Decoding Imagined Speech of Daily Use Words from EEG Signals Using Binary Classification

The simplest form of communication between people is done through speech. However, there are situations in which this communication is not possible, hence, there is great interest in decoding imagined speech. To address this problem, this work proposes a method for recognizing imagined speech from electroencephalographic signals applying artificial intelligence models, focusing on covering daily use words. We used the OpenBCI system with a reduced number of channels using dry electrodes and machine learning methods. Mean, variance, skewness, RMS and kurtosis were extracted as characteristics for each channel of the EEG. A Decision Tree and a Support Vector Machine were tested to classify the words via a one-vs-rest approach. The SVM gave the best results for the task of imagined speech classification, with an accuracy of 92.5%, 76.3%, 70% and 70% for “Descansar", “Baño”, “Comida” and “Okay" respectively. The Decision Tree presents lower accuracy results, but allows us to identify the decision criteria used for each classification.