Convention Paper Presented at the 120th Convention 2006 May Paris, France

Transcription

1 Audo Engneerng Socety Conventon Paper Presented at the 2th Conventon 26 May 2 23 Pars, France Ths conventon paper has been reproduced from the author's advance manuscrpt, wthout edtng, correctons, or consderaton by the Revew Board. The AES takes no responsblty for the contents. Addtonal papers may be obtaned by sendng request and remttance to Audo Engneerng Socety, 6 East 42 nd Street, New York, New York , USA; also see All rghts reserved. Reproducton of ths paper, or any porton thereof, s not permtted wthout drect permsson from the Journal of the Audo Engneerng Socety. A System for Rapd Measurement and Drect Customzaton of Head Related Impulse Responses Smone Fontana, Angelo Farna 2 and Yves Grener 3 Ecole Natonale Supéreure des Télécommuncatons, Pars, France fontana@ts.enst.fr 2 Unverstà d Parma, Parma, Itala farna@pcfarna.eng.unpr.t 3 Ecole Natonale Supéreure des Télécommuncatons, Pars, France grener@ts.enst.fr ABSTRACT Head-Related Impulse Responses (HRIRs) measurement systems are qute complex and present long acquston tmes for an accurate samplng of the full 3D space. Therefore HRIRs customzaton has become an mportant research topc. In HRIRs customzaton some parameters (generally anthropometrc measurements) are obtaned from new lsteners and ad-hoc HRIRs can be retreved from them. Another way to get new lsteners parameters s to measure a subset of the full 3D space HRIRs and extrapolate them n order to obtan a full 3D database. Ths partal acquston system, of course, should be rapd and accurate. In ths paper we present a system whch allows for rapd acquston and equalzaton of HRIRs for a subset of the 3D grd. Then a technque to carry out HRIR customzaton based on the measured HRIRs wll be descrbed.. INTRODUCTION Bnaural hearng systems am to reproduce at one lstener s ears the same sound feld that he would perceve n a target lstenng space. Ths s done by drect bnaural recordng n the target lstenng space, or n a synthetc way, by the convoluton of a dry sgnal wth the target Room Impulse Response and the lstener Head Related Impulse Responses correspondng to

2 specfc locatons; n the followng we wll focus on ths last approach. Head Related Impulse Responses (HRIRs) are defned as: Pφ, θ ( f ) HRIR( f, φ, θ ) =, P ( f ) REF where P(t) s the pressure at the eardrum of a subject due to a source comng from a drecton defned by θ and φ and P REF (t) s some reference pressure, usually taken at the poston of the lstener head centre. Ths dvson also performs equalzaton, f the two pressures are obtaned wth the same measurement system. The Fourer Transform of an HRIR s called Head Related Transfer Functon (HRTF). As ponted out n [], achevement of spatal consstency requres renderng statc, dynamc, and envronmental cues. In fact t s known that the ntroducton of envronmental cues, such as the room mpulse response or some knd of reverberaton, provdes sound externalzaton n syntheszed bnaural recordngs. Statc cues provde localzaton and rely essentally on HRTFs. HRTFs are strctly ndvdual, and can consderably vary from subject to subject; localzaton degradaton due to non-ndvdualzed HRTFs has been observed, namely n term of front/back resoluton. Ths effect can be reasonably be reduced by lettng the lstener movng ts head to obtan front-back dstncton. Dynamc cues are provded through HRIR real tme nterpolaton and head trackng. Indvdualzed HRIRs are obtaned from measurements of mpulse response of sources placed on a 3D pont grd. The most complete (publc) databases for real subjects HRIRs are the CIPIC ([2]) and the IRCAM Lsten ([3]) databases, whch present HRTFs sets for 45 and 5 subjects. The databases nclude 3D HRIRs measurements wth dfferent spatal resoluton and anthropometrc measurements for the lsteners. The underlyng measurement systems are qute complex and present long acquston tmes. HRIR customzaton tres to overcome the full measurement process. The perfect approach s to solve the wave equaton wth the real subject head, for example by BEM ([4]). Ths technque stll presents an mportant acquston tme. Several approached strateges are under nvestgaton: they can be classed n selecton methods, structural methods, decomposton methods and nterpolaton methods. In selecton methods the customzed HRIRs set s chosen between some HRIRs sets contaned n prerecorded databases, on the bass of nteractve choce of the new lstener, or morphologcal proxmty between the new lstener and the database subject ([], [5]). In structural methods ([6]) the HRTFs are obtaned as cascade flters representng physcal scatterng on the dfferent body parts that can be tuned accordng to the morphology of the new lstener. Decomposton methods ([7]) assocate Prncpal Components Representaton of HRTFs to a reduced set of morphologcal parameters, n order to synthesze HRTFs from anthropometrc measurements. In nterpolaton methods ([8]) a pre-clusterng phase reduces the number of HRIRs to be measured on the new lstener, and nterpolaton s then performed to obtan the full HRIR set. The technque presented n ths paper can be consdered a selecton method based on HRTFs proxmty. A way to perform customzaton n ths case s to measure a subset of the full 3D HRIR space and extrapolate t to a full 3D space. A way to do ths s to compare the measured HRIR subset to the correspondng HRIR subset present n exstng database, select the most fttng one and use the correspondng full 3D set as customzed set. The database selecton s not made on the bass of perceptual parameters or anthropometrc measurements, but drectly on the proxmty of the new lstener s HRTFs to the database HRTFs. Ths s why we call ths customzaton approach Drect Customzaton. If the HRIRs subset s composed by a consequent number of measurements ponts, t could be nterestng to ntegrate these measurements wth the selected set more than smply use t as a reference. Ths means that, nstead of usng the measured subset as a tool to select a whole database set, t could preferable to use the measured HRTFs (the new lstener own HRTF) and complete these wth the HRTF of the most fttng database set. In ths case the two subsets have to be rendered homogeneous, correctng the effects of the dfferent measurements systems. In ths case we wll talk of Integratve Drect Customzaton. To be appled n HRIR customzaton, the HRIR subset measurement system should be rapd and accurate. In ths paper we present a rapd measurement system that employs a crcular loudspeaker array n an anechoc AES 2th Conventon, Pars, France, 26 May 2 23 Page 2 of 8

