Localization Performance of Real and Virtual Sound Sources

Transcription

1 M.Sc.E.E. Jan Abildgaard Pdrsn AM3D A/S Riihimäkivj 6 DK-9200 Aalborg Dnmark M.Sc.E.E. Torbn Jørgnsn Trma A/S Hovmarkn 4 DK-8520 Lystrup Dnmark jap@am3d.com / toj@trma.dk ABSTRACT This papr dscribs how a 3D-Audio systm for us in fightr aircrafts was valuatd in an xprimnt, by comparing localization prformanc btwn ral and virtual sound sourcs. Virtual sound sourcs from 58 slctd dirctions wr valuatd, whil 16 of ths dirctions wr also valuatd using ral sound sourcs, i.. loudspakrs. 13 pilots from th Royal Danish Air Forc and 13 civil prsons wr usd in th tst. Th localization prformanc was split into a constant and a stochastic diffrnc btwn th prcivd dirction and th dsird dirction (stimulus). Th constant diffrnc is a localization offst and th stochastic diffrnc is a masur for th localization uncrtainty. Stimuli lngth of both 250 ms and 2 s nabld invstigation of th importanc of had movmnts, i.. using had tracking. Ral and virtual sound sourcs could b locatd with an uncrtainty of 10o and 14o dgrs for azimuth whil th uncrtainty for lvation was 12o and 24o (ral and virtual sound sourcs). No significant localization offst was found for azimuth, whil an avrag offst for lvation of 3o 6o dgrs was found using long stimuli. A significant diffrnc btwn th localization offst obtaind in diffrnt dirctions was found spcially for lvation, whr th offst was found to hav a strong corrlation to th stimuli lvation. 1 INTRODUCTION 3D-Audio is usd in fightr aircrafts to nhanc th situational awarnss. This includs 3D-Audio indication of th dirctions of dtctd approaching missils as wll as channl sparation of svral simultanous sound signals,.g. spch from ground control, wingman and alarm signals. Th 3D-Audio systm includd hadphon playback systm, had trackr and a digital signal procssor (DSP). This papr dscribs th dsign and th rsults of a psychoacoustic tst, which was usd to valuat such a 3D-Audio systm. In particular th localization prformanc was valuatd, i.. th capability of positioning sound sourcs in 3D spac at prdtrmind positions, which nabls listnrs to localiz sound sourcs at dsird positions with som uncrtainty. Th xprimnt focusd only on th localization prformanc in rlation to dirction - distanc was not considrd. Sparat work focusing on distanc prcption has bn prsntd arlir by othr authors [1], [2], [3]. Localization prformanc has bn a subjct for rsarch for many yars and a long list of authors has prsntd work including localization prformanc of ral sound sourcs, phantom sound sourcs and Pdrsn, J.A.; Jørgnsn, T. (2005) Localization Prformanc of Ral and Virtual Sound Sourcs. In Nw Dirctions for Improving Audio Effctivnss (pp ). Mting Procdings RTO-MP-HFM-123, Papr 29. Nuilly-sur-Sin, Franc: RTO. Availabl from: RTO-MP-HFM

2 virtual sound sourcs [4], [5], [6], [7], [8], [9], [10], [11]. Many of ths paprs focus on th localization prformanc of virtual sound sourcs, i.. binaural sound rproduction systms, whr tst subjcts war hadphons and th signals fd to ach ar has bn rcordd using a dummy had or has bn procssd by had-rlatd transfr functions (HRTF) [5]. Eithr th rcording or th procssing will nabl listnrs to localiz sound sourcs in full 3D as if th listnr was prsnt at th rcording position or at th synthsizd listning position. Had tracking nabls th signals fd to th two ars to chang accordingly to had movmnts, so that th prcption of th virtual sound sourc rmains whil moving th had. Th prcption of th virtual sound sourc not only rmains bcaus of th had tracking systm, it is improvd dramatically [5], [10], [11]. Espcially front-back ambiguity is almost solvd in 3D-Audio systms, which includs a had trackr [11]. Whn valuating th localization prformanc of 3D-Audio systms diffrnt stratgis has bn proposd. On way is to dtrmin th diffrnc btwn th dirction of stimuli (dsird dirction) and th prcivd dirction. This will show how larg th rrors can b whn trying to position a virtual sound sourc in 3D spac, i.. an absolut masur of th angl btwn th dsird and obsrvd position. In a givn application this could b compard to th actual nds or rquirmnts for localization prformanc. Howvr an obvious rfrnc would b ral lif localization prformanc, i.. localization prformanc of ral sound sourcs [9]. Most 3D-audio systms ar implmnting an approximation to a ral sound fild, which mans that if th localization prformanc of th 3D-Audio systm is approximatly th sam as for ral sound sourcs, thn th localization prformanc of th 3D-Audio systm is optimal. Th xprimnt dscribd in this papr includd a listning tst, whr th sam tst subjcts valuatd th localization prformanc of both a 3D-audio systm and a stup of ral sound sourcs. Car was takn to nsur that all conditions wr qual for th two valuations. Th only diffrnc was th fact that tst subjcts did war hadphons whn thy valuatd th 3D-Audio systm, which was not th cas whn thy valuatd th ral sound sourcs. 2 STRATEGY FOR THE EXPERIMENT 2.1 Purpos of th xprimnt Th purpos of this xprimnt was to provid information about th localization prformanc of a 3D- Audio systm in rlation to usag in Mission Critical systms. Spcifically th prformanc masurs for systms, whr alarm signals ar givn a dirction corrsponding to objcts that nd attntion, wr of intrst. It was a dsir to dtrmin th localization prformanc in situations that rsmbl th ral lif us of an auditory display for Mission Critical systms. Th localization prformanc for virtual sourcs was compard to th localization prformanc for ral sound sourcs. In this xprimnt hadphons wr usd for virtual sound sourcs and loudspakrs wr usd for ral sound sourcs. 2.2 Psychoacoustic mthod First th subjcts hard on sound signal from on dirction ithr via hadphons (virtual sourcs) or via on loudspakr (ral sourcs) and wr thn askd to point to th dirction, whr thy prcivd th sound sourc, i.. th singl stimuli mthod. Simply pointing to th dirction whr thy prcivd th sound sourc to com from was an intuitiv mthod, which should dcras th nd for vry long and intnsiv training sssions. A lasr pointr was mountd on th pointing dvic in such a way, that a light dot gav visual fdback to th subjct to nsur, that th answrd dirction was indd th dirction, whr thy prcivd th sound 29-2 RTO-MP-HFM-123

