Brain-Computer Interfaces (BCIs) are systems able to transform the brain signals into commands to control a device. Different instruments are used to acquire brain signals, in this study electroencephalography (EEG) was employed to record the brain electrophysiological signals. In particular, the use of EEG is of great interest due to its simple operation and low cost. Former EEG-based BCIs used an external stimulus in which the brain activity related to such stimulus is known [Farwell and Donchin, 1988]. The present study is based on the use of an internal stimulus related to language, known as imagined speech, which is the action of imagining the diction of a word without emitting nor articulating any sound [Torres-García et al., 2016]. The use of imagined speech may provide a new communication channel for computers. Beyond aiding people with disabilities, this approach can change their interactions with electronic devices. An imagined speech based BCI requires the processing of brain signals to detect brain activity related to a specific word. Previous EEG-based BCIs have proposed methods for feature extraction and analysis. Some have attempted to represent EEGs as multidimensional data and have analyzed these representations in different ways [Ji et al., 2016, Zhao et al., 2009, Li and Zhang, 2010, Lee et al., 2007]. This study proposed a new method that integrates the dimensions of an EEG signal to find appropriate patterns for imagined speech discrimination. Therefore, a multiple variable (multivariate) representation will be proposed, as well as a deep learning architecture. These methods will be compared in subsequent experiments.