Recunoașterea vorbirii imaginate prin analiza semnalelor EEG / Imaginary speech recognition by EEG signal analysis

Communication represents an important element for the human relations. Unfortunately, there are cases in which this ability can no longer be performed, or speech cannot be articulated due to conditions such as stroke, lock-down syndrome etc. This doctoral thesis aims to study artificial intelligence systems with the purpose of recognizing words and phonemes from EEG signals in order to synthesize imagined speech. The study used the Kara One database containing phonemes and words acquired during imagined speech, covering a wide range of pronunciation mechanisms. The database was preprocessed by removing the contaminated signals, and two methods of removing ocular artifacts were compared at this stage: the adaptive filter and the removal of ocular components using PCA. During the research, the authors conducted a comparative study on multiple feature extraction methods: (a) linear predictive coefficients, (b) mel-cepstral coefficients, (c) time-domain cross-covariance, and (d) frequency-domain cross-covariance. The extracted features were qualitatively and quantitatively analyzed by projecting two-dimensional maps of the feature space in clusters based on their similarity using the SOM neural network. In order to obtain a system with high performance, the study introduced deep learning classification methods such as: CNN and CNNLSTM. The best results were obtained using in the feature extraction stage the frequency-domain cross-covariance together with the CNNLSTM neural network reaching a maximum accuracy of 43.98% for the classification of eleven classes: seven phonemes and four words.