Brain‐to‐text

Abstract

The need to improve communication for individuals with severe physical and verbal impairments is critical. This thesis addresses the development and evaluation of a brain‐computer interface (BCI) system based on electroencephalography (EEG) that translates human ideas into text. Our main goal is to improve the lives of people and help them achieve independence. By using a carefully selected vocabulary, the research focuses on integrating cutting‐edge machine learning models and feature extraction techniques to efficiently process and interpret EEG data. At the core of the system architecture lies a specifically designed Transformer model, which handles the complex sequential patterns of EEG signals. The study investigates feature selection techniques and measures the effectiveness of the Transformer model in a variety of data scenarios, including settings with mixed and individual participants. Important assessment criteria like test loss, F1‐score, recall, accuracy, and precision are rigorously observed. With an accuracy of 0.9, the Transformer model is highly capable and effectively captures the subtle dynamics required for accurate text translation. However, when tested with data from another participant, whose brain waves were not included in the training set, the results suggest difficulties in generalising the model to various people, indicating that individual calibration may be required in real‐world scenarios. This was identified by the harsh decline in accuracy to 0.31. Ultimately, this research significantly contributes to the field of neurotechnology by pushing the limits of existing EEG technology and machine learning approaches, thus opening up new avenues for enhancing the quality of life for people with communication impairments. The study’s practical implications include possible uses in assistive communication devices, which give people with disabilities greater autonomy. Future studies may concentrate on making the model more broadly applicable and minimising the requirement for individual calibration.