audibility of head-above-torso orientation in head-related...

Audibility of head-above-torso orientation inhead-related transfer functions

Fabian Brinkmann, Reinhild Roden, Alexander Lindau, Stefan WeinzierlAudio Communication Group, TU Berlin, Germany.

SummaryHead-related transfer functions (HRTFs) incorporate the fundamental cues for human spatial hear-ing, which are often applied to auralize sound fields based on numerical room acoustic modelingtechniques. HRTFs are typically available for various directions of sound incidence and a fixed head-above-torso orientation (HATO). Consequently, if head movements are tracked in order to exchangeHRTFs in interactive auralizations, they correspond to a listener turning head and torso simultane-ously, while – in reality – listeners usually turn their head independently above a fixed torso. Earlierstudies showed informal evidence that the influence of the head-above-torso orientation might beaudible. A convincing perceptual evaluation was, however, not available so far. In the present studywe used an ABX listening test design to assess whether correctly accounting for head-above-torsoorientations will be audible in a dynamic binaural simulation. Eleven subjects compared interactivesimulations with constant and variable head-above-torso orientations for different directions of soundincidence and for different audio stimuli. Results show that differences were audible for all testedsource positions and stimuli thereby underlining the perceptual relevance of correct head-above-torsoorientations for binaural synthesis.

PACS no. 43.60.Sx, 43.66.Pn

1. Introduction

Numerical modeling techniques used to simulate thesound field in room acoustical environments are typ-ically auralized by superposing the direct sound andeach reflection at a listening position with correspond-ing HRTFs to obtain binaural room impulse responses(BRIRs, [1, p. 272]). In order to obtain a stableand plausible representation of the simulated environ-ment, it is important to account for head movementsof the listener, which help to resolve front-back con-fusions [2], and are naturally used to judge attributessuch as the envelopment, timbre or the width and lo-cation of a sound source [3, 4]. However, HRTFs areusually measured for different angles of sound inci-dence relative to a fixed dummy head or head andtorso simulator, so that head movements at the repro-duction stage will necessarily correspond to a listenermoving head and torso (Fig. 1, left). In a physiologi-cally typical situation, however, the head would be ro-tated independently above a fixed torso (Fig. 1, right).

The effect of the torso on HRTFs was extensivelystudied by Algazi et al. [5] for static binaural synthesis

(c) European Acoustics Association

and a neutral head-above-torso orientation. Authorsfound that in case the direct path from the soundsource to the ear is blocked by the torso, shadowingoccurs for frequencies above approximately 100 Hz,causing increasing attenuation of up to 25 dB. Forother directions of sound incidence, the torso wasshown to act as a reflector causing comb-filters withan amplitude of up to ±5 dB to interfere with themagnitude spectrum of the HRTF. The exact posi-tions of peaks and dips of the comb filter thus dependsmainly on the source elevation. For a source above thelistener the first dip was shown to occur already at afrequency as low as 700 Hz. Above 3 kHz however, theHRTF is dominated by distinct head and pinnae cueswith a magnitude of ±20 dB [6, 7].From an analysis of HRTFs measured for varioushead-above-torso orientations, Guldenschuh et al. [8]found, that most prominent torso reflections occur ifear, shoulder, and source are approximately aligned,and if the source elevation is within 20◦ below to 40◦above the horizontal plane. It was further hypothe-sized, that effects caused by the torso are audible atleast for critical source positions.Despite the dominating role of head and pinnae ef-fects on the HRTF, Genuit [9] already assumed torsoeffects to hold localization cues at frequencies below3.5 kHz. Algazi et al. [7] conducted localization exper-iments with stimuli low-passed at 3 kHz, and showed

Brinkmann et al.: Audibility of head orientation in HRTFsFORUM ACUSTICUM 20147-12 September, Krakow

Figure 1. Illustration of head movements with constant(left) and variable (right) head-above-torso orientation.

that such cues aid in detecting the elevation of soundsources outside the median plane.In summary, several previous studies cited above sup-port the hypothesis that accounting for correct head-above-torso orientations is necessary for a percep-tually transparent binaural synthesis. Thus, in thepresent study, we physically and perceptually exam-ined differences between dynamic auralizations allow-ing for either constant or variable head-above-torsoorientations.

