This paper presents a Brain Computer Interface methodology for unspoken speech recognition based on Electroencephalography (EEG). Each phase within this approach is presented and discussed, followed by the noise elimination methodology and ends up by features extraction and data classification. The presented work consists of database construction with features vectors that will be classified into different classes by applying an articial neural network with three layers. The proposed approach provides results with high percentage of recognition (93% Testing, 95% Validation) when applied on two English words ON/OFF acquired from 2 different resources.