Toward an Imagined Speech-Based Brain Computer Interface Using EEG Signals

Individuals with physical disabilities face difficulties in communication. A number of neuromuscular impairments could limit people from using available communication aids, because such aids require some degree of muscle movement. This makes brain–computer interfaces (BCIs) a potentially promising alternative communication technology for these people. Electroencephalographic (EEG) signals are commonly used in BCI systems to capture non-invasively the neural representations of intended, internal and imagined activities that are not physically or verbally evident. Examples include motor and speech imagery activities. Since 2006, researchers have become increasingly interested in classifying different types of imagined speech from EEG signals. However, the field still has a limited understanding of several issues, including experiment design, stimulus type, training, calibration and the examined features. The main aim of the research in this thesis is to advance automatic recognition of imagined speech using EEG signals by addressing a variety of issues that have not been solved in previous studies. These include (1)improving the discrimination between imagined speech versus non-speech tasks, (2) examining temporal parameters to optimise the recognition of imagined words and (3) providing a new feature extraction framework for improving EEG-based imagined speech recognition by considering temporal information after reducing within-session temporal non-stationarities. For the discrimination of speech versus non-speech, EEG data was collected during the imagination of randomly presented and semantically varying words. The non-speech tasks involved attention to visual stimuli and resting. Time-domain and spatio-spectral features were examined in different time intervals. Above-chance-level classification accuracies were achieved for each word and for groups of words compared to the non-speech tasks. To classify imagined words, EEG data related to the imagination of five words was collected. In addition to words classification, the impacts of experimental parameters on classification accuracy were examined. The optimization of these parameters is important to improve the rate and speed of recognizing unspoken speech in on-line applications. These parameters included using different training sizes, classification algorithms, feature extraction in different time intervals and the use of imagination time length as classification feature. Our extensive results showed that Random Forest classifier with features extracted using Discrete Wavelet Transform from 4 seconds fixed time frame EEG yielded that highest average classification of 87.93% in classification of five imagined words. To minimise within class temporal variations, a novel feature extraction framework based on dynamic time warping (DTW) was developed. Using linear discriminant analysis as the classifier, the proposed framework yielded an average 72.02% accuracy in the classification of imagined speech versus silence and 52.5% accuracy in the classification of five words. These results significantly outperformed a baseline configuration of state-of-the art time-domain features.