2. HRTF measurements

Before being able to assess the effect of head-above-torso orientation, an appropriate HRTF data set wasmeasured with the head and torso simulator FABIAN,which is equipped with a software-controlled neckjoint, allowing for a precise control of the head-above-torso orientation in multiple degrees of freedom [10].Transfer functions were measured in the fully anechoicchamber of the TU Berlin (V=1850 m3, fc=63 Hz).HRTFs were measured for head orientations in a rangeof ±82◦ azimuth with a resolution of 0.5◦ for six soundsources each distributed on the frontal hemisphere.Each head orientation was measured (a) for a con-stant and (b) for a variable head-above-torso orien-tation. The range of head movements was chosen ac-cording to the maximum range of motion [12, 13]. Thespatial resolution was assumed to be imperceptiblyfine, as it was smaller than the worst-case localization

Figure 2. Photo of the HRTF measurement setup takenwhile adjusting the source position with the help of a lasermounted below FABIAN’s the left ear.

blur of 0.75◦ reported by Blauert [14, p. 39]. In total,329 HRTFs were measured per source position and forboth, constant and variable head-above-torso orienta-tion.Sound source positions were chosen to be both typi-cal (e.g. in front of a listener) and particularly crit-ical with respect to a strong shoulder/torso effect.The latter is the case for source positions causing thetorso to act either as a strong reflector or as an ob-stacle for the sound field at the ears. Hence, soundsource positions are particularly critical when beingaligned with shoulder and ear, or when resulting ina strongly shadowed contralateral ear. Consequently,sources were placed as follows: (1) ϕ = 0◦ azimuth;ϑ = 90◦ elevation, (2) 315◦; 30◦, (3) 0◦; 0◦, (4) 45◦;0◦, (5) 90◦; 0◦, and (6) 315◦; −30◦. Hereby, azimuthangles of 90◦ denote sources to the left of FABIAN,while positive elevations refer to sources above thehorizontal plane. Exact source positions were deter-mined using a laser attached to FABIAN’s neck joint.Sources were placed at distances between 2.1 m and2.6 m, thus avoiding proximity effects [15, 16] andensuring that reflections from the speakers could beremoved by windowing.For excitation, we used sine sweeps between 50 Hzand 21 kHz and of an FFT order of 16. A bass em-phasis was applied for SNR optimization at low fre-quencies [17]. A peak-to-tail SNR of about 90 dBwas achieved by averaging across two repetitions ofthe sweep signal. Measurements were conducted usingGenelec 8030a active studio speakers, and time vari-ability of the loudspeakers’ frequency response couldbe reduced to ±0.2 dB by means of an one hour awarm-up procedure.Controlling time variability wasparticularly important because HRTFs measured atdifferent points in time were to be directly comparedto each other in a highly sensitive listening test lateron. The measurement setup is shown in Fig. 2.Subsequent to the HRTF measurements, FABIANwas removed and its DPA 4060 miniature electret


condenser microphones were detached for conductingreference measurements. The positions of the micro-phones were adjusted to be identical to FABIAN’sinteraural center. Finally, HRTFs were calculatedby spectral division of the measured HRTF andthe reference spectrum, simultaneously compensat-ing for transfer functions of loud-speaker and micro-phone. Further processing of HRTFs was conductedas described in Brinkmann et al. [18], however low-frequency extrapolation was excluded as it was be-lieved to be irrelevant in this context.

3. Physical evaluation

This section presents a physical evaluation of thetorso’s influence on HRTFs as a function of head-above-torso orientation and source position. Observeddifferences between head movements with constantand variable head-above-torso orientation are dis-cussed and a critical subset of source positions is se-lected for perceptual evaluation in a subsequent lis-tening test.

