The P300 speller is a brain-computer interface (BCI) system designed to communicate language by presenting language stimuli and detecting event related potentials in a subject's electroencephalogram (EEG) signal. The target patient population is prone to fatigue, so reducing or removing this training step could increase the amount of time available to the subject for actual BCI use. We present an expectation maximization approach that trains the classifier in an unsupervised manner. A general classifier is created from a set of multiple subjects and it is then refined using the subject's unlabeled data and knowledge from the language domain. The method was tested offline on a data set of 15 healthy subjects and achieved similar performance to fully supervised methods for all subjects. This suggests that this method could be used in the place of the training step for BCI systems.