We present an efficient implementation of brain-computer interface (BCI) based on high-frequency steady state visually evoked potentials (SSVEP). Individual shape of the SSVEP response is extracted by means of a feedforward comb filter, which adds delayed versions of the signal to itself. Rendering of the stimuli is controlled by specialized hardware (BCI Appliance). Out of 15 participants of the study, nine were able to produce stable response in at least eight out of ten frequencies from the 30-39 Hz range. They achieved on average 96±4% accuracy and 47±5 bit/min information transfer rate (ITR) for an optimized simple seven-letter speller, while generic full-alphabet speller allowed in this group for 89±9% accuracy and 36±9 bit/min ITR. These values exceed the performances of high-frequency SSVEP-BCI systems reported to date. Classical approach to SSVEP parameterization by relative spectral power in the frequencies of stimulation, implemented on the same data, resulted in significantly lower performance. This suggests that specific shape of the response is an important feature in classification. Finally, we discuss the differences in SSVEP responses of the participants who were able or unable to use the interface, as well as the statistically significant influence of the layout of the speller on the speed of BCI operation.