3.1. Method

Differences in HRTFs were examined with respect tointeraural time and level differences (ITD, ILD), aswell as with regard to spectral coloration. Therefore,ILDs were estimated as RMS level differences betweenleft and right ear, whereas ITDs were calculated asdifferences in onset times between left and right earusing onset detection on ten times up-sampled head-related impulse responses (HRIRs) [19].In order to obtain an impression of the spectral dif-ferences, the log-ratio of the magnitude responses be-tween the HRTFs for constant and variable head-above-torso conditions was calculated (in dB) as

∆HRTF(f) = 20lg|HRTFconst(f)||HRTFvar(f)|

, (1)

where f is the frequency in Hz. For convenience, thedependency of the HRTF on head orientation, sourceposition, and left and right ear was omitted.For a better comparability across source positions, asingle value measure was calculated based on Min-naar et al. [20], who described the error between areference and an interpolated HRTF by averagingabsolute magnitude differences at 94 logarithmicallyspaced frequencies, and adding results for left andright ear. This physical measure was a good predictorfor listening test results, where subjects had to detectdifferences between original and interpolated HRTFs.However, instead of calculating the error for discretefrequencies we used a Gammatone filter bank, as sug-gested by Schärer and Lindau [21]. The error level (indB) in one filter band is given by

∆HRTF(fc) = 20lg

∫C(f, fc) |HRTFconst(f)| df∫C(f, fc) |HRTFvar(f)| df

, (2)

Figure 3. HRTFs measured for the right ear and a sourceat (315◦; 30◦) with constant (top) and variable (middle)head-above-torso orientation. Difference in magnitude re-sponses are shown at the bottom.

where C is a Gammatone filter with center frequencyfc in Hz as implemented in the Auditory Toolbox [22].The error level ∆HRTF(fc) was calculated for N = 39auditory filters between 70 Hz and 20 kHz. Then, theresults for left and right ear were added and averagedacross fc resulting in a single value error measure ∆G(in dB) for each pair of HRTFs with constant andvariable head-above-torso orientation

∆G =1

N

∑fc

(|HRTFl(fc)|+ |HRTFr(fc)|

)(3)

3.2. Results

On average, differences in ITD and ILD between headmovements with constant and variable head-above-torso orientation were well below difference thresholdsof 10 µs and 0.6 dB [14, pp. 153], whereas maximumdeviations of 11.4 µs and 0.95 dB only slightly exceedit.Two exemplary HRTF sets for head movements withconstant and variable head-above-torso orientationare depicted in Fig. 3 for the right ear and a sourceat (315◦; 30◦). For both cases, a comb filter causedby the shoulder reflection is visible for frequenciesup to approx. 1 kHz. Above 1 kHz, strong peak andnotch patterns caused by pinnae resonances are mask-ing the visibility of the shoulder comb filter. How-ever, when calculating the spectral difference accord-ing to (1) identical head-to-source orientation causethe high frequency pinna cues to be cancelled out


(Fig. 3, bottom). Thus, the comb filter becomes visi-ble again over the full frequency range. As expected,observed differences are nearly negligible for head-to-source orientations in the vicinity of 0◦, as in this casehead-above-torso orientations for constant and vari-able head movements are very similar. For other headorientations comb-filter-like structures stretch across0.7 to 20 kHz. They result as the difference of twomismatched comb filters which are related from todifferent distances between ear and shoulder, in thecases of either constant or variable head-above-torsoorientations. Similar difference patterns were observedfor other source positions, with a tendency for largerdeviations to occur at the contralateral ear. Devia-tions due to shadowing of the torso were present be-low 700 Hz, but less pronounced even for the source at−30◦ elevation. This is in good accordance with Al-gazi et al. [5], where strongest influences of shadowingin HRTFs were found for sound sources below −40◦

elevation and the contralateral ear (measurements ofa KEMAR mannequin).The differences according to (3) are shown in Fig. 4 forall source positions and head-to-source orientations.As mentioned above, deviations are small in the vicin-ity of 0◦ whereas elsewise they reach a maximum ofup to 2.5 dB. Moreover, a tendency for the error toincrease with decreasing source elevation can be seen:The smallest error of ≤ 1 dB is found for the source at(0◦; 90◦). In this case, the shoulder reflection is weakfor both constant and variable head-above-torso ori-entations as most energy is reflected away from theear. Intermediate differences of up to 1.5 dB occurfor the sources on the horizontal plane and 30◦ ele-vation, most likely caused by strong shoulder reflec-tions. The largest error of 2.5 dB was found for thesource at −30◦ elevation and for head-to-source ori-entations larger than 45◦ azimuth, because the ear ispartly shadowed by the torso in one case (constanthead-above-torso orientation). Confirming the obser-vations of Guldenschuh [8], the error for sources onthe horizontal plane can be seen to increase with de-creasing source azimuth.