3 to originat from. Diffrnt and indpndnt random squncs of dirctions and sound sourc typ (ral or virtual) wr usd for ach subjct. Th main xprimnt was prcdd by a structurd training sssion, which introducd th tst subjcts to th concpt of dirction of sound sourcs as wll as running tst sssions using both virtual and ral sound sourcs. Th aim of th training sssions wr to minimiz th incrasd varianc introducd by using subjcts, who wr naiv in th task, mthods and procdurs usd in th xprimnt. Visibility of th ral sound sourcs, i.. loudspakrs, would hav introducd a bias into th xprimnt, bcaus subjcts would only point to dirctions, whr thy could s a loudspakr not in dirctions in btwn. This bias was avoidd by installation of an acoustically transparnt curtain (sid, top and bottom) surrounding th subjcts in such a way that no loudspakrs wr visibl. 2.3 Localization Prformanc and dfinition of dirction Th localization prformanc was split into a constant and a stochastic diffrnc btwn th prcivd dirction and th dsird dirction, i.. stimulus. Th constant diffrnc is a localization offst and th stochastic diffrnc is a masur for th localization uncrtainty. A givn dirction rlativ to th had of th listnr was charactrisd by th two Eulr angls: azimuth and lvation. Azimuth dfins a rotation about th z-axis in a coordinat systm, whr origo is th cntr of th listnrs had, z-axis pointing upward and y-axis pointing straight forward, i.. x-axis pointing to th right hand sid. Azimuth is 0 for a dirction straight ahad, i.. in th dirction of th y-axis. Azimuth is positiv whn turning in th dirction to th lft (countr clockwis),.g. azimuth is +90 o for a dirction dirctly to th lft of th listnr and 270 o for a dirction dirctly to th right of th listnr. Elvation is a rotation around th x-axis, whr a positiv angl mans upward and a ngativ angl mans downward rlativ to th horizontal plan for th listnrs had,.g. +90 o is right abov th listnrs had and -90 o is dirctly blow th listnr. Elvation is 0 o in th horizontal plan. 2.4 Virtual and ral sound sourcs Th prcivd dirction of a sound will b diffrnt from th dsird/ral dirction of th prsntd sound. This is tru both whn using virtual and ral sound sourcs. This mans, that whn judging th localization prformanc of a 3D-Audio systm (virtual sourcs), on could compar th rsults to rsults obtaind using ral sound sourcs placd in th dsird dirction rlativ to th subjct. A larg numbr of dirctions wr usd for virtual sound sourcs bcaus th 3D-Audio systm was th main focus of th xprimnt. Azimuth and lvation wr considrd to b two indpndnt variabls in th xprimnt, which mant that sparat slction of valus for azimuth and lvation wr prformd. At first 107 dirctions wr considrd for us in th xprimnt for virtual sourcs, howvr du to lft/right symmtry of th basic st of dirctions this numbr was lowrd to 58 asymmtric dirctions, which basically covrd th sam dirctions as th basic st. This dcision was mad to lowr th tim consumption of th xprimnt and bcaus vrification of lft/right symmtry was not part of th purpos of this xprimnt. For practical rasons, a significant lowr numbr of ral sound sourcs wr usd: 16 loudspakrs. Ths 16 dirctions for th ral sourcs wr a subst of th 58 dirctions usd for virtual sourcs. A tst sssion was ithr a loudspakr sssion or a virtual sourc sssion, so that tst subjcts did not hav to put on hadphons or rmov thm during a tst sssion. RTO-MP-HFM

4 2.5 Controlld paramtrs in th xprimnt Th controlld paramtrs of th xprimnt wr: Azimuth angl Elvation angl Tst subjct Stimuli Sound sourc typ: virtual or ral 3 DESIGN OF THE EXPERIMENT 3.1 Stimuli Spctral and tmporal charactristics Whit nois was slctd as th basic stimuli. Both short and long bursts wr usd: 250 ms and 2 s. Short bursts of whit nois wr usd to prvnt subjcts from bing abl to turn thir had as part of localizing th sound sourcs. This stimulus was usd to valuat th static localization prformanc. Earlir work has provd that th had movmnt has a dramatic impact on th localization prformanc, i.. th diffrnc btwn static and dynamic sound localization prformanc [5], [10], [11], [13]. Dynamic localization prformanc, whr thn valuatd using th long nois bursts, whr subjcts had tim to us had movmnts as part of th localization task. During th training sssions only long bursts wr usd. If a tst sssion lastd longr than 15 min. a brak was automatically insrtd by th program, which controlld th tst. Aftr a brak of approximatly 5 min. th tst sssion was continud. Training sssions: Long bursts of whit nois: 2.00 s. Exprimnt sssions: Short bursts of whit nois: 0.25 s. Long bursts of whit nois: 2.00 s. Maximum 15 min. without braks Th lvl of th stimuli was adjustd to 84 db SPL(LIN) = 75 db SPL(A) at th listning position whil playing th nois signal through a loudspakr. A dummy had, Valdmar [5], [6], [7], wr placd on th listning position and usd to nsur th sam rproduction lvl using loudspakrs and hadphons RTO-MP-HFM-123

5 3.1.2 Dirctions of virtual and ral sound sourcs Azimuth and lvation had to b indpndnt sinc thy wr rgardd as two controlld paramtrs of th xprimnt. Basically it was dcidd to us 15 valus for azimuth and 9 valus for lvation. This corrsponds to th 107 dirctions, bcaus azimuth ar undfind for th two lvation valus +90 o and - 90 o.th valus for azimuth angl wr 0 o, 24 o, 48 o, 72 o, 96 o, 120 o, 144 o, 168 o, 192 o, 216 o, 240 o, 264 o, 288 o, 312 o and 336 o. Th valus for lvation angl wr -90 o, -66 o, -44 o, -22 o, 0 o, 22 o, 44 o, 66 o and 90 o. Ths 107 dirctions wr thn rducd to 58 simply by rmoving ithr th lft or th right dirction in a coupl, which wr lft/right symmtric. Th usd 58 dirction for virtual sound sourcs is shown in figur 3.1. Figur 3.1: Dirctions for th 58 virtual sound sourcs without lft/right symmtry. This mans that th numbr of valus usd for azimuth rducd to 8: Azimuth: 0 o, 24 o, 72 o, 120 o, 168 o, 216 o, 264 o, 312 o Elvation: -90 o, -66 o, -44 o, -22 o, 0 o, 22 o, 44 o, 66 o, 90 o 16 of th 58 dirctions wr thn slctd for ral sound sourcs, i.. loudspakrs wr mountd in ths 16 dirctions. Figur 3.2 shows ths 16 dirctions in a similar way as in figur 3.1. Finally tabl 3.1 givs th complt list of th 58 dirctions including Dirction ID, lvation angls, azimuth angls and an indication of Ral Sourc, which mans that such a dirction wr usd both as a virtual and a ral sound sourc. RTO-MP-HFM

6 Figur 3.2: Dirctions for th 16 ral sound sourcs (loudspakrs) shown as rd squars, which ar a subst of th 58 dirctions usd for virtual sourcs (blu disks) RTO-MP-HFM-123

7 Tabl 3.1: Dirction of all 58 virtual and 16 ral sound sourcs. Dirction ID Elvation Azimuth Ral Sourc 1-90 o 0 o o 0 o X 3-66 o 24 o o 72 o o 120 o X 6-66 o 168 o o 216 o o 264 o X 9-66 o 312 o o 0 o o 24 o o 72 o X o 120 o o 168 o o 216 o X o 264 o o 312 o X o 0 o X o 24 o o 72 o o 120 o o 168 o X o 216 o o 264 o o 312 o o 0 o X 27 0 o 24 o X 28 0 o 72 o o 120 o X 30 0 o 168 o o 216 o o 264 o X 33 0 o 312 o o 0 o o 24 o o 72 o X o 120 o o 168 o o 216 o o 264 o o 312 o o 0 o o 24 o o 72 o o 120 o o 168 o o 216 o o 264 o o 312 o X o 0 o o 24 o o 72 o o 120 o o 168 o o 216 o X o 264 o o 312 o o 0 o X RTO-MP-HFM

