The interest for using of brain computer interfaces as a communication channel has been increasing nowadays, however, there are many challenges to achieve natural communication with this tool. On the particular case of imagined speech based brain computer interfaces, here still exists difficulties to extract information from the brain signals. The objective of this work is to propose a representation based on characteristic units focused on EEG signals generated in imagined speech. From this characteristic units a representation will be developed, which, more than recognize a specific vocabulary, allows to extend such vocabulary. In this work, a set of characteristic units or bag of features representation is explored. These type of representations have shown to be useful in similar tasks. Nevertheless, to determine an adequate bag of features for a specific problem requires to adjust many parameters. The proposed method aims to find an automatic signal characterization obtaining characteristic units and later generating a representative pattern from them. This method finds a set of characteristic units from each class (i.e. each imagined word), which are used for recognition and classification of the imagined vocabulary of a subject. The generation of characteristic units is performed by a clustering method. The generated prototypes are considered the characteristic units, and are called codewords. Each codeword is an instance from a general dictionary called codebook. For evaluating the method, a database composed of the electroencephalograms from twenty seven Spanish native speakers was used. The data consists of five Spanish imagined words ("Arriba", "Abajo", "Izquierda", "Derecha", "Seleccionar") repeated thirty three times each one, with a rest period between repetitions, this database was obtained in [Torres-García et al., 2013]. The proposed method achieved comparable results with related works, also have been tested in different approaches (i.e. transfer learning). Bag of features is able to incorporate frequency, temporal and spatial information from the data. Also, different representations which consider information of all channels, and feature extraction methods were explored. In further steps, is expected that extracted characteristic units of the signals allow to use transfer learning to recognize new imagined words, these units can be seen as prototypes of each imagined word.