4. Perceptual evaluation

To test whether or not differences are audible be-tween head movements with either constant or vari-able head-above-torso orientation, a highly sensitiveABX listening test was conducted. The setup allowedfor instantaneous and repeated comparison betweenHRTF sets using a dynamic binaural auralization ac-counting for head movements of the listeners.

4.1. Method

Seven men and four women with a median age of 29years and normal hearing participated in the test.Hearing level, assessed by means of fixed tone au-diograms averaged across octave frequencies between

Figure 4. Differences between HRTFs with constant andvariable head-above-torso orientation calculated accordingto (3). Results for sources 1 to 6 are shown from top tobottom.

500 Hz and 4 kHz, was below 25 dB(HL) for all sub-jects (Classified no hearing impairment by the WorldOrganisation [23]). All subjects had a musical back-ground; nine subjects had participated in listeningtests before.Following the ABX paradigm, three stimuli (A, B,and X ) were presented to the subjects, whose taskwas to identify whether A or B equaled X. Subjectswere encouraged to switch between stimuli and taketheir time at will before giving an answer, while headmovements were randomly assigned to A, B, and X.A single test consisted of 23 ABX sequences. Acrosssubjects, 23 · 11 = 253 observations were made un-der each condition. To obtain cumulated type 1 andtype 2 error levels of 0.05 after accounting for mul-tiple testing by Bonferroni correction, results for onetested condition will be significantly above chance at≥147 observed correct answers, indicating audibilitybetween head movements. Insignificant results (<147correct answers) imply group averaged detection ratesto be significantly below 65% [24].For reproduction of binaural signals, the BKsystem[25] consisting of a low noise DSP-driven amplifierand extraaural headphones was used in conjunctionwith a Polhemus Patriot tracker for monitoring thesubjects head positions. The fast convolution enginefWonder [10] including an approach for post-hoc ITDindividualization [19] was used for dynamic binauralsimulation, allowing for head movements in the rangeof ±82◦ azimuth. fWonder was also used for the ap-plication of a headphone compensation filter, whichwas designed using regulated inversion based on head-phone transfer functions measured on FABIAN [26].


In order to limit the duration of the experiment, asubset of three sound sources was selected for percep-tual evaluation. By drawing on results of the physicalevaluation, particularly critical and uncritical sourcepositions at (0◦; 90◦), (90◦; 0◦), and (315◦; 30◦) wereselected. Two different audio stimuli were used: Pinknoise was chosen in order to reveal spectral differ-ences, an excerpt of anechoic male speech was used asa familiar and typical ’real-life’ sound with a time-variant behavior. The combination of three soundsources and two audio contents lead to six conditionswhich were assessed individually by each subject. Theexperiment was split in two blocks while successivelypresenting the three source conditions per audio con-tent. Source positions were randomized within blocks,and the sequence of blocks was balanced across sub-jects.The test was conducted in a quiet listening room(Leq,A = 38 dB SPL). Subjects were placed on a re-volving chair to comfortably reach and hold arbitraryhead orientations. The whisPER listening test envi-ronment1 [27] was used for the test, with the userinterface displayed on a touchpad. Training was con-ducted prior to the listening test in order to familiarizesubjects with the interface and stimuli. Subjects max-imally needed 1.5 hours for rating and were encour-aged to take breaks at will in order to avoid fatigue.

4.2. Results

Individual results, and results averaged across sub-jects are shown in Fig. 5 for all test conditions. Av-erage results, as given by the white horizontal bars,indicate a clear distinguishability of head movementswith constant and variable head-above-torso orienta-tion. For all tested conditions, even for relatively un-critical source positions at (0◦; 90◦), the results weresignificantly above chance.When asked for the most prominent perceived dif-ference between head movements with constant andvariable head-above-torso orientation, subjects namedcoloration in case of the noise content and colorationand/or localization for the speech sample.

