Genetic algorithm-optimized machine learning approaches for EEG-based silent speech decoding

The phases of human communication consist of speech perception, production, and imagination. The objective of this work is to understand and analyse the changes that occur in the neural signals during the hearing phase by examining electroencephalogram (EEG) patterns of the subject for different sentences. We propose optimising the decoding process using Genetic Algorithms (GA). Six different experiments are performed on Dataset 3 of coSpeech EEG Database. Both handcrafted features and CNN-based features are used for classification. GA is used for two purposes - channel selection as well as feature selection. Two classifiers - decision trees and SVMs are used for sentence classification. A benchmark accuracy of 41.92% is obtained using the proposed methods. Accuracy improves in the alpha, beta and gamma frequency sub-bands (41.79%, 40.92%, 40.27% respectively). Channel selection using GA reduces the computational load significantly (∼ 90%) while producing comparable results (34.37%, 33.20%, 32.93% in the alpha, beta and gamma sub-bands). This work highlights that EEG is a viable, non-invasive way to decode speech from subjects, which would help people with speech disorders communicate in a better way without exertion. Silent speech decoding has applications in assisting speech-impaired individuals, ensuring private communication, and enhancing human-computer interaction.