3 room and a fast-and-easy-to-apply bnaural mcrophone n open meatus confguraton. The overall measurement process takes less than 5 mnutes; the subject can leave wth a full 3D customzed HRTF set on hs USB key mnutes after hs arrval Effect of samplng tme wndow at 5 Hz 2 samples samples samples Accuracy s obtaned through electronc chan equalzaton and a frequency-dependent tme wndowng post-processng step to wndow out paraste reflectons. Low frequences HRTF resoluton s guaranteed by employng a proper tme wndow. 2log H(f) 3 2 After a prelmnary secton to show the lnk between low frequency resoluton and space-tme wndowng, we wll present the acoustcal propertes of the measurement system n secton 3. The measurement process s descrbed n secton 4, whle the post processng s treated n secton 5, where we present the equalzaton and the frequency-dependent tme wndowng. The drect customzaton process s detaled n secton PRELIMINARY REMARKS 2.. Tme-frequency remarks Audo sgnals are wdeband, n the range 2-2 Hz. They can be thought as composed by snusodal components, accordng to the Fourer Theorem. Samplng a wdeband audo sgnal means samplng each one of ts snusodal components. To correctly sample an audo sgnal, t s necessary to use a samplng frequency bgger than the Nyqust frequency, n order to avod alasng. Usng the correct samplng frequency guarantees the correct representaton of hgh-frequences components. What s sometmes forgotten s that the samplng tme wndowng length s an mportant parameter for the resoluton of low-frequences. Ths s of scarce mportance n long audo fles, but for shorter fles (some ms long, as n HRIRs), ths parameter becomes mportant. Let us suppose that we sample an audo fle, and nvestgate how each frequency components s reproduced after samplng. In fgure we show a 5 Hz component sampled at 44 Hz on 2,, samples. It s qute clear from the fgure that wth 2 samples the 5 Hz component s not resolved, whle resoluton gradually mproves augmentng the number of samples Frequency Fgure. Samplng wndow effect at 5 Hz In fgure 2 t s easy to see that a samplng wndow of 5 ponts resolves the Hz frequency. Ths s due to the rato between the lobe wdth and the observed frequency, and not to a reducton of the lobe wdth. 2log H(f) Effect of samplng tme wndow at Hz 5 samples samples Frequency Fgure 2. Samplng wndow effect at Hz The man lobe due to the convoluton of the deal delta functon wth the snc functon related to the rectangular analyss wndow presents a wdth nversely proportonal to the samplng wndow NT s, where N s the number of samples and Ts the samplng tme. In fgure 3 we verfy the theoretcal result keepng fxed the frequency of nterest (say f ) at 5 Hz and plottng the frequency resoluton AES 2th Conventon, Pars, France, 26 May 2 23 Page 3 of 8

4 B R = f 3dB as a functon of the samples number. B -3dB s the wdth of the prncpal lobe at -3dB. Frequency resoluton: Bandwdth 3dB/f Frequency resoluton at 5 Hz METRES MS Hz Hz Table. Space-tme correspondence The goal of the measurement system s obvously to keep the largest number of samples avodng (as long as t s reasonably possble) paraste reflectons. Ths can be done by smply wndowng out the mpulse response whch s thought to be spurous or, n an optmzed way, by frequency dependent tme wndowng, as we ll see n secton Samples Number Fgure 3. Frequency resoluton at 5 Hz. We observe the expected /N behavour of the functon and conclude that a.5 frequency resoluton at 5 Hz s guaranteed wth a samplng wndow of at least 24 samples Space-tme remarks HRIRs measurement systems are usually located n enclosed spaces (rooms, anechoc rooms, etc.) and are composed by more or less extended measurement equpment (mcrophones, loudspeakers, loudspeakers supports, etc.). All these elements can potentally provoke nterference n the measurements. Keepng long mpulse responses on one sde mproves lowfrequency resoluton, but on the other sde can dangerously make Room-and-Head Related Impulse Responses of supposed only-head Related Impulse Responses. In Table we gve some values of correspondence between samples number, mpulse response length and space radus spanned by the response. For an N- samples mpulse response to be pure the correspondng radus should be free of dffractng or reflectng objects. 3. SYSTEM CHARACTERIZATION The am of ths secton s to gve an acoustcal descrpton of the system, n order to prove ts possblty to provde pure HRIR and eventually defne the better strategy to employ for equalzaton. The electroacoustc chan s composed by a TASCAM MWE 24x24 drect-to-dsk recordng system, lnked by Ethernet to a portable computer for audo downlnk and uplnk. The TASCAM output s nput a 6 channels Yamaha 65 power amplfer, whch feeds 6 Tannoy system 6 loudspeakers. The sgnal s recorded by a Sennheser MKE 22 bnaural mcrophone, lnked to a Behrnger preamp. The two-channel output s nput to the TASCAM. The measurement system s composed by electronc equpment n an acoustc envronment (an anechoc room) that s supposed to be transparent to the measurement. The sweep s electroncally fltered by the playback equpment and recordng transfer functon, and by the acoustc transfer functon that represents the acoustcal path from the loudspeaker to the mcrophone, and that ncludes reflectons and scatterng by the loudspeakers and the loudspeakers supports (and not by the walls, consdered as perfectly absorbng). AES 2th Conventon, Pars, France, 26 May 2 23 Page 4 of 8

