Nature © Macmillan Publishers Ltd 19988Role of spectral detail insound-source localizationAbhijit Kulkarni & H. Steven ColburnHearing Research Center and Department of Biomedical Engineering,Boston University, Boston, Massachusetts 02215, USA.........................................................................................................................Sounds heard over headphones are typically perceived inside thehead1(internalized), unlike real sound sources which are perceivedoutside the head (externalized). If the acoustical waveforms from areal sound source are reproduced precisely using headphones,auditory images are appropriately externalized and localized1±4.The ®ltering (relative boosting, attenuation and delaying of com-ponent frequencies) of a sound by the head and outer ear providesinformation about the location of a sound source by means of thedifferences in the frequency spectra between the ears as well as theoverall spectral shape. This location-dependent ®ltering is explicitlydescribed by the head-related transfer function (HRTF) from soundsource to ear canal. Here we present sounds to subjects throughopen-canal tube-phones and investigate how accurately the HRTFsmust be reproduced to achieve true three-dimensional perception ofauditory signals in anechoic space. Listeners attempted to discrimi-nate between `real' sounds presented from a loudspeaker and`virtual' sounds presented over tube-phones. Our results showthat the HRTFs can be smoothed signi®cantly in frequency withoutaffecting the perceived location of a sound. Listeners cannotdistinguish real from virtual sources until the HRTF has lost mostof its detailed variation in frequency, at which time the perceivedelevation of the image is the reported cue.Although previous studies1±4showed that real and virtual sounds(matched to have identical acoustic waveforms in the ear canals)were indistinguishable when the subject was wearing headphones,they did not address the question of the sensitivity of the subjects tothe details of the HRTF. Here we presented virtual sounds tosubjects' ears through open-canal `tube-phones' (Fig. 1), so that`real' sounds were essentially unaffected by the presence of the tube-phones and sounds from real and virtual sources could be directlycompared. The paired-comparison experiments required listenerssimply to report the order of the stimuli: real ®rst or virtual ®rst.Subjects were given practice and trial-by-trial feedback so that smalldifferences in the location or the realism of the sounds could be usedfor judgements. In the measurements with smoothed HRTFs, thestimulus spectrum was randomized so that non-spatial attributes(such as timbre) could not be used for discrimination.In the ®rst experiment, four subjects were tested for their abilityto distinguish natural stimuli from virtual stimuli, which wereconstructed to match the natural stimuli exactly. Source spectrawere not varied (beyond the stochastic nature of the noise wave-forms) in these initial validation experiments. We tested fourazimuthal locations (0, 45, 135 and 180 degrees, where 0 degreesis straight ahead) in separate sets of trials. None of the subjects wasable to distinguish natural from virtual stimuli. Performance of eachsubject at each location was within the bounds expected fromchance performance. (The percentages of correct values obtainedare plotted with the data from the smoothing experiment in Fig. 2.)These results were consistent with the subjects' impressions thatthey perceived both types of stimuli as completely natural and wereunable to perceive any differences between the virtual and the free-®eld stimuli. These basic results show the adequacy of our methodsfor measurements of HRTFs and for virtual-stimulus presentationas well as the naturalness of the resulting perceptions.In a second set of experiments, we found that, provided that theoverall interaural time delay (ITD) was maintained to be consistentwith that in the natural sound waveforms, virtual sounds synthe-sized through HRTFs with the original magnitude spectra (that is,that of the natural stimulus) but simpli®ed phase spectra were alsoindistinguishable from the natural sounds. This lack of dependenceof the spatial percept on the frequency-dependent detail of the ITDis consistent with previous results1,5,6and led us to develop ourprocedure in which the modi®ed HRTFs were given the minimum-phase response7for each magnitude, supplemented by a pure ITDthat was calculated to make the ITD of the virtual stimulus matchthe ITD of the real stimulus.In all of our other experiments, the magnitude spectra ofempirical HRTFs measured from individual listeners were system-atically smoothed and the dicriminability of the resulting virtualstimuli from free-®eld stimuli was assessed. This smoothing wasperformed by expressing the HRTF log-magnitude spectrum as aFourier series and reconstructing the HRTF spectrum using atruncated series (see Methods). The nature of the smoothingoperation is shown in Fig. 3 for a representative HRTF for sixvalues of the smoothing parameter. For these experiments, thesignal components of the test stimuli (the source waveforms) wereletters to natureNATURE|VOL 396|24/31 DECEMBER 1998|www.nature.com 747Figure 1 The tube-phone apparatus. a, The apparatus attached to a customizedlatex-rubber hollow shell to ®t the ear canals of listeners. b, The apparatusinserted in the ear canal of a listener. Note that the shell provides a snug ®t ¯ushwith the walls of the ear canal with minimal distortion of ear-canal acoustics.Nature © Macmillan Publishers Ltd 19988randomly varied in each trial. The interval-to-interval variation (seeMethods) in the stimulus waveform introduced an uncertainty inthe quality of the sounds in each presentation and preventedlisteners from using non-spatial differences in sound quality (suchas timbre) to discriminate between natural and virtual auditoryimages. We determined a priori that this variation did not alter thelocation of the image from a loudspeaker.The performance of the four subjects tested is shown in Fig. 2 forfour azimuthal locations (0, 45, 135 and 180 degrees). The perfor-mance (in per cent correct) is plotted as a function of the smoothingparameter (the number of Fourier coef®cients used in the HRTFreconstruction). Performance was largely unaffected by the spectralsmoothing; performance was always within the bounds of chanceperformance for reconstructions having more than 16