8 3.1.3 Procssing signals to crat virtual sound sourcs Signals for right ar and for lft ar wr convolvd in ral tim with th Had-Rlatd Transfr Functions (HRTF) which corrspondd to th dirction of th dsird sound sourc position rlativ to th orintation and position of th listnrs had. This was nabld by fitting th hadphon with a Had Trackr, which continuously did snt th positions and orintations to th program, which was prforming procssing of th signals. Th usd HRTF databas did hold masurmnts from diffrnt dirctions, i.. 2 o angular rsolution [5]. Th complt ral tim 3D-Audio systm including HRTF databas, had trackr, convolution, binaural rvrbration tc. was a product dvlopd by AM3D A/S [12]. Th procssing also includd qualization for th transfr function of th usd hadphon [5]. A calibration of th Had Trackr was prformd bfor th start of ach tst sssion. 3.2 Physical stup Rproduction systm for ral sound sourcs Th physical stup consistd of 16 ral sound sourcs, i.. loudspakrs, positiond in a distanc of 2.13 m (7 ft) from a listning position, s figur 3.3. A curtain cylindr surroundd th listning position to avoid visibility of th loudspakrs. Th light in th room was switchd off during all tsts and light was shind on th curtain from insid th cylindr, which nsurd no visibility of objcts outsid th cylindr. Th curtain cylindr was fittd both with a circular top and a circular bottom. Th hight of th curtain cylindr was 3.70 m and th diamtr was 2.90 m. Th tst subjcts wr standing on an adjustabl platform, which lvatd th ars of all subjcts to a hight of 2.15 m abov floor lvl. Figur 3.3 shows th usd lvation angls, whil figur 3.4 shows a top viw, whr th usd azimuth angls ar shown. Th physical stup was stablishd in a larg Tlvision Studio at th facilitis of AM3D A/S. Th adjustabl platform masurd 30 cm by 30 cm, which constraind th possibility for tst subjcts to mov away from th rfrnc listning position. Howvr th subjcts wr allowd to turn thir whol body around in ordr to localiz th sound and for prcis pointing toward th position, whr thy prcivd th sound to originat from. Th sound prssur rsponss of all 16 loudspakrs wr masurd at th listning position, and individual qualization wr applid to ach loudspakr to nsur vry similar rspons from all 16 loudspakrs. Th rsulting amplitud rsponss of all th loudspakrs wr flat in th rang from 140 Hz to 15 khz RTO-MP-HFM-123

9 Ø2.90 m h=3.70 m Listning position 0.30 m Floor Figur 3.3: Sid viw of th physical stup, whr 16 loudspakrs wr positiond around a listning position in a distanc of 2.13 m. A curtain cylindr (gry) surroundd th listning position. Diamtr of th cylindr was 2.90 m and th hight was 3.70 m. Both th top and bottom of th cylindr was fittd with curtain disks to clos up th cylindr. Tst subjcts wr standing on a platform (grn) to nsur an ar hight of 2.15 m. Figur 3.5 shows th complt stup including th curtain cylindr and th loudspakrs mountd around it. This pictur was takn from th lft hand sid, i.. from th dirction, whr azimuth was +90 dgrs. It follows from this, that th loudspakrs sn in th lft sid of th pictur, wr th loudspakrs dirctly in front of th listnr, i.. in th dirction, whr azimuth was 0 dgrs. RTO-MP-HFM

10 Lft Back Front 2X Right Figur 3.4: Top viw of th physical stup, whr 16 loudspakrs wr positiond around a listning position in a distanc of 2.13 m. A curtain disk (gry) forms th two nds of th closd cylindr. Th diamtr of th disks was 2.90 m. Thr loudspakrs wr pointing almost upwards (lvation = -66 o ), which is indicatd by th upward facing loudspakrs. Two Loudspakrs wr pointing downwards or almost downwards, which is indicatd by th downward facing loudspakrs. All loudspakrs wr positiond in a distanc of 2.13 m from th listning position, but show up at diffrnt projctd distancs in this 2D top viw diagram du to diffrnt valus of lvation angl for th 16 diffrnt loudspakrs. Figur 3.5: Pictur of th curtain cylindr with loudspakrs mountd around it RTO-MP-HFM-123

11 3.2.2 Rproduction systm for virtual sound sourcs Th rproduction stup for virtual sourcs consists of signal procssing insid th 3D-Audio systm and th tst subjct waring hadphons, which wr quippd with a had trackr. Basd on th position and orintation of th tst subjcts had combind with th dsird dirction of th virtual sound sourc digital filtrs wr applid to th short or long whit nois burst as dscribd in sction Th slctd hadphon was a Byrdynamic DT 990 PRO, which has a spcifid frquncy rang of 5 Hz 35 khz. Th had trackr updatd th position and orintation of th hadphons with a rat of 60 Hz. Th maximum latncy was 35 ms. Each hadphon sssion startd with a calibration of th had trackr by instructing th subjct to look straight forward towards a fixd point on th curtain, which was in th sam hight as th ys of th subjct and in th dirction dfind as zro azimuth Pointing dvic Th tst subjcts had to show thir prcivd dirction by pointing. For this a toy gun quippd with a tracking dvic was usd. Th tst subjcts wr instructd to point by aiming at a targt using both hands and straight arms. This nsurd that th toy gun was hld in a position dirctly in front of th tst subjcts, i.. not of th ithr right or lft hand sid. Th tst subjct wr also instructd to turn thir whol body as part of th localization task and whn pointing thy should b facing th position, whr thy prcivd th sound to originat from. This should prvnt subjct from pointing and answring in an imprcis way,.g. by shooting ovr thir shouldr to point to a position bhind thm. Subjcts wr instructd to notic whr thy wr pointing by obsrving whr th lasr dot shind on th curtain. Th updat rat of th trackr mountd on th toy gun was also 60 Hz, and this trackr was calibratd at th start of ach sssion by instructing th tst subjct to point and aim at th sam fix point on th curtain as dscribd in sction This was prformd at th sam tim as th calibration of th had trackr Automatd tst systm Prsntation of stimuli in th prdtrmind dirctions, adjusting th signal procssing according to th position and orintation of th had trackr, prforming all th ncssary signal procssing and rcording th answrd dirctions by rading th data from th gun trackr was all prformd by an intgratd systm, which was build around a PC including tst systm softwar and signal procssing softwar, had trackr control and powr amplifirs. Figur 3.6 shows th structur of this systm. Th systm includd a usr intrfac which was usd by th tst oprator to control th tst. RTO-MP-HFM