5 We measured the mpulse response of the electroacoustc chan, wth the sne sweep method ([9]), n order to obtan the mpulse response and transfer functon from the TASCAM output to the TASCAM nput. 3.. Electronc chan In order to characterze the only electronc chan we put one loudspeaker n an anechoc room (the ENST anechoc room, 4.2x4.5x4.6m) and a Schoeps MK2 mcrophone n the loudspeaker axs at the dstance of meter. The absorpton by the ar on the drect path s consdered as belongng to the electronc chan. The obtaned transfer functon s plotted n fgure 4, where t s compared to the one measured n smlar condtons at IRCAM. 2 Tannoy System 6 transfer functon dfferences to the dfferent playback and measurement equpment, and the dfferent measurement condtons. It s mportant to remark that Tannoy clams a flat frequency response (-3dB values) n the range 52Hz- 2KHz. The response of the MKE22 s qualtatvely deduced by the prevous measurements, comparng the transfer functon measured wth the MKE22 and the prevous transfer functon. The two functons are shown n fgure 5. We can observe that MKE22 seems to loose senstvty n the low frequency range: t presents -9dB at 5 Hz compared to the MK2-measured transfer functon (24 samples). From these prelmnary observatons, we can observe that the system transfer functon contans sgnfcant energy (at -3dB) n the range -9 Hz; 52 ponts HRIRs guarantee.5 resoluton at Hz. In the followng we present a technque that allows for low frequences resoluton down to 5 Hz (whch s the clamed lower frequency bound for Tannoy loudspeakers), that corresponds to 24 samples HRIRs. MKE22 MK2 measured transfer functons 2log H(f) IRCAM Tranfer functon Measured Tranfer functon f Fgure 4 System Transfer Functon, measured wth the MK2 mcrophone (db/hz). 2log H(f) MK2 MK22 The measured transfer functon represents the contrbutons of the cascade of the electronc chan components. We measured separately the contrbuton of the TASCAM, drectly connectng ts output and nput. The response reveals to be flat n the nterest frequency range. The response of the loudspeaker, preamp and amp has been obtaned wth a Schoeps MK2 mcrophone (assumng ts frequency response flat). The response we found s qute smlar to the one measured at IRCAM. The major dfferences are sharpest decays at the range boundary, more energy n the low frequences and a more marked notch around 5 Hz, as t s possble to see n fgure 4. We attrbute these f Fgure 5. System Transfer Functon, measured wth the MKE22 mcrophone. We observed the degree of anechocty of the room on the mpulse response of the system (fgure 6). It s possble to see that the frst sgnfcant reflecton s 4 db under the man peak and then supposed naudble: the room can reasonably be thought as anechoc n ths loudspeaker confguraton. AES 2th Conventon, Pars, France, 26 May 2 23 Page 5 of 8

6 room mpulse response 2 3 2logabs(h(t)) tme Fgure 6. Anechoc Room mpulse response (db/ms). The reflecton takes place around samples 43 and corresponds to a reflecton on a wall (3.2 meters path, for meter drect path) Electroacoustc chan In ths secton we am to characterze the full electroacoustc chan, ncludng all the paths from a loudspeaker to the mcrophone, n presence of the complete measurement system, for two potental measurement systems. Fgure 8. New measurement system Even f the structure was elegant and ergonomc, t only allowed for azmuthal HRIR measurements and, most of all, t provoked an acoustcal nterference that acoustcal correcton dd not completely correct, and that affected n qute a dramatc way the purty of the measured HRIRs. We chose to redesgn the system n order to allow for a more sgnfcant samplng of the 3D spatal grd, and to reduce the amount of nterference due to the loudspeaker support. 2*log h(t) Ancent Impulse Response Tme 2 Impulse Response comparson New Impulse Response 2*log h(t) Fgure 7. Ancent measurement system We mounted a frst measurement system n the anechoc chamber of the ENST, whch measures 4.2 x 4.5 x 4.6 m. The structure was composed by an alumnum support and a set of 2 Tannoy system 6 passve loudspeakers (see fgure 7) Tme Fgure 9. Old and New system mpulse response comparson The actual system s shown n fgure 8. We reduced the number of measurement postons to 6, augmented the structure radus, used normal, thn, loudspeaker support, consdered dfferent elevatons, and used glass wool to solate loudspeakers supports and loudspeakers face. The new system presents a cleaner mpulse response: AES 2th Conventon, Pars, France, 26 May 2 23 Page 6 of 8

7 the frst reflecton s db less compared to the ancent mpulse response (fgure 9). Frst sgnfcant reflecton s also translated n tme from 27 to 43 samples after the drect front. In fgure we report the dfferences between the one-loudspeakers and the 6 loudspeakers transfer functon. The effect of the structure s clearly vsble around sample 4, where a reflecton aganst the n-front loudspeaker takes place. The drect front response tal (untl 39 samples) stays below 35 db and we wll consder t as mmune from paraste reflectons. 2*log h(t) 2*log h(t) Room Impulse Response Tme Impulse Response comparson Room and structure Impulse Response Tme Fgure. One loudspeaker and full structure mpulse response comparson 3.3. System calbraton A calbraton of the loudspeaker dstance from the center reference pont has been carred out, measurng the flght tme from each loudspeaker to the mcrophone postoned at the center of the structure. The maxmum dstance dfference s 4 samples, whch correspond to a msplacement of 3 cm. A calbraton for loudspeaker ampltude has equally been made, for a reference taken at the center of the structure. The transfer functons at the center of the structure can vary from one loudspeaker to another, as t s shown n fgure, consderng only the drect front. Ths s due to the dfferences n loudspeaker manufacture and postons but mostly to the loudspeaker dfferent elevaton from the mcrophone, whch determnes outof-axs coloraton. Loudspeakers set at the same elevaton present more smlar functons (fgure 2), whch let us assume that loudspeakers transfer functons are qute constant for the 6 loudspeakers. Analogue results have been obtaned for the same loudspeaker at two recordng postons (for example at the two ears poston). The dfference n ths case can be due to capsule dfferences too. 2*log H(f) Loudspeaker and 4 Tranfer Functons To centre Loudspeaker Loudspeaker f Fgure. Loudspeakers response comparson System senstvty to lttle varatons of measurement ponts We put the MKE22 at the vrtual poston of the ears, whch are marked wth the reference support that s vsble n fgure 8. Then we place t agan, at the vrtual poston of a bgger head. 2*log H(f) Loudspeaker 4 Loudspeaker 5 Loudspeaker 4 and 5: Tranfer functon to centre f Fgure 2. Equal elevaton loudspeaker response comparson AES 2th Conventon, Pars, France, 26 May 2 23 Page 7 of 8

