Head-related transfer functions (HRTFs) contain information that is vital both for the analysis of acoustic cues used in sound localization and for the application of virtual-sound synthesis techniques. As the HRTFs measured for a representative set of source positions constitute a large body of data, the application of suitable data reduction techniques is of crucial importance. An obvious method, which takes the properties of the human ear into account, is the averaging of the frequency spectrum within equidistant intervals on a logarithmic frequency scale. This method was perceptually evaluated in an experiment in which six listeners compared original (individualized) HRTFs, obtained for ten source positions, with HRTFs averaged within 1/3-, 1/6-, or 1/12-octave bands. Sound stimuli were bursts of pink noise with either a flat spectrum or with ± 2 or ± 4 dB roving within 1/3-octave bands. It appeared that differences could only be perceived for the HRTFs averaged within 1/3-octave bands. Neither roving level nor virtual source position had a significant effect on the results. It was furthermore found that data reduction of the HRTFs using principle component analysis is somewhat more efficient when HRTFs are first averaged within 1/6-octave bands.