12 Storag Audio Gnrator Tst data Control data: Start, Stop. Tst Excution Application Global Virtual Sound Sourc Position 3D-Audio procssing softwar Pointing Dvic Trackr Hadphon Amplifir Global Had position and orintation Hadtrackr Figur 3.6: Structur of th automatd tst systm 3.3 Procdur for th xprimnt Slction of tst subjcts Th 3D-Audio systm undr tst is targtd towards applications in fightr aircrafts. This ld to th dsir to tst if fightr pilots hav spcial rquirmnts for a 3D-Audio systm. In practis this was nsurd by rcruiting half of tst subjcts from th Royal Danish Air Forc: 13 pilots from th Air Forc Bass in Aalborg and Karup, Dnmark. Th othr half of tst subjcts was civil prsons, primarily studnts from Aalborg Univrsity, Dnmark. This ld to a total numbr of tst subjcts of 26 prsons Scrning and instruction of tst subjcts All tst subjcts had a tst of haring thrshold, i.. audiogram. All tst subjcts wr nsurd to fall within a rang of haring loss btwn +20 db to -10 db, basd on th standard haring thrshold (ISO 389). Thn th tst subjcts had thir pictur takn, dtrmination of lading y, collction of prsonal facts (nam, ag, sx, xprinc in listning tsts, ar hight abov th floor and total hight) and thy had to rad a 2 pag instruction, which dscribd thir task in th xprimnt. A short discussion btwn th tst oprator and th tst subjct should minimiz th risk of misundrstandings. Th hight of th platform in th bottom of th curtain cylindr was thn adjustd according to th masurd hight of th ars of th tst subjcts to nsur that th hight of th ars whn standing on th platform was 2.15 m in all cass. Th subjcts wr not prmittd to viw th loudspakrs in th stup prior or during th xprimnt RTO-MP-HFM-123

13 3.3.3 Tst sssions Th whol xprimnt was split into 6 tst sssion for ach tst subjct: Hadphon familiarization (training: 16 dirctions, 3 rptitions, long stimuli) Virtual sound sourcs, 2 s stimuli Virtual sound sourcs, 250 ms stimuli Loudspakr familiarization (training, 16 dirctions, 3 rptitions, long stimuli) Ral sound sourcs, 2 s stimuli Ral sound sourcs, 250 ms stimuli Th ordr of ths 6 sssions was mad random for ach prson by random slction of 1 out of 4 possibl squncs: Squnc A: Virtual sound sourcs first 2 s stimuli first Squnc B: Virtual sound sourcs first 250 ms stimuli first Squnc C: Ral sound sourcs first 2 s stimuli first Squnc D: Ral sound sourcs first 250 ms stimuli first Following this trminology thn squnc A is similar to th squnc givn in th start of sction Each sound sourc dirction was usd 3 tims for ach subjct using both hadphons and loudspakrs. Th total numbr of stimuli and answrs wr: Hadphon familiarization: 26 prsons x 16 dirctions x 3 rptitions = 1248 stimuli Virtual sourcs: 26 prsons x 58 dirctions x 3 rptitions x 2 stimuli lngth = 9048 stimuli Loudspakr familiarization: 26 prsons x 16 dirctions x 3 rptitions = 1248 stimuli Ral sourcs: 26 prsons x 16 dirctions x 3 rptitions x 2 stimuli lngth = 2496 stimuli Total: stimuli, which was qual to 39 hours of ffctiv tst (approx. 10 sc/stimuli) This was qual to 1.5 hour of ffctiv tsting tim for 1 tst subjct, but scrning, instruction and pauss ld to a total tsting tim for 1 tst subjct of approx. 2.5 hours. Th automatic control softwar insrtd a paus in any tst sssion aftr ach 15 min. of tsting tim. Th 2.5 hour for 1 tst subjct corrspondd to a total tsting tim of 65 hours On tst sssion Th procdur for any on sssion was: Th tst subjct was positiond in th rfrnc position: standing on th platform in th cntr of th curtain cylindr facing th fix point on th curtain, which indicatd zro azimuth and zro lvation. Th subjcts calibratd both th had trackr and th gun trackr by aiming with two-handd grip and straight arms towards th fix point on th curtain and by looking straight ahad towards th sam fix point. Th calibration was ffctuatd by prssing th triggr on th toy gun. In th cas of ral sound sourcs thn only th gun trackr was calibratd. RTO-MP-HFM

14 Bfor ach stimulus, th tst subjct rportd Rady to th systm by prssing th triggr aftr rturning to th rfrnc position facing th fix point. Each stimulus signal was prsntd Answrd dirction was givn by th subjct through aiming with two-handd grip and straight arms toward th prcivd dirction and prssing th triggr. 4 RESULTS 4.1 Basic data analysis Th localization rror, calculatd as th diffrnc btwn th answrd dirction and th dsird dirction, was rgardd as a random variabl. Th distribution was assumd to b a Normal distribution. 2 Unbiasd stimators for man (avrag), µ, and varianc, σ, wr calculatd from th raw dirctional data obtaind, and standard dviation,σ, of th rror was simply calculatd as th squar root of th stimatd varianc [14]. A 95% confidnc intrval cntrd at th stimatd man (avrag) having a width qual to 3.92 tims th standard dviation was thn dtrmind as a modl dscribing th basic data [14]. Th basic data analysis was prformd using data from all 26 tst subjcts but limitd to 1 spcific dirction, 1 lngth of stimuli and 1 spcific rproduction mthod. Data from virtual sound sourcs, 2 s stimuli in th dirction ID=28 ar plottd in figur 4.1. Dirction ID=28 is locatd in th horizontal plan, i.. lvation=0 o. Th azimuth for th dirction is 72 o (to th lft rlativ to front dirction). For such a spcific subst of data 78 answrs wr obtaind (26 prsons x 3 rptitions). Figur 4.1: 78 answrs for th dirction indicatd by th black horizontal and vrtical lins. Each answr is indicatd by a blu circl. Data wr obtaind using virtual sourcs and 2 s stimuli RTO-MP-HFM-123

15 Th azimuth and lvation valus for ach answr wr thn mad rlativ to th stimuli dirction by subtracting th stimuli azimuth and lvation. Figur 4.2 shows ths rlativ answrs, i.. th localization rrors. Th rd rctangl indicats 95% confidnc intrvals for azimuth and lvation, whil th avrag azimuth rror, µ a, and avrag lvation rror, µ, ar indicatd by black horizontal and vrtical lins. Th statistical paramtrs for ths data ar givn in tabl 4.1. Figur 4.2: Localization rror calculatd from th 78 answrs obtaind using virtual sourcs and 2 s stimuli in dirction 28 (azimuth=72 o,lvation=0 o ). Th rd rctangl indicats 95% confidnc intrval for azimuth and lvation rrors. Each localization rror is indicatd by a blu circl. Black horizontal and vrtical lins indicat th global avrag azimuth and lvation rrors. Tabl 4.1: Statistical charactristics for th localization rror obtaind in th dirction ID=28 using virtual sourcs and 2 s stimuli. Azimuth Elvation Avrag rror ( µ ) -0.1 o 2.6 o Standard dviation (σ ) 6.0 o 6.5 o 95% confidnc intrval [-11.9 o ; 11.6 o ] [-10.1 o ; 15.4 o ] Equations 4.1 and 4.2 giv th stochastic modl for th localization rror both for azimuth and lvation. rror rror azimuth lvation 2 N( µ, σ ) (4.1) a a 2 N( µ, σ ) (4.2) RTO-MP-HFM

