The use of stimulus evoked electroencephalography (EEG) potentials is the mainstream of brain-computer interface (BCI) spellers due to their high performance. On the other hand, to establish non-stimulus based BCI spellers which allow users to input an imagined arbitrary character, other streams using electrocorticography or functional magnetic resonance imaging (fMRI) have been emerged since high spatial discrimination ability seems to be essential for non-stimulus based BCI spellers. Considering the necessity of non-invasiveness and real-time data acquisition for BCI spellers, however, EEG would be ideal if its low spatial discrimination ability can be increased. In this study, we attempted to use EEG cortical currents to address the issue. Since EEG cortical currents are considered to be source signals of EEG signals, they have higher spatial discrimination than EEG signals. We estimated EEG cortical currents from EEG signals with a variational Bayesian method that uses fMRI data as a hierarchical prior. Using vowel imagery tasks, the binary classification (vowels /a/ and /i/) using a sparse logistic regression method revealed that mean classification accuracy of EEG cortical currents were higher than that using EEG signals. Besides the possibility of EEG cortical currents for non-stimulus based BCI spellers, this approach showed importance of parameters tuning for EEG cortical current estimation to obtain best performance, which might indicate the individual difference of brain areas for vowel processing.