8 The Interaural Tme Dfference (ITD) s coherent wth the new head sze, and a value s guarantee n both cases for azmuth zero, as t s possble to see n fgure 3. Only mnor spectral dfferences (<db) are reported for the transfer functons correspondng to the same ear poston, n the two measurement stuatons. seat heght to make the back of the ears touch the reference supports, and then the MKE 22 (fgure 4) s appled. 25 ITD n two dfferent stuatons 2 5 ITD(samples) loudspeaker Fgure 3. ITD for two heads of dfferent dmensons System characterzaton: conclusons Expermental data, obtaned through the observaton of the 6 mpulse responses let us fx the drect front boundary to 35 samples after the drect front arrval. Ths length guarantees a.5 frequency resoluton at 5 Hz. An equalzaton step s necessary to remove the coloraton due to electroacoustc chan and guarantee 5 Hz frequency resoluton through longer wndowng. The senstvty of the system to dfferent loudspeakers and measurement poston has been consdered comparng the correspondng transfer functons. Even f some dfferences have been observed, a compromse has to be found between the use of a dfferent equalzer for each loudspeaker and each poston, and the use of a reduced number of equalzers robust to varatons and easly scalable wth dfferent measurement geometry choces. 4. MEASUREMENT The measurement ponts are marked wth a fxed support, so that they not vary from one measurement to the other. The person s smply asked to st and tune the Fgure 4. MKE22 mcrophone The 6 sgnals are played through the correspondng loudspeaker. The overall measurement process takes less than 5 mnutes. The recorded wave fle (44 Hz) conssts n the sequence of the 6 recorded chrps. The sgnal s downloaded from the TASCAM and drectly provded to the processng module. In ths prelmnary phase we measured four sets of HRTFs. MKE22 performs open meatus HRIR measurements: occluson of the ear channel by a plug would results n an unnatural hgh frequency reflecton on the plug, as the capsule s not drectly fxed on t. Closed meatus technques are usually preferred because ear channel doesn t provde more spatal nformaton and only adds ndvdual nformaton. For these reasons the blockedentrance pressure s more lkely to be representatve of a populaton []. In order to fully characterze the ndvdual hearng propertes, open meatus measurements can be used. Open meatus measurements wth the MKE22 are taken 6mm outsde the ear channel entrance and nclude the ear channel resonance. The frst constrant does not deterorate the spatal nformaton contaned n the HRTF ([9]) and the resonance can be removed at the headphones equalzaton stage. AES 2th Conventon, Pars, France, 26 May 2 23 Page 8 of 8

9 5. PROCESSING The processng module carres out the followng tasks: mpulse response extracton, equalzaton and frequency dependent tme wndowng, HRTF customzaton (fgure 6). The processng s performed n dffered tme. The overall processng tme s 2 seconds The extracton module performs the sweep deconvoluton, the synchronzaton wth the recorded clock and the HRIR selecton through peak fndng and wndowng, to solate wth a rectangular wndow the 6x2 24 samples mpulse responses. The equalzaton and customzaton modules are treated n detal n the next sectons. 5.. Equalzaton The goal of a HRIRs measurement system s to obtan the free feld mpulse response from a source n a gven poston to the lstener s ears. Ths s the condton that guarantees the good workng of the convoluton process, followng the Green Theory. In practce we obtan the response to an mpulse fltered by the electronc chan n en enclosed space. Both the electronc coloraton and the room reflectons are spurous elements to be elmnated from the measurements. The electroacoustc chan s depcted n fgure 5. The full electroacoustc chan can then be expressed as H( f, r) = H ( f, r) HRTF( f, r), where H S = HTOHY HT H measurement poston. δ(t) TASCAM - output HTO (f) YAMAHA HY (f) S R H M H(f,θ) H B H TI, and r s the h(t,θ) TASCAM - nput HTI (f) BEHRINGER HB (f) If we want to separate the acoustcal (two dmensonalspace, tme) and the electronc (mono dmensonal, tme) parts of the chan, we assume that we can wrte the transducers transfer functons as HT ( f, r) = HTr ( f, r ) H ( f, r) Tr, and HM ( f, r) = HMr ( f, r ) HMr( f, ), r where H Tr ( f, r ) and H (, ) Mr f r are ntended to be measured at a reference poston r n free feld. In ths way we can wrte the transfer functon of the electronc chan f, r ) as H H E ( ( f ) HY ( f ) HTr ( f, r ) H ( f, r ) H ( f ) H ( f ) Mr B TI TO, We remark that H E ( f, r ) does not depend on the measurement poston. The transfer functon of the acoustc chan wthout the lstener ( f, r) as so that H H AWL ( f, r) H ( f, r) H ( f, r) Tr Mr R, H( f, r) = H ( f, r ) H ( f, r) HRTF( f, r) E AWL In an deal system H should be equal to HRTF, and then we need to equalze the measured response, whch suffers from nterference of the room and the electroacoustc chan. Of course, t s possble to nvert drectly the product of the two transfer functons. To quantfy the performances of equalzaton we defne, smlarly as n [], the equalzaton effcency n the frequency doman as the cepstral dstance between the target functon and the equalzed transfer functon. We suppose that the target functon s the Fourer transform of a delta functon that s a flat spectrum. We assume the value of the constant spectrum to be the mean value of the equalzed response. TANNOY HT (f, θ) ROOM HR (f, θ) LISTENER HRTF (f, θ) MKE22 HM (f, θ) ~ ~ j k) = * log H ( k) *log H, where ( Fgure 5. Electro-acoustc chan ~ H ( k ) s the equalzed electroacoustc response. Other choces for the target functon ([5]) are possble. AES 2th Conventon, Pars, France, 26 May 2 23 Page 9 of 8