16 4.2 Dirction offst and uncrtainty across all dirctions Th analysis procss dscribd in sction 4.1 was rpatd for all 58 dirctions usd for virtual sourcs. Figur 4.3 shows th avrag azimuth rror, µ, for all 58 dirctions, which shows that th avrag a rror, µ, must b modlld as a stochastic variabl itslf. This stochastic variabl is also assumd to b a a Normal distribution. Th black lin indicats th ovr all avrag for azimuth rror, i.. a global avrag azimuth rror, µ, and th 95% confidnc intrval for th avrag azimuth rror is indicatd by rd horizontal lins. a Figur 4.3: Avrag azimuth rror calculatd for th 58 dirctions usd for virtual sound sourcs indicatd by blu circls. Th basic data wr obtaind using virtual sourcs and 2 s stimuli. Th black horizontal lin indicats th global avrag azimuth rror, and th rd horizontal lins indicat th 95% confidnc intrval for th avrag azimuth rror. Equation 4.3 givs th stochastic modl for th distribution of avrag azimuth rror, µ a, at diffrnt dirctions, as shown in figur 4.3. Th standard dviation of th avrag rror is trmd: σ avrag, a. Th global avrag azimuth rror, µ a, global, was calculatd as givn in quation 4.4 basd on th avrag azimuth rror, µ (i a ), in all 58 dirctions. 2 µ a N( µ a, global, σ avrag, a ) (4.3) 1 µ (4.4) 58 = ) a, global µ a ( i 58 i= 1 Th localization rror can now b modlld as th sum of two stochastic variabls, both Normal distributd, which in turn yilds on Normal distributd stochastic variabl with an incrasd varianc. Equation 4.5 thn givs th total stochastic modl for th localization rror, whr σ is th total a, total RTO-MP-HFM-123

17 standard dviation of th azimuth rror, which can b calculatd using quation 4.6 [14]. This was don by assuming that th stochastic lmnt within isolatd dirctions was indpndnt from th stochastic lmnt btwn dirctions [14]. rror azimuth 2 = N( µ, σ ) (4.5) a, global a, total σ = σ + σ (4.6) a, total avrag, a a Figur 4.4 shows th avrag lvation rror for th 58 dirctions usd for virtual sourcs, 2 s stimuli. Hr it should b notd, that th avrag rror sms to corrlat to th dirction ID. In fact th avrag lvation rror was found to b a function of th stimuli lvation and th global avrag lvation rror, µ,global. Figur 4.5 shows th rlationship btwn th avrag lvation rror, µ, th stimuli lvation, S, and th global avrag lvation rror, µ, global, which is also dscribd in Equation 4.7. o µ = α S + µ, global = 0.25 S (4.7) α is an Elvation rror factor, which dscribs how larg a fraction of an incras in stimuli lvation is rflctd in an incrasd lvation rror, i.. largr stimuli lvation yilds largr lvation rror. Figurs 4.6 and 4.7 show th standard dviations for azimuth rrors and lvation rrors as a function of stimuli dirction ID. Figur 4.8 shows how th standard dviations for lvation rrors incrasd whn stimuli movd away from th horizontal plan (lvation=0 o ). Figur 4.4: Avrag lvation rror calculatd for th 58 dirctions usd for virtual sound sourcs indicatd by blu circls. Th basic data wr obtaind using virtual sourcs and 2 s stimuli. Th black horizontal lin indicats th global avrag lvation rror, and th rd horizontal lins indicat th 95% confidnc intrval for th avrag lvation rror. RTO-MP-HFM

18 Figur 4.5: Avrag lvation rror calculatd for 58 dirctions plottd as a function of stimuli lvation. Th basic data wr obtaind using virtual sourcs and 2 s stimuli. Th black straight lin indicats a simpl rlationship btwn avrag lvation rror and stimuli lvation. Figur 4.6: Standard dviations for azimuth rror calculatd for th 58 dirctions usd for virtual sound sourcs indicatd by blu circls. Th basic data wr obtaind using virtual sourcs and 2 s stimuli. Th black horizontal lin indicats th avrag standard dviation across dirctions RTO-MP-HFM-123

19 Figur 4.7: Standard dviations for lvation rror calculatd for th 58 dirctions usd for virtual sound sourcs. Th basic data wr obtaind using virtual sourcs and 2 s stimuli. Th black horizontal lin indicats th avrag standard dviation across dirctions. Figur 4.8: Standard dviations for lvation rror calculatd for 58 dirctions plottd as a function of stimuli lvation. Th basic data wr obtaind using virtual sourcs and 2 s stimuli. Th black horizontal lin indicats th avrag standard dviation across dirctions. RTO-MP-HFM

20 Figur 4.9 shows all 4524 answrs obtaind using virtual sound sourcs and 2 s stimuli. Th black horizontal and vrtical lins show th global avrag for azimuth rror, µ, and lvation rror, a, global µ,global. Th rd rctangl indicats 95% confidnc intrvals for azimuth and lvation, which ar basd on th statistical paramtrs givn in tabl 4.2. Figur 4.9: Localization rror calculatd from 4524 answrs obtaind using virtual sourcs and 2 s stimuli in all 58 dirctions. Th rd rctangl indicats 95% confidnc intrval for azimuth and lvation rrors. Each localization rror is indicatd by a blu circl. Black horizontal and vrtical lins indicat th global avrag azimuth and lvation rrors. Tabl 4.2: Statistical charactristics for th localization rror obtaind using virtual sourcs and 2 s stimuli. Azimuth Elvation Avrag rror ( µ ) -0.4 o 5.7 o global Standard dviation ( σ ) 7.4 o 15.7 o total 95% confidnc intrval [-14.8 o ; 14.1 o ] [-25.2 o ; 36.5 o ] Th localization offst was thn rmovd by calculating th localization rror as th diffrnc btwn th answrd dirctions and th avrag of th answrd dirctions for a givn stimuli dirction. This is quivalnt to compnsating for th localization offst in th diffrnt dirctions, s figurs 4.3, 4.4 and 4.5. It follows from this that th avrag localization rror bcoms zro in all stimuli dirctions, and th global avrag localization rror vanishs to zro. Th standard dviation of th avrag localization rror also bcoms zro bcaus th localization rror was constant (zro) in all stimuli dirctions. Though th RTO-MP-HFM-123

21 total standard dviation for th localization rror, σ total, now bcoms qual to th standard dviation for th localization rror in th individual dirctions,σ, bcaus no incras in varianc is found du to variation of th avrag localization rror across stimuli dirctions. Figur 4.10 shows all 4524 answrs obtaind using virtual sound sourcs and 2 s stimuli aftr compnsation for th localization offst found in th 58 individual stimuli dirctions. Th black horizontal and vrtical lins show th global avrag for azimuth rror, µ a, global, and lvation rror, µ, global, which ar both trivially qual to 0 o aftr compnsation. Th rd rctangl indicats 95% confidnc intrvals for azimuth and lvation, which ar basd on th compnsatd statistical paramtrs givn in tabl 4.3. Figur 4.10: Compnsatd localization rror calculatd from 4524 answrs obtaind using virtual sourcs and 2 s stimuli in all 58 dirctions. Th rd rctangl indicats 95% confidnc intrval for azimuth and lvation rrors. Each localization rror is indicatd by a blu circl. Black horizontal and vrtical lins indicat th global avrag azimuth and lvation rrors Tabl 4.3: Statistical charactristics for th localization rror obtaind using virtual sourcs and 2 s stimuli compnsatd for th localization offst found in ach of th 58 stimuli dirctions. Azimuth Elvation Avrag rror ( µ ) 0.0 o 0.0 o global Standard dviation (σ ) 7.3 o 12.1 o 95% confidnc intrval [-14.2 o ; 14.2 o ] [-23.7 o ; 23.7 o ] RTO-MP-HFM

