speech brain-computer interfaces (BCIs) aim to restore communication for individuals with speech loss.
• Speech detectors are crucial for BCIs to differentiate speech intent from silence.
• Existing detectors must account for non-speaking language-related brain activity like reading or listening.

Purpose of the Study:

• To analyze brain activity across various speech modes: speaking, listening, imagining speaking, reading, and mouthing.
• To develop and evaluate a speech mode classifier using electrocorticography (ECoG) data.
• To assess the transferability of trained classifiers across participants for single- and multi-electrode configurations.

Main Methods:

• Collected ECoG data from 29 participants performing different speech-related tasks.
• Developed linear classifiers for speech mode detection.
• Evaluated classification accuracy for single- and multi-electrode setups.
• Assessed cross-participant classifier transferability for binary and multiclass scenarios.

Main Results:

• High classification accuracies achieved: 88.89% for single-electrode and 96.49% for multi-electrode classifiers distinguishing speaking, listening, and silence.
• Optimal electrode locations identified on the superior temporal gyrus and sensorimotor cortex.
• Single-electrode classifiers demonstrated successful transfer across recording sites.
• Multi-electrode classifiers showed better transferability for binary tasks compared to multiclass tasks.

Conclusions:

• Accurate speech detection is vital for reliable speech BCIs, preventing false outputs and enabling use outside laboratory settings.
• Cross-participant transfer of classifiers is valuable for reducing training time, especially when subject training is difficult.
• The developed speech mode classifiers hold significant potential for advancing practical speech BCI technology.