10 Clock L(t) R(t) Deconvoluton Synchro EXTRACTION J = max( j( k) ) max These values should be better qualfed wth perceptual tests or wth psychoacoustc weghts, but we can use t as reference Electro-acoustc chan equalzaton Selected 3D HRIR set Fgure 6. Processng chan We also defne the standard devaton, J, as: J = Pre-wndowng Peak search Frequency Smoothng Drect Front Equalzaton Room Equalzaton 24 samples response 2 samples response WINDOWING 52 samples response FD-tme-wndowng CUSTOMIZATION K K k = j( k) CIPIC LISTEN 2 EQUALIZATION WINDOWING, One could thnk that the best equalzaton strategy would be to store the twelve mpulse responses from each loudspeaker to each capsule, and buld 2 dfferent equalzers to apply n post-processng to each measurement. Ths s a possble technque, but not optmal, nor scalable. In fact lttle errors on reference poston and measurement ponts can lead, due to a complex room nterference pattern, to severe deteroratons n the equalzaton process, most of all for hgh frequences. Moreover changng n the loudspeaker setup would need another mpulse response measurement sesson. We wll use two equalzers, obtaned from measurements on one loudspeakers (say loudspeaker ), for the two channels left and rght, and use them for the equalzaton of all the loudspeakers. We consder exact and smoothed equalzaton. In exact equalzaton we drectly use the equalzers ssued from nverse electroacoustc channel computaton, followng the algorthm presented n [2]. The basc dea of smoothng s that t s not worth to try to perfectly correct the transfer functon of each loudspeaker at one pont: the matchng of the nverse flter for one poston s not necessarly optmal for lghtly dfferent postons (due, for example, to dfferent loudspeaker postons). Because of ths, t could be nterestng to smooth the reference mpulse response, so as ths becomes more representatve of the mpulse response n proxmty of the measurement pont or of the responses of loudspeakers at dfferent postons. Ths s, of course, a cause of deteroraton of the performances for the reference pont, but t should reveal better for robust equalzaton. Smoothng s performed n the frequency doman, followng [], that s the one mplemented n the Aurora Krkeby Inverse Flterng plug n ([4]). The smoothng s performed on half octave wndows. where K s the number of ponts on whch we perform the FFT. Another relevant parameter s the maxmum devaton AES 2th Conventon, Pars, France, 26 May 2 23 Page of 8

11 5..2. Separated equalzaton When we apply equalzaton on the electroacoustc transfer functon wth smoothng, we do not control the percentage of smoothng on the drect front and on the room. We could thnk that applyng perfect equalzaton for the drect front and a smoothed equalzaton on the structure response could lead to an optmal equalzaton. We can obtan H E (r ), the transfer functon of the electronc chan, by measurng the response of a sngle loudspeaker wth a mcrophone postoned at a dstance of meter, orented to the center of the speaker and elevaton, n order to avod not-n-axs coloraton, n an anechoc room that s supposed equvalent to free feld. The electronc transfer functon represent the drect wave front contrbuton, and can be equalzed n qute a drastc way, because t s supposed not to depend on the error on the measurement poston (r s the real angle at whch the measurement takes place). On the other sde, H E lghtly depends on errors on the measurement, due, for example to dfferent head dmensons, or mperfect replacement or placement of the bnaural mcrophone, or to the loudspeaker poston. Once the drect front has been equalzed for all the loudspeakers, we can obtan the room equalzers. To do ths, we just consder the drect-front equalzed mpulse response for a loudspeaker, obtan the Krkeby nverse flter from ths partally equalzed response, and then use t as we dd for drect front equalzaton, for the equalzaton of all the drect-front-equalzed transfer functons Results The obtaned results are shown n table 2. Electronc Equalzaton of the drect front presents the best results. It s performed on a 35 samples wndowed verson of the HRIR and results are reported only for the drect front equalzaton. As n ths case the response s not senstve to measurement postons, perfect equalzaton can be used n a robust way. Ths explans the superor qualty of equalzaton. In Electronc Equalzaton smoothng the transfer functon s a suboptmal operaton that does not mprove robustness, but only determnes performances deteroraton. Drect front equalzaton only allows for 35 samples-long HRIRs. In order to obtan longer HRIRs, an equalzaton of the response tal s necessary, as explaned. Electroacoustc equalzaton performs jontly the two equalzatons. We can compare Electroacoustc Equalzaton and Acoustc Equalzaton of the prevous drect-front equalzed mpulse response. Electronc J (db) J max (db) Perfect Smoothng Electroacoustc Perfect Smoothng Acoustc Perfect Smoothng.2 5 Table 2. Equalzaton Strateges Comparson It s possble to observe that smoothng plays an mportant role (as forecast) n Electroacoustc Equalzaton, provdng robustness to equalzaton (see J max ). Perfect equalzaton work perfectly for the frst loudspeaker (the one taken for the reference transfer functon): ths explans the nferor J value compared to smoothed equalzaton. For the other loudspeakers smoothng provde better performances, as t s possble to see n fgure 7. For Acoustc Equalzaton of the full drect-frontequalzed response we can observe that perfect equalzaton leads to better results only for the mean value of J, and a smoothed equalzaton provdes n general worst performances than the electroacoustc equalzaton. These results can be explaned not by the worst global performances of the separated equalzaton, but by ts more marked senstvty to partcular bad cases, namely n very hgh frequences (above KHz). Performances are n general lghtly better than those of electroacoustc equalzaton, but we observed that n some partcular cases (namely loudspeaker 5), the equalzaton process leads to strong artfacts, AES 2th Conventon, Pars, France, 26 May 2 23 Page of 8