22 4.3 Accuracy of physical stup and dirctional data rcording A spcial sssion was run in ordr to dtrmin th accuracy of th physical stup and th systm for rcording th dirctional data, i.. th positions of th ral sound sourcs (loudspakrs) and th trackr systm, which dtrmins th position and orintation of th toy gun. A tst subjct was placd on th platform insid th curtain cylindr and was instructd to point to ach of th 16 loudspakrs on at a tim. Th tst subjct was assistd in this task by a prson outsid th curtain, who nsurd that th lasr dot shind in th cntr of th loudspakr (+/- 0.5 cm). Th light in th room was turnd on during this sssion to nabl visibility of th loudspakr from insid th cylindr. Th obtaind dirctional data was thn analyzd in a similar way as dscribd in sction 4.2. Tabl 4.4 givs th rsults of this analysis. σ avrag is a masur for th accuracy of th positions of th loudspakrs rlativ th thortical corrct positions as dfind in tabl 3.1.σ on th othr hand is a masur for th accuracy of th systm for rcording th dirctional data, i.. gun trackr. Th uncompnsatd confidnc intrval givs th total localization uncrtainty including both th uncrtainty of loudspakr positions and th gun trackr. Th compnsatd confidnc intrval is th uncrtainty only for th gun trackr. This compnsation for localization offst was dscribd in sction 4.2. Two mthods wr trid out for dtrmining th answrd dirction: Dirction givn by th orintation of th toy gun Dirction givn by a straight lin passing through th cntr of th had of th tst subjct and th calculatd position of th lasr dot on th curtain. Similar rsults wr obtaind by th two mthods, which ld to usag of th orintation of th toy gun for all dirctions in this papr. Tabl 4.4: Statistical paramtrs for accuracy of physical stup and dirctional data rcording. Azimuth Elvation Avrag localization rror ( µ ) -0.1 o -0.9 o global Standard dviation for th avrag ( σ ) 1.6 o 2.2 o avrag Standard dviation in ach stimuli dirction (σ ) 0.9 o 0.6 o Total standard dviation ( σ ) 1.8 o 2.3 o total 95% confidnc intrval [-3.7 o ; 3.5 o ] [-5.4 o ; 3.6 o ] 95% confidnc intrval (compnsatd) [-1.8 o ; 1.8 o ] [-1.2 o ; 1.2 o ] RTO-MP-HFM-123

23 4.4 Summry of rsults for sound sourc typ and stimuli lngth Tabl 4.5, 4.6, 4.7 and 4.8 givs a summry of th rsults obtaind using th 4 combinations of long/short stimuli and virtual/ral sound sourcs. Tabl 4.5: Summry of analysis rsults obtaind using virtual sourcs and 2 s stimuli. Azimuth Elvation Avrag localization rror ( µ ) -0.4 o 5.7 o global Elvation rror factor (α ) Standard dviation for th avrag ( σ ) 1.4 o 10.1 o avrag Standard dviation in ach stimuli dirction (σ ) 7.3 o 12.1 o Total standard dviation ( σ ) 7.4 o 15.7 o total 95% confidnc intrval [-14.8 o ; 14.1 o ] [-25.2 o ; 36.5 o ] 95% confidnc intrval (compnsatd) [-14.2 o ; 14.2 o ] [-23.7 o ; 23.7 o ] Tabl 4.6: Summry of analysis rsults obtaind using ral sourcs and 2 s stimuli. Azimuth Elvation Avrag localization rror ( µ ) 0.1 o 2.9 o global Elvation rror factor (α ) Standard dviation for th avrag ( σ ) 1.3 o 4.8 o avrag Standard dviation in ach stimuli dirction (σ ) 4.9 o 6.2 o Total standard dviation ( σ ) 5.1 o 7.8 o total 95% confidnc intrval [-9.9 o ; 10.0 o ] [-12.4 o ; 18.2 o ] 95% confidnc intrval (compnsatd) [-9.6 o ; 9.6 o ] [-12.1 o ; 12.1 o ] RTO-MP-HFM

24 Tabl 4.7: Summry of analysis rsults obtaind using virtual sourcs and 250 ms stimuli. Azimuth Elvation Avrag localization rror ( µ ) -2.8 o 0.9 o global Elvation rror factor (α ) Standard dviation for th avrag ( σ ) 13.6 o 20.5 o avrag Standard dviation in ach stimuli dirction (σ ) 21.4 o 19.2 o Total standard dviation ( σ ) 25.4 o 28.1 o total 95% confidnc intrval [-52.5 o ; 46.9 o ] [-54.2 o ; 55.9 o ] 95% confidnc intrval (compnsatd) [-42.0 o ; 42.0 o ] [-37.7 o ; 37.7 o ] Tabl 4.8: Summry of analysis rsults obtaind using ral sourcs and 250 ms stimuli. Azimuth Elvation Avrag localization rror ( µ ) 0.7 o 4.4 o global Elvation rror factor (α ) Standard dviation for th avrag ( σ ) 6.0 o 8.7 o avrag Standard dviation in ach stimuli dirction (σ ) 8.8 o 11.2 o Total standard dviation ( σ ) 10.7 o 14.2 o total 95% confidnc intrval [-20.2 o ; 21.6 o ] [-23.5 o ; 32.2 o ] 95% confidnc intrval (compnsatd) [-17.3 o ; 17.3 o ] [-22.0 o ; 22.0 o ] 4.5 Pilots vs. civil prsons Sparat analysis has bn prformd on data from th 13 pilots and from th 13 civil prsons to invstigat if diffrncs could b found. Tabl 4.9 givs th uncompnsatd confidnc intrvals and tabl 4.10 givs th confidnc intrvals, which ar compnsatd for localization offst RTO-MP-HFM-123

