Vowel Imagery Decoding toward Silent Speech BCI Using Extreme Learning Machine with Electroencephalogram.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Silent speech Brain-Computer Interfaces (BCI) aim to decode intended speech without vocalization.
• Classifying electroencephalography (EEG) signals from imagined speech is crucial for developing effective silent speech BCIs.
• Single-trial classification presents a significant challenge due to inherent signal variability.

Purpose of the Study:

• To classify electroencephalography (EEG) data from imagined speech within a single trial.
• To evaluate the performance of different machine learning algorithms for this classification task.
• To explore the potential of EEG-based imagined speech classification for future BCI applications.

Main Methods:

• Recorded EEG data from five subjects imagining distinct vowels (/a/, /e/, /i/, /o/, /u/).
• Segmented single-trial EEG data and extracted statistical features (mean, variance, standard deviation, skewness).
• Applied sparse regression for feature selection and classified features using Support Vector Machines (SVM) and Extreme Learning Machines (ELM) with different kernels.

Main Results:

• Extreme Learning Machines (ELM) and its variants demonstrated superior classification accuracy compared to Support Vector Machines (SVM) with a radial basis function kernel and Linear Discriminant Analysis (LDA).
• The proposed method successfully classified imagined speech in single EEG trials.
• A majority voting scheme across thirty segments per trial determined the final classification label.

Conclusions:

• EEG responses to imagined speech can be successfully classified in single trials.
• Extreme Learning Machines, particularly with radial basis function and linear kernels, show significant promise for imagined speech classification.
• This research contributes to the advancement of silent speech Brain-Computer Interface (BCI) systems.