12 concentrated n very hgh frequences, where the room equalzaton s partcularly crtc. Performng complete electroacoustc equalzaton appears to smooth these llcondtoned cases. To obtan pure HRIRs, these reflectons should be wndowed out. On the other sde, wndowng the mpulse response at sample 35 would not allow for a frequency resoluton below 5 Hz. mean devaton DRC perfect DRC smoothed EQUALIZATION loudspeaker Fgure 7. J for electroacoustc equalzaton Conclusons Takng a look to one equalzed mpulse response (fgure 8), we see that Electroacoustc Equalzaton doesn t elmnate completely the reflectons around sample 4. The resdual weak reflectons present n the obtaned mpulse responses determne a weak frequency comb flterng effect that can affect the local behavor of the mpulse response though preservng the global propertes of the measured HRTFs. Another consderaton can then be made. The reflectons on the loudspeakers and the measurement structure only affect components above a certan frequency. It s known that scatterng on an object takes place for wavelengths comparables wth the object dmensons. In ths case we could fx ths frequency to 4 Hz, for whch half the wavelength (around 4 cm) s comparable wth the loudspeaker dmenson. We could thnk to flter the response tal above sample 35 wth a low-pass flter wth cutoff frequency of 4 Hz, and keep ths part of the mpulse response, that could be mxed wth the all-pass drect front response. We just keep the hgh frequency response untl sample 35, whch ncludes all the useful reflectons. The mx wth the low-passed mpulse response from sample 35 to sample 24 guarantee the low frequences resoluton for snusodal components below 4 Hz, that are not nterested by structure scatterng 5.2. FREQUENCY DEPENDENT WINDOWING Hard Frequency Dependent Tme Wndowng Lookng at the spectrogram (fgure 9) of one of the mpulse responses, we can clearly see reflectons around 9.5, 9.7,, 6 and 7 ms, even f the frst sgnfcant reflecton s -36 db under the man front..5 Esualzed mpulse response Ampltude.2. Frequency samples Fgure 8. Drect front equalzed mpulse response Tme Fgure 9. Drect front equalzed mpulse response spectrogram AES 2th Conventon, Pars, France, 26 May 2 23 Page 2 of 8

13 If we hgh-pass flter the mpulse response, wth a cutoff frequency of 6 Hz, and we look at the spectrogram (fgure 2), we can see that reflectons are stll present and easy to dentfy. On the other hand, flterng the mpulse response wth a low pass flter wth cutoff frequency of 4 Hz, we can see (fgure 2) how the reflectons have dsappeared, and there s not strong dscontnuty n the slow-varyng low frequences mpulse response around ms. mpulse response. Ths phenomenon s clearly vsble n fgure 22, where we report the spectrogram of the H- FDTW processed mpulse response. One way to solve ths problem s to use what we call Soft -FDTW that allows for smoother tme-frequences transtons. We consder as useful nformaton the mpulse response from sample to 35 for hgh frequences (drect front), then we wndow the drect front wth a Tukey wndow, whch allows avodng sharp transtons at the cuttng ponts, preservng the useful nformaton. In a smlar way, we flter the mpulse response tal wth a low pass flter and wndow t from sample 35 to sample 24, wth another Tukey wndow. We mx the two mpulse responses wth 2 samples overlappng. 2 x Fgure 2.Drect front equalzed low-passed mpulse response spectrogram Frequency Frequency-dependent-soft wndowng S-FDTW has been carred out, wth the Dens Sbragon toolbox for Dgtal Room Equalzaton ([5]) that also provdes C-wrtten functons for homomorphc deconvoluton, frequency dependent rngng truncaton and peak lmtng Tme Fgure 2.Drect front equalzed hgh-passed mpulse response spectrogram Ths processng s usually called Frequency Dependent Tme Wndowng (FDTW) and t s usually used for room equalzaton, where the amount of correcton s drectly proportonal to the length of the mpulse response and nversely proportonal to the consdered frequency band. We call ths knd of processng, where only one pont of transton s appled Hard FDTW. One drawback of ths process s not the temporal sharp transton, whch s avoded by Tukey Wndowng and overlappng, but the tme-frequency sharp transton between an all-pass mpulse response and a low passed Fgure 22. H-FDTW mpulse response spectrogram AES 2th Conventon, Pars, France, 26 May 2 23 Page 3 of 8

14 In the DRC toolbox, frequency dependent wndowng s mplemented followng a band wndowng approach or a sldng low pass flterng approach. The two technques present smlar results, even f the second technques s sad to be more stable and flexble. In both technques two wndows are defned at the frequency range boundary: a longer wndow for low-frequences versons of the mpulse response and shorter wndows for hgher frequency versons of the mpulse response. The frst procedure smply apples a flter bank to the mpulse response and uses dfferent tme wndows for each subband sgnal. These sgnals are then summed back together to obtan the S-FDTW verson of the mpulse response. The second procedure flters the mpulse response wth a low pass flters wth a cutoff frequency that decreases wth the tme wndow length. In the frst procedure the wndow length for each band s computed as W = A* ( F + Q) Where W s the wndow length, F the normalzed frequency, A and Q are chosen n order to ft the boundary wndow length. The parameter named WE s the wndow exponent and determnes the decreasng slope for wndow length reducton. Smaller WE values determnes bggest fracton of the tme frequency plan to be suppressed, as t s shown n fgure 23. WE F c =, WE A* ( W + Q) where F c s the cut-off frequency of the sldng flter. The S-FDTW mpulse response spectrogram s reported n fgure 24. It clearly presents a smoothed transton to the mpulse response tal n the tme-frequency plan. Frequency Tme Fgure 24. S-FDTW mpulse response spectrogram 6. HRIR DIRECT CUSTOMIZATION The two retaned publc ndvdual HRIRs databases are the CIPIC and the IRCAM LISTEN ones. The two databases are brefly presented n the followng. The drect customzaton and the ntegratve drect customzaton module are then descrbed. 6.. CIPIC Database The CIPIC database contans HRIRs measured for 47 subjects. The measurements have been made at the CIPIC Interface Laboratory at the Unversty of Calforna. For each subjects a set of 27 relevant anthropometrc measurements has been provded and two sets, one for the rght ear and one for the left ear of HRIRs sampled on a grd of 75 ponts. Fgure 23. Wndow exponent for S-FDTW In the second procedure The grd does not represent a unform 3D samplng. The measurements are gven for 5 dfferent elevatons and 25 azmuths, followng the conventon showed n fgure 25. θs called azmuth and ts range s [-8, 8], more AES 2th Conventon, Pars, France, 26 May 2 23 Page 4 of 8