25 Tabl 4.9: 95% confidnc intrvals no compnsation for localization offst. Azimuth Elvation virtual sourcs 2 s pilots [-13.1 o ; 9.5 o ] [-26.1 o ; 35.2 o ] virtual sourcs 2 s civil prsons [-15.2 o ; 17.5 o ] [-24.2 o ; 37.7 o ] ral sourcs 2 s pilots [-8.0 o ; 6.7 o ] [-11.6 o ; 17.9 o ] ral sourcs 2 s civil prsons [-10.3 o ; 11.9 o ] [-12.8 o ; 18.3 o ] virtual sourcs 250 ms pilots [-57.2 o ; 48.2 o ] [-52.4 o ; 57.5 o ] virtual sourcs 250 ms civil prsons [-47.6 o ; 45.3 o ] [-55.6 o ; 54.1 o ] ral sourcs 250 ms pilots [-21.6 o ; 21.6 o ] [-25.3 o ; 36.0 o ] ral sourcs 250 ms civil prsons [-18.4 o ; 21.1 o ] [-21.2 o ; 28.0 o ] Tabl 4.10: Compnsatd 95% confidnc intrvals. Azimuth Elvation virtual sourcs 2 s pilots [-11.0 o ; 11.0 o ] [-22.7 o ; 22.7 o ] virtual sourcs 2 s civil prsons [-15.9 o ; 15.9 o ] [-24.2 o ; 24.2 o ] ral sourcs 2 s pilots [-6.9 o ; 6.9 o ] [-11.4 o ; 11.4 o ] ral sourcs 2 s civil prsons [-10.6 o ; 10.6 o ] [-12.0 o ; 12.0 o ] virtual sourcs 250 ms pilots [-41.1 o ; 41.1 o ] [-38.3 o ; 38.3 o ] virtual sourcs 250 ms civil prsons [-40.5 o ; 40.5 o ] [-36.0 o ; 36.0 o ] ral sourcs 250 ms pilots [-16.4 o ; 16.4 o ] [-24.0 o ; 24.0 o ] ral sourcs 250 ms civil prsons [-16.9 o ; 16.9 o ] [-18.8 o ; 18.8 o ] 4.6 Front/back rvrsals Front-back ambiguity has bn found to b a significant problm for 3D-Audio systms, howvr systms including a had trackr hav bn shown to almost solv this problm [11]. Th front-back ambiguity or front/back rvrsal occurs whn a sound vnt is prcivd in a position in front of th prson whn th RTO-MP-HFM

26 sound sourc in fact is positiond bhind th prson. Th charactristics of th misprcivd position rlativ to th tru position is a dirct font-to-back or back-to-front mirroring in a vrtical plan that passs through th two ars of th prson (front and back hmisphr). A spcial analysis to uncovr this problm has bn prformd, whr th localization rror was calculatd to b th smallr of two possibilitis: Th diffrnc btwn th prcivd dirction and th stimulus dirction Th diffrnc btwn th mirror imag of th prcivd dirction and th stimulus dirction Tabl 4.11 givs th rsults for this analysis, whr front/back% is a masur of how many tims th mirrord position was usd compard to th total numbr of positions, i.. how oftn front/back rvrsals wr idntifid. Tabl 4.11: Localization offst compnsatd 95% confidnc intrvals for azimuth, whr front/back rvrsals mirrord back to positions closr to th stimuli dirctions. Azimuth - normal Azimuth - mirrord Front/back% virtual sourcs 2 s [-14.2 o ; 14.2 o ] [-12.7 o ; 12.7 o ] 5.1% ral sourcs 2 s [-9.6 o ; 9.6 o ] [-8.2 o ; 8.2 o ] 4.2% virtual sourcs 250 ms [-42.0 o ; 42.0 o ] [-27.2 o ; 27.2 o ] 21.3% ral sourcs 250 ms [-17.3 o ; 17.3 o ] [-14.3 o ; 14.3 o ] 9.1% 5 DISCUSSION On of th intrsting findings in th rsults was th strong corrlation btwn avrag lvation rror and stimuli lvation, which was shown in figur 4.5. Equation 4.7 formulatd a simpl rlationship btwn stimuli lvation, S, and th avrag lvation rror, µ. Th global avrag lvation rror, µ,global, formd an offst for th avrag lvation rror as sn in quation 4.7. On way to xploit this knowldg is to compnsat th stimulus dirction in such a way, that taking th avrag lvation rror into account will yild a prcivd dirction, which closly approximats a dsird lvation, D. Equation 5.1 stats th rlationship btwn th lvation rror, rror, stimulus lvation, S, and th answrd lvation, A. Th avrag lvation rror, µ, can thn b calculatd from quation 5.2, whr E[x] is th xpctd valu oprator [14]. rror = A S (5.1) µ = E µ = E [ A S ] c (5.2) [ A ] S Th xpctd valu of th answrd lvation, i.. th avrag answrd lvation, E[A ], should b qual to th dsird lvation, D. Combining this with quation 4.7 lads to quation 5.3, which givs th RTO-MP-HFM-123

27 rlationship btwn th dsird lvation, D, and th stimuli lvation, S. Isolating th stimuli dirction, S, lads to quation 5.4, which spcifis th ndd stimuli dirction for a givn dsird lvation in ordr to compnsat for th found rlationship btwn avrag lvation rror and stimuli dirction. D µ + S = α S + µ, global + S = ( 1+ α) S + µ, global = (5.3) D µ,global S = (5.4) 1+ α Both th global avrag lvation rror, µ, global, and th lvation rror factor,α, can b found in tabls 4.5, 4.6, 4.7 and 4.8. Using quation 5.4 for calculating th ndd stimuli lads gnrally to stimuli lvation angls largr than th dsird lvation angls,.g. for a dsird lvation angl of 60 o rquir a stimuli lvation angl of 72.4 o using virtual sourcs and 2 s stimuli (tabl 4.5). Howvr stimuli lvation angls can t go abov +90 o or blow -90 o. This fact imposs a limit onto th rang of possibl dsird lvation angls. Stting th stimuli lvation, S, to +90 o and -90 o in quation 5.3 lads to a lowr and uppr limit for dsird lvation angls to o and o. Th impact of this statmnt is that positions nar th pols of th sphr, i o lvation and +90 o, rprsnts a spcial challng, which rquir furthr rsarch. Why should stimuli lvation angl b largr than th dsird lvation angl? On possibl answr could b that humans ar mor spcializd in localizing in th horizontal plan du to vryday lif, whr th majority of sound vnts tak plac in th sam lvl abov th ground as our ars. This combind with th fact that th human ars ar locatd in a way, which favours th discrimination from lft to right rathr than down to up, may xplain why humans has a dfault localization nar th horizontal plan? Th rsults givn in tabl 4.5, 4.6, 4.7 and 4.8 supports a statmnt saying that: Th mor a human is in doubt about whr th sound originats from th mor rstrictd rang xists for th answrd lvation angls. E.g. using virtual sourcs and short stimuli (250 ms) yilds a possibl rang for th dsird lvation angls to [-44.1 o ; 45.9 o ]. Th xprinc is that sound localization bcoms mor difficult for shortr stimuli and for virtual sourcs compard to ral sourcs. Tabl 5.1 givs th rang of possibl dsird lvation angls in th 4 cass originating from tabls 4.5, 4.6, 4.7 and 4.8. Tabl 5.1: Rangs for possibl dsird lvation angls. Possibl dsird lvation angls virtual sourcs 2 s [-61.8 o ; 73.2 o ] ral sourcs 2 s [-75.9 o ; 81.7 o ] virtual sourcs 250 ms [-44.1 o ; 45.9 o ] ral sourcs 250 ms [-74.4 o ; 83.2 o ] Using azimuth and lvation to spcify a dirction in 3D may rprsnt a problm: whn calculating th avrag lvation rror nar th pols of th sphr, thn th answrd dirctions ar limitd/clippd to RTO-MP-HFM

