Silent EEG-Speech Recognition Using Convolutional and Recurrent Neural Network with 85% Accuracy of 9 Words...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Silent speech recognition using electroencephalography (EEG) is crucial for advancing brain-computer interface (BCI) development.
• Existing BCI technologies often exclude individuals with neurodegenerative diseases or communication impairments.

Purpose of the Study:

• To investigate the feasibility of silent speech recognition in EEG data for BCI applications.
• To develop accurate EEG-based classifiers for silent word commands.
• To assess the transferability of learned models to new individuals and the general population.

Main Methods:

• Recorded EEG data from 270 healthy subjects during silent speech of nine Russian words (commands and a pseudoword).
• Utilized machine learning models to classify silent speech patterns from EEG signals.
• Evaluated classifier performance on individual subjects and assessed cross-subject generalizability.

Main Results:

• Demonstrated statistical similarities in brain activity patterns for certain silent words, particularly those indicating directed movement.
• Achieved 85% accuracy for classifying nine silent words and 88% for binary classification after individual training.
• Showed that subject-specific datasets yield more accurate classifiers than larger, group-collected datasets.
• Confirmed the transferability of learning outcomes from limited EEG data samples to the general population.

Conclusions:

• Silent speech recognition via EEG is a viable method for BCI development.
• Individualized EEG-based silent speech recognition models can achieve high accuracy.
• Transferable learning from EEG data enables the creation of BCI input devices for individuals with communication disabilities, even without prior training on them.