In this paperHidden Markov Models (HMMs) are trained to classify half-second segments of six channels, EEG data into one of five classes, corresponding to five cognitive tasks performed by four subjects. The designed HMMs took takes into account the variability of EEGs during different sessions. Based on Bayesian Inference Criterion (BIC), the proposed HMM training algorithm is able to select the optimal number of states corresponding to each set of EEG training records. The previous classifying approaches are used in order to classify between a combinations of two mental tasks. However, in this work we will classify five mental tasks in the same time using HMMs, where EEG signals are represented as autoregressive (AR) models. The training procedures and the validation results of the models are reported and studied. The obtained results give a correct and a promising classification rates for all subjects which is the objective of this research work.