Exploring Diverse Augmentation Strategies for Imagined Speech Detection in Brain-Computer Interface using Machine Learning Models

Imagined speech refers to the inner articulation of words without vocalization. Augmentation techniques in Electroencephalography (EEG)-imagined speech datasets are essential to compensate for limited data, enhance model generalization, and ensure robust performance. They mitigate overfitting and improve the effectiveness of models trained on these datasets. This work demonstrates the application of diverse data augmentation techniques in imagined speech tasks, where individuals mentally simulate speech without vocalization. The proposed augmentation methods including overlapping, time warping + overlapping, gaussian noise, fourier transform surrogate (FTSurrogate), time reverse, channel shuffle, time warping + jitter and no augmentation methods aim to enhance the robustness and generalization of machine learning models. Leveraging these techniques, several classification experiments using a variety of supervised learning algorithms to distinguish imagined speech patterns are conducted. This work highlights the importance of data augmentation and machine learning (ML) in imagined speech tasks. This work has applications in cognitive neuroscience and assistive rehabilitation technologies. Our multi-layer perceptron (MLP) result has shown promising outcomes. The results are analyzed in terms of accuracy, f1 score and kappa.