28 th rang +/- 90 o. E.g. for stimuli lvation angl = +90 o, whr tst subjcts will point to dirctions distributd in an ara around th pol of th sphr. Any position away from th pol of th sphr will yild a ngativ lvation rror th lvation rror can t b positiv in this situation by th dfinition of lvation angl. It follows from this, that vn if th answrd dirctions ar vnly distributd around th pol of th sphr, thn th avrag lvation rror will not b zro, i.. it will b ngativ. This ld to a spcial analysis, whr th avrag localization rror wasn t spilt into azimuth and lvation rror it was prformd as an avrag of complx numbrs, which rprsntd th localization rror in polar form. Howvr this didn t chang th rsult dramatically, but th standard dviation for lvation rror did dcras. Figur 4.9 and 4.10 confirms in a graphical way th fit btwn th obtaind localization rrors and th stochastic modl including th calculatd paramtrs. Figur 4.10 shows how compnsation for th localization offst do dcras th ovrall localization uncrtainty in a significant way. It follows from this, that it is vry important to sparat th varianc du to diffrnt localization offsts in diffrnt dirction from th varianc du to localization uncrtainty in ach dirction. Tabl 4.4 shows that both th accuracy of th physical stup and th accuracy of th dirctional data rcording systm wr sufficint compard to th rsults obtaind in th xprimnt. Standard dviation for azimuth using ral sourcs and 2 s stimuli was found to b 4.9 o (tabl 4.6), which is comparabl to rsults obtaind by othr authors: 1.5 o (in th stimuli rang +/-30 o ) [4], 8 o (avrag absolut localization rror) [9] and 1 o to 3 o [8]. Th standard dviation for lvation rrors using ral sourcs and 2 s stimuli was found to b 6.2 o, which again could b compard to th avrag absolut localization rror found by [9]: 8 o. Diffrncs in choic of stimuli, stimuli lngth, physical stup, rang and distribution of sound sourc positions may b possibl xplanations for ths diffrncs. It should b notd, that this xprimnt covrd th full sphr of dirctions, i.. combinations of azimuth and lvation covring th full sphr. Virtual sourcs and stimuli lngth of 2 s yildd a standard dviation of 7.3 o for azimuth and 12.1 o for lvation. Th avrag absolut localization rror for virtual sourcs was found by [9] to b 11 o and 10 o by [13]. Using short stimuli (250 ms) this was found to incras to 21.4 o for azimuth and 19.2 o for lvation. This compars to th data prsntd by [13]: 18 o, which was a tst rstrictd to th horizontal plan. Th standard dviations obtaind for pilots gnrally wr found to b smallr compard to civil prsons. A significantly lowr standard dviation was found for azimuth rrors for pilots compard to civil prsons: 5.0 o vs. 8.0 o. Th ability to focus on a task at hand and to us th input from th auditory systm may b an xplanation for this rsult? Th front/back rvrsal rat was found to b 4.2% (2 s stimuli) and 9.1% (250 ms stimuli) for ral sourcs, which is lowr than th rsults obtaind by [9]: 12%. Howvr this shows that vn whn using ral sourcs th auditory systm has a significant rat of front/back rvrsals. Th rat for virtual sourcs was found to b 5.1% (2 s stimuli) and 21.3% (250 ms). In [9] th front/back rvrsal rat was found to b 20% for virtual sourcs and in [13] th rat was found to b 4% (2 s stimuli) and 10% (250 ms). It is clar, that short stimuli, i.. no had movmnt, gratly incrass th problm with front/back rvrsals both for ral and virtual sourcs, howvr th incras was substantially largr for virtual sourcs. Th main rsults of th papr show that th uncrtainty for azimuth for virtual sourcs dos compar to th ral sourcs. Howvr th uncrtainty for lvation is significantly highr for th virtual sourcs. Actually th uncrtainty for lvation for virtual sourcs (2 s stimuli) compars wll with ral sourcs for short stimuli (250 ms). Th uncrtainty both for azimuth and lvation incrasd for shortr stimuli, whr had movmnts can t b usd in th localization task. This was particularly clar for virtual sourcs RTO-MP-HFM-123

29 6 CONCLUSION Th localization uncrtainty was much highr for short stimuli (0.25 sc.) compard to long stimuli (2 sc.). Th long stimuli nabld had movmnts to b usd in th localization task. Pilots had a lowr localization uncrtainty compard to civil prsons. Had movmnts, i.. long stimuli, gratly rducd th problm of front/back rvrsals, spcially for virtual sound sourcs. No significant localization offst was found for azimuth, whil an offst for lvation of 3 6 dgrs was found using long stimuli. A significant diffrnc btwn th localization offsts (avrag rror) obtaind in diffrnt dirctions was found spcially for lvation, whr th offst was found to hav a strong corrlation to th stimuli lvation. It follows from this, that a significant part of th uncrtainty for lvation can b rmovd by compnsating for ths individual offsts in th diffrnt dirctions (Compnsatd for localization offst). A simpl formula was dvisd for calculating th ndd stimuli lvation from a dsird lvation angl, which gnrally introducs largr stimuli lvation angls compard to th dsird lvation angl. Du to th ndd compnsation of lvation angl a rstrictd rang of dsird lvation angls was dtrmind, which showd a problm of how to position sound sourcs nar th pols of th sphr, i.. nar lvation = +90 o and -90 o. Th rsults of th analysis of th obtaind dirctional data compard wll to rsults prsntd by othr authors undr th obsrvation that significant diffrncs xistd btwn th diffrnt xprimnts: stimuli, stimuli lngth, rproduction stup, rang and distribution for azimuth and lvation angls and sound sourc typ. This xprimnt covrd th full sphr of dirctions: both azimuth and lvation. Th localization uncrtainty for virtual sound sourcs was found to b largr than for ral sound sourcs, spcially for lvation. Howvr th localization prformanc of virtual sound sourcs was comparabl to th prformanc of ral sound sourcs: 95% confidnc intrval (compnsatd for offst): Ral Sound Sourcs: Azimuth=[-9.6 ; 9.6] Elvation=[-12.1 ; 12.1] Virtual Sourcs: Azimuth=[-14.2 ; 14.2] Elvation=[-23.7 ; 23.7] 7 ACKNOWLEDGMENT Th authors would lik to thank Douglas S. Brungart, Wright-Pattrson Air Forc Bas, for taking part in th arly stag discussions about th physical stup for th xprimnt. Aalborg Univrsity is thankd for contributing to th xprimnt by providing loudspakrs and powr amplifirs. Th authors would also lik to thank thir collagus from AM3D A/S and Trma A/S for contributing to th work, stting up and carrying out th xprimnt, for inspiration, for thir xprinc and for dcisiv discussions: Michal Holmbrg, Stffn Vang Madsn, Ptr Thilgaard Hounum, Prbn Schmidt Nilsn, Kim Eriksn, Simon Attrmann Brinch, Robrt Stpin and Pr Schmock. Finally th authors would lik to thank th 26 tst subjct for taking part in th xprimnt, i.. 13 pilots from th Royal Danish Air Forc, from th Air Forc Bass in Aalborg and Karup, Dnmark, and 13 civil prsons, primary studnts in Acoustics at Aalborg Univrsity, Dnmark. RTO-MP-HFM