15 densely sampled n the mdsagttal [-45, 45] plane; ϕs the elevaton and ts range s [-45,23.625], unformly sampled wth a step of degrees. These values don t match all the measurement postons of our system. In table 3 we show the values of ENST system ponts, and the ones of CIPIC that more match these postons. We can assume that the obtaned HRIR can be wrtten n the form HRTF CIPIC ϕθ = ~ ( f ) H ( f ) H ( f ) HRTF ( f ) ϕθ Where HRTF CIPIC ϕθ ( f ) s the Fourer Transform of the CIPIC HRIR, and ~ CIPIC Hϕθ ( f ) H ϕθ ( f ) = H ϕθ ( f ) s the global CIPIC transfer functon that takes nto account the CIPIC measurement system transfer functon and the CIPIC equalzaton for the real HRIR. ϕθ ϕθ 6.2. IRCAM LISTEN Database Fgure 25. Interaural-polar coordnates system In order to obtan the maxmum precson, we should nterpolate CIPIC HRIRs, or change the measurements ponts n ENST system. In ths phase we just elmnate the [9,5] pont, that presents an unacceptable dfference n azmuth, and use the remanng 5 postons as ndex. We tolerate an error of localzaton of 5 degrees for the azmuth and.875 for elevaton. The measurements have been made wth the Snapshot System by Aureal Semconductors, and Etymotc Research ER-7C probe mcrophones (n closed meatus confguraton). θ, ϕenst θ, ϕcipic,. 3, 3, (9,5 8,6.875) 6,2 65, ,2-65,5,875-6,5-65,6.875 Table 3. Measurement poston comparson The ear channel resonance s not ncluded n the measurements. 6.4-cm dameter Bose Acoustcmass loudspeakers have been used. The measured HRIRs have been wndowed wth a 2 wndow (4.5 ms) n order to elmnate paraste reflectons (even f some floor or knees reflectons are present for some postons). The responses have been compensated wth a perfect equalzer obtaned from free feld measurement at the poston of the center of the head. Some lowfrequency compensaton has been ntroduced (see the documentaton of the CIPIC database for more detals). The IRCAM LISTEN database contans HRIRs measured for 5 subjects. The measurements have been made n the IRCAM anechoc room n Pars. For each subjects a set of 27 relevant anthropometrc measurements have been provded and two sets, one for the rght ear and one for the left ear of HRIRs. The samplng grd conssts of elevaton angles startng at -45 endng at +9 n 5 steps vertcal resoluton. The steps per rotaton vary from 24 to only (9 elevaton). As a whole, there are 87 measurement ponts. The provded ponts match ENST system measurement ponts. The measurement has been carred out wth a Tannoy system 6 loudspeaker, usng the sweep sne method. A par of Knowles FG3329 probe mcrophones has been used n closed meatus confguraton. The measurement technque and equpment are smlar to the one used n the present paper. Dffuse feld equalzaton [3], and wndowng wth a 52 samples wndows have been carred out (see the LISTEN web ste for more nformaton). As we dd wth CIPIC, we defne the global IRCAM transfer functon as ~ IRCAM Hϕθ ( f ) Hϕθ ( f ) = Hϕθ ( f ) DIRECT Customzaton As we dd for the two databases, we defne as ~ ENST Hϕθ ( f ) Hϕθ ( f ) = Hϕθ ( f ) the ENST global transfer functon (that also contans the ear channel resonance due to open meatus measurement). In Drect customzaton we assume that AES 2th Conventon, Pars, France, 26 May 2 23 Page 5 of 8

16 H ENST CIPIC IRCAM ϕθ ( f ) = Hϕθ ( f ) = Hϕθ ( f ) = ( f ), whch s coherent wth the fact that each one of these transfer functon s ntended to be an equalzed verson of the electroacoustc chan transfer functon. Let us suppose to have N measured reference K pont HRTFs and M numbers of HRIRs sets n the database. No dfference between left and rght ears. Drect customzaton follows the scheme presented n fgure 26. A proxmty crteron W between the reference and each one of the databases HRTF sets s obtaned from a metrc characterzng the dstance between two HRTFs. Ths metrc can be defned smply as the cepstral dstance between the two functons, D j = K K k = (log( HRTF ( k)) log( HRTF j ( k))) The proxmty crteron between the measured reference set and the canddate full database set can be defned, for example, as the mean of the metrcs for each reference poston. The mnmum value of the proxmty crteron over all the databases ears provdes the customzed set ndex. Note that wth ths technque we made no dstncton between rght ears and left ears, and that the two customzed sets for the new lstener can come from ears belongng to dfferent subjects n the database INTEGRATIVE Drect Customzaton Drect customzaton s a method to choose, on the bass of a reduced reference HRTF subset, a 3D set of HRTF. The chosen set s then used n ts ntegralty, and the reference HRTFs subset s dscarded. The reference HRTFs are of course the own lstener HRTFs and one could thnk to keep them and just complete the reduced 3D subset, especally f the number of measured personal HRTFs can be consequent. In drect customzaton the global transfer functons are consdered equal to the flat spectrum. Even f ths assumpton s not completely true (see for example the results n secton 5), the customzaton process only means that we are comparng HRTFs measured wth two dfferent systems, and then use only one of them: The lstener would be usng HRTFs that present an 2 addtonal coloraton due to the measurement system used for measurement. On the other sde, ntegratng two HRTFs subsets, ssued from dfferent measurement systems, would represent a problem due to dfferent coloraton determned by equalzaton resduals, whch n general are not the same for the two systems. Ths would lead to dfferent percepton of sound comng from locatons assocated to dfferent subsets, wthout a physcal reason to be. The two systems have to be adapted. To do ths, once the customzed set has been found, a system equalzer s bult. Let us suppose to have the HRTFs ssued from two measurement systems, say A and B, where A s the subset reference measurement systems that provde personal HRTF and B a customzaton database. We can wrte: HRTF HRTF A B A ( f ) = H ( f ) HRTF ( f ), and ϕθ B ( f ) = H ( f ) HRTF ( f ) =:N. ϕθ The nverse flter obtaned from B s ϕ ( f ) =. B Hϕθ ( f ) HRTF ( f ) Applyng t to the measured reference HRTFs we obtan H A Φ ( f ) = HRTF ( f ) ϕ ( f ) =, and H HRTF B A ϕϑ B ϕϑ A ( f ) Φ ( f ) = H HRTF ( f ). We call Φ the database equalzer: applyng Φ to the database customzed (sub)set makes the two subsets homogeneous and then ntegraton becomes possble. Usng the database equalzer the ear channel resonance s corrected, too. We have to note that Φ depends on the measurement poston, because measurement systems present loudspeaker-to-mcrophone transfer functons that are varable (see secton 3). To obtan a sgnfcant mean database equalzer, we can use ϕϑ N Φ( f ) = ϕ ( f ). N AES 2th Conventon, Pars, France, 26 May 2 23 Page 6 of 8

