Calibration of brain-machine interfaces exploiting event-related potentials has to be performed for each experimental paradigm. Even if these signals have been used in previous experiments with different protocols. We show that use of signals from previous experiments can reduce the calibration time for single-trial classification of error-related potentials. Compensating latency variations across tasks yield up to a 50% reduction the training period in new experiments without decrease in online performance compared to the standard training.