5. Discussion and Conclusion

In this study, we assessed the audibility of differencesoccurring during head movements with constant orvariable head-above-torso orientation. To this endwe conducted a physical evaluation of spectral andtemporal deviations, as well as a highly sensitiveABX listening test. Results show that differenceswere audible for all tested source positions and audiocontents, stressing the importance of accounting forcorrect head-above-torso orientation if aiming at anauthentic auralization, i.e. an auralization that is

1 http://dx.doi.org/10.14279/depositonce-31

(0°;90°) (90°;0°) (315°;−30°) (0°;90°) (90°;0°) (315°;−30°)

37.5

50

62.5

75

87.5

10025

28 11

2

52 5

noise speech

perc

enta

ge o

f cor

rect

ans

wer

s

(0°;90°) (90°;0°) (315°;−30°) (0°;90°) (90°;0°) (315°;−30°)

7

9

11

13

15

17

19

21

23

num

ber

of c

orre

ct a

nsw

ers

Figure 5. Listening test results of 11 subjects and for eachtested condition. Black dots indicate percentage/numberof correct answers; depicted numbers show how many sub-jects had identical results. Mean scores (given by solidwhite lines) above dashed line are significantly abovechance.

indistinguishable from a corresponding real soundfield.Differences between head movements with staticand variable head-above-torso orientation were evenaudible for a source at(0◦;90◦), which exhibited thesmallest deviations in the physical evaluation. Thus,the outcome of the listening test can be expected tohold for all source positions measured in the currentstudy.From results of the physical evaluation, we supposeperceived differences to be mostly due to spectraldifferences, because observed deviations in ITDs andILDs were below the threshold of audibility for thevast majority of head-above-torso orientations andsource positions. Perceived differences in localization,as reported for the speech stimulus, might also bedue to spectral cues, with mismatched comb filtersin HRTFs exciting different directional bands [14,pp. 93] and thus evoking differences in perceivedelevation. This assumption would be in accordancewith Algazi et al. [7], who found torso and shoulderrelated cues to be involved in the perception ofelevation for sources outside the median plane.Moreover, results of our physical evaluation supportfindings of earlier studies regarding the comb-filterlike effect of the shoulder reflection on the HRTF,which was found to be most prominent for if soundsource, shoulder, and ear are aligned [8, 5, 7].

In summary, the results from our study emphasizethe importance of providing correct head-above-torsoorientations in dynamic auralizations. However, thiswould require HRTFs data containing various head-above-torso orientations and a fine resolution of soundsource positions, as provided by a recently measuredHRTF data set [18]. Because such data sets are costlyto produce, a perceptually transparent interpolationbetween HRTFs with different head-above-torso ori-


entations would be all the more desirable. This wasexamined by the authors in a follow-up study [28].

Acknowledgement

This work was done within the Simulation and Evalu-ation of Acoustical Environments (SEACEN) researchgroup, and funded by the German Research Founda-tion (DFG WE 4057/3-1).

References

[1] M. Vorländer: Auralization. 1st Edition, Springer,Berlin, Heidelberg, 2008.

[2] P. Minnaar, S.K. Olesen, F. Christensen, H. Møller:The importance of head movements for binaural roomsynthesis. Proc. 2001 Int. Conf. Auditory Display. Es-poo, Finnland.

[3] W.R. Thurlow, J.W. Mangels, P.S. Runge: Head move-ments during sound localization. J. Acoust. Soc. Am.42(2):489-493 (1967).

[4] C. Kim, R. Mason, T. Brookes: Head movements madeby listeners in experimental and real-life listening ac-tivities. J. Audio Eng. Soc. 61(6):425-438 (2013).

[5] V.R. Algazi, R.O. Duda, R. Duraiswami, N.A.Gumerov, Z. Tang: Approximating the head-relatedtransfer function using simple geometric models of thehead and torso. J. Acoust. Soc. Am. 112(5):2053-2064(2002).

[6] M.B. Gardner: Some monaural and binaural facets ofmedian plane localization. J. Acoust. Soc. Am. 54(6):1489-1495 (1973).

[7] V.R. Algazi, C. Avendano and R.O. Duda: Elevationlocalization and head-related transfer function analysisat low frequencies. J. Acoust. Soc. Am. 109(3): 1110-1122 (2001).

[8] M. Guldenschuh, A. Sontacchi, and F. Zotter: HRTFmodelling in due consideration variable torso reflec-tions. Proc. 2008 Acoustics. Paris, France.