17 or frequency smoothng technques, as we dd for electroacoustc chan equalzaton. 2 j M DATABASE REFERENCES 2 N W W W W 2 = = = N N N D = N D 2 = j D j N = = Fgure 26. Drect customzaton prncple 7. CONCLUSIONS AND FUTURE WORKS N N M D M N = mn W j j Ths paper presents a rapd measurement system for ndex HRIR to use for drect customzaton. The acoustc characterstcs of the system have been descrbed wth detal. The post processng module for mpulse response extracton, equalzaton and frequency-dependent wndowng has been tested n J order to obtan some nformaton about the performance of acoustcal channel correcton. The reported results show that the smoothed equalzed electroacoustc transfer functons present.94 mean devaton from the flat spectrum and max devaton of 4.84 db;.75 and 2.8 db are the correspondng values for drect front equalzaton and soft frequency dependent tme wndowng. These values let us thnk that measured HRIRs can be consdered as representatve of the only scatterng by the head and the torso of the lstener, and provde bnaural cues for bnaural synthess. The utlzaton of these measured HRIRs as ndex for drect and ntegratve drect customzaton has been descrbed. Localzaton and externalzaton qualty of measured HRIRs have been proved wth some nformal tests on 4 subjects. In a future work we wll carry out a measurement campagn wth several subjects, and a perceptual tests campagn. The am s to make clearer the followng ponts:. The perceptual value of Equalzaton. 2. The perceptual value of low frequency resoluton enhancement. 3. The perceptual value of drect customzaton, by drect n stu comparson of the real playback and the bnaural playback wth the measured HRIRs and the customzed HRIRs. 8. REFERENCES [] D.N.Zotkn, R.Duraswam, L.S.Davs, Customzable Audtory Dsplays, Proc. 22 Internatonal Conference on Audtory Dsplays, Kyoto, Japan, July 2-5, 22. [2] V. R. Algaz, R. O. Duda, D. M. Thompson and C. Avendano, The CIPIC HRTF Database, Proc. 2 IEEE Workshop on Applcatons of Sgnal Processng to Audo and Electroacoustcs, pp. 99-2, Mohonk Mountan House, New Paltz, NY, Oct. 2-24, 2. [3] AES 2th Conventon, Pars, France, 26 May 2 23 Page 7 of 8

18 [4] B. Katz, Boundary element method calculaton of ndvdual head-related transfer functons. I Rgd model calculaton", JASA, (5) November 2. [5] J.C Mddlebrooks, E.A. Macpherson. Z.Onsan, Psychophyscal customzaton of drectonal, transfer functons for vrtual sound localzaton, JASA, 8(6), December 2. [6] C. Brown, O. Duda A structural model for Bnaural Sound Synthess, IEEE Transactons on speech and audo processng, 6(5), September 998 [7] C.Jn, P.Leong, J. Leung, A. Corderoy, S.Carlle, Enablng ndvdualzed vrtual audtory space usng morphologcal measurements, Proc. Frst IEEE Pacfc-Rm Conf. on Multmeda, 2 [8] C. Fahn, Y.C. Lo, On the clusterng of Head- Related-Transfer Functons used for 3-D Sound Localzaton, Journal of Informaton Scence and Engneerng 9, 4-57 (23) [9] A. Farna, Smultaneous Measurement of Impulse Response and Dstorton wth a Swept Sne Technque, AES 8 th Conventon, 2 February 9-22, Pars, France [] D. Hammersho, H. Moller, Sound transmsson to and wthn the human ear canal, JASA, (), July 996. [] P. Hatzantonou, J.N Mourjopoulos, Generalzed Fractonal-Octave Smoothng of Audo and acoustc Responses, JAES, 48(4), 2, Aprl [2] O.Krkeby, P.Rubak, A. Farna, Desgn of Cross- Talk Cancellaton Networks by usng fast deconvoluton, 6 th AES Conventon, Munch, 8- May 999. [3] J.M.Jot, V.Larcher, O.Warusfel, Dgtal sgnal processng ssues n the context of bnaural and transuaral stereophony, 98 th AES Conventon, Pars, France, Feb , 995, preprnt no [4] [5] AES 2th Conventon, Pars, France, 26 May 2 23 Page 8 of 8