[9] K. Genuit: Ein Modell zur Beschreibung vonAußenohrübertragungseigenschaften. Doctoral Thesis.Techn. Hochsch. Aachen, Germany (1984).

[10] A. Lindau, T. Hohn, S. Weinzierl: Binaural resynthe-sis for comparative studies of acoustical environments.Proc. 122th AES Convention 2007. Vienna, Austria.

[11] A. Lindau, S. Weinzierl: On the spatial resolution ofvirtual acoustic environments for head movements onhorizontal, vertical and lateral direction. Proc. 2009EAA Symposium on Auralization.

[12] C.T. Morgan, A. Chapanis, J.S. Cook, M.W. Lund:Human Engineering Guide to Equipment Design.McGraw-Hill, New York, 1963.

[13] P. Schöps, N. Seichert, M. Schenk, U. Petri andE. Senn: Alters- und geschlechtsspeziïňĄsche Bewe-gungsausmaße der Halswirbelsäule,âĂİ Phys. Rehab.Kur. Med. 7(3): 80-87 (1997).

[14] J. Blauert: Spatial Hearing. The psychophysics of hu-man sound localization. MIT Press, 2nd Edition, Mas-sachusetts, USA, 1997.

[15] D.S. Brungart, W.M. Rabinowitz: Auditory localiza-tion of nearby sources. Head-related transfer functions.J. Acoust. Soc. Am. 106(3):1465-1479 (1999).

[16] H. Wierstorf, M. Geier, A. Raake, S. Spors: A freedatabase of head-related impulse response measure-ments in the horizontal plane with multiple distances.Proc. 130th AES Convention 2011 (Engineering Brief),London, UK.

[17] S. Müller, P. Massarani: Transfer function measure-ment with Sweeps. Directors’s cut including previouslyunreleased material and some corrections. J. AudioEng. Soc. (Original release) 49(6):443-471 (2001).

[18] F. Brinkmann, A. Lindau, S. Weinzierl, G. Geissler,S. van de Par: A high resolution head-related transferfunction database including different orientations ofhead above the torso. Proc. AIA-DAGA 2013, Interna-tional Conference on Acoustics, Meran, Italy, 596-599.

[19] A. Lindau, J. Estrella, S. Weinzierl: Individualizationof dynamic binaural synthesis by real time manipula-tion of the ITD. Proc. 128th AES Convention 2010,London, UK.

[20] P. Minnaar, J. Plogsties, F. Christensen: Directionalresolution of head-related transfer functions requiredin binaural synthesis. J. Audio Eng. Soc. 53(10):919-929 (2005).

[21] Z. Schärer and A. Lindau: Evaluation of equalizationmethods for binaural signals. Proc. 126th AES Con-vention 2009, Munich, Germany.

[22] M. Slaney: Auditory Toolbox. Technical Report#1998-010, Interval Research Corporation. Availableat https://engineering.purdue.edu/~malcolm/interval/1998-010, accessed May 2014.

[23] World Health Organisation: Grades of hearing im-pairment. Available at, www.who.int/pbd/deafness/hearing_impairment_grades/en, accessed May 2014.

[24] L. Leventhal: Type 1 and type 2 errors in the sta-tistical analysis of listening tests. J. Audio Eng. Soc.34(6): 437-453 (1986).

[25] V. Erbes, F. Schulz, A. Lindau, S. Weinzierl: An ex-traaural headphone system for optimized binaural re-production. Proc. 2012 German annual acoustic con-ference (DAGA), 313-314.

[26] A. Lindau, F. Brinkmann: Perceptual evaluation ofheadphone compensation in binaural synthesis basedon non-individual recordings. J. Audio Eng. Soc.60(1/2):54-62 (2012).

[27] S. Ciba, A. Wlodarski, H.J. Maempel: WhisPER –A new tool for performing listening tests. Proc. 126thAES Convention 2009, Munich, Germany. Available athttp://dx.doi.org/10.14279/depositonce-30.

[28] F. Brinkmann, R. Roden, A. Lindau, S. Weinzierl:Interpolation of head-above-torso orientation in head-related transfer functions. Proc. 2014 Forum Acus-ticum. Krakow, Poland (Accepted for publication).

audibility of head-above-torso orientation in head-related...

Documents