The  International Congress and Exposition on Noise Contr ol Engineering Predicting anno ance judgments fr om psyc hoacoustic metrics Identiab le ver sus neutraliz ed sounds W
99K - views

The International Congress and Exposition on Noise Contr ol Engineering Predicting anno ance judgments fr om psyc hoacoustic metrics Identiab le ver sus neutraliz ed sounds W

Ellermeier A Zeitler and H astl Sound Qualit Researc Unit Departmen of Acoustics Aalb org Univ ersit redrik Ba jers ej B5 DK9220 Aalb org Denmark ec hnisc he Akustik Mensc hMasc hineKomm unik ation TU unc hen German weazacousticsdk fastlmmkeitumde

Download Pdf

The International Congress and Exposition on Noise Contr ol Engineering Predicting anno ance judgments fr om psyc hoacoustic metrics Identiab le ver sus neutraliz ed sounds W

Download Pdf - The PPT/PDF document "The International Congress and Expositi..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.

Presentation on theme: "The International Congress and Exposition on Noise Contr ol Engineering Predicting anno ance judgments fr om psyc hoacoustic metrics Identiab le ver sus neutraliz ed sounds W"— Presentation transcript:

Page 1
The 33 International Congress and Exposition on Noise Contr ol Engineering Predicting anno ance judgments fr om psyc hoacoustic metrics: Identifiab le ver sus neutraliz ed sounds W. Ellermeier A. Zeitler and H. astl Sound Qualit Researc Unit, Departmen of Acoustics, Aalb org Univ ersit redrik Ba jers ej B-5, DK-9220 Aalb org Ø, Denmark ec hnisc he Akustik, Mensc h-Masc hine-Komm unik ation, TU unc hen, German [we;az]; Abstract [267] It is common practice to predict erceiv ed noise anno ance means of regression mo dels using instrumen tal

psyc hoacoustic metrics as predictors. The alidit of this approac has een criticized for not taking in to accoun non-sensory ariables suc as the meaning of the sound. The presen study in estigates to whic exten judgmen ts of an- no ance reflect sensory attributes in terms of psyc hoacoustic metrics as opp osed to cognitiv and emotional ariables related to the sound source. new signal-pro cessing metho whic substan tially reduces the iden tifiabilit of sound sources as applied to set of 40 en vironmen- tal and pro duct sounds. In the listening exp erimen t, indep enden groups of

participan ts 25 eac h) pro vided anno ance judgmen ts of either the original or neutralized ersion of the sounds using the metho of category-sub division scaling. In the second part of the exp eri- men t, the participan ts rated the affectiv meaning of the sounds on concept-sp ecific seman tic differen tial. Instrumen tal analyses of the sounds included the calculation of psyc hoacoustic metrics of loudness, sharpness, roughness, fluctuation strength, and tonal prominence whic ere en tered in to regression mo dels to predict the outcome of the listening tests. It

turned out that while instrumen tal metrics fared ell in predicting erall anno ance, they did not accoun for the discrepancies in judgmen ts of original ersus neutralized sounds, suggesting that these actually reflect non-sensory effects mediated the ’meaning of the sound. INTR ODUCTION Psyc hoacoustic metrics (indices of loudness, sharpness, roughness, and the lik e) when prop erly computed, and com bined, usually go long in predicting the anno ance reactions pro duced en vironmen tal sounds. They cannot, ho ev er, accoun for non-sensory influences en tering in to the

judgmen of sound. These migh include the attitude to ards the source, effects of familiarit preferences, and user exp ectations ab out protot ypical sounds. Suc non-auditory influences on psyc hoacoustical judgmen ts are sometimes summarized as effects of the me aning of the sound [1 2]. Due to astl’s recen prop osal of signal-pro cessing algorithm [3] that mo difies the acous- tic prop erties of giv en sound so that it is ery lik ely to ecome unrecognizable (and th us ’meaningless in the sense discussed), no ha handle on quan tifying the relativ con- tribution of

acoustic and non-acoustic factors to the judgmen ts made in listening test. The adv an tage of astl’s metho er other alternativ es (suc as filling the temp oral en elop of 1/6
Page 2
the original sound with broadband noise) is that it tak es oth temp oral and sp ectral prop erties in to accoun t, and is designed to preserv the temp oral loudness pattern of the original. That is accomplished first sub jecting the sound to ourier time transform (FTT), then applying some sp ectral broadening to the elemen ts of the FTT pattern, and subsequen tly re-syn thesizing the sound an

in erse FTT [3]. In an earlier rep ort [4], fo cused on the effects of this ’neutralization pro cedure on loudness judgmen ts, sho wing that while iden tifiabilit as greatly reduced (from 93 to 13%), the effects on loudness ere relativ ely small, reac hing significance only for three in sample of 40 sounds. In the presen rep ort, will analyze the effects of the neutralization pro cedure on ratings of annoyanc Clearly while the pro cedure guaran tees loudness (as measured appropriate metrics) to unc hanged, when in estigating anno ance ratings of original and

neutralized sounds, oth acoustical differences (c hanges in roughness, for example due to the sp ectral broadening, see [5 ]), and differences in ’meaning will confounded. disen tangle these effects, in first step, anno ance ratings of oth kinds of sounds will obtained from indep enden groups of sub jects. Subsequen tly the results will related to (a) psyc hoacoustic metrics capturing the sensory effects of the sounds, and to (b) seman tic-differen tial ratings capturing their connotativ ’meaning. METHOD 2.1 ar ticipants total of 50 studen ts at Aalb org

Univ ersit et een 19 and 31 ears of age participated in the exp erimen t. They ere audiometrically screened with the requiremen that their pure-tone thresholds did not exceed the normal curv more than 20 dB in the frequency range from 0.25 to kHz. Subsequen tly half of the participan ts ere randomly assigned to judge the anno ance of the original, half to judge the anno ance of the neutralized sounds. 2.2 Apparatus and Stim uli The original sounds ere recorded using Br uel Kjær (P ortable PULSE 3560 C) fron tend connected to (mono) microphone (Br uel Kjær yp 4165 or 4179) placed at appropriate

distances from 0.3 to from the source. The files ere con erted to 16-bit, 44.1 kHz format to pla ed from regular (RME Digi96 Pro) sound card the output of whic as amplified (Behringer HA 4400) efore eing presen ted diotically to the sub jects listening in double- alled sound-atten uating ham er via headphones (Bey erdynamic DT 990). ourt sounds ere selected for the exp erimen to highly iden tifiable in the original condition: Most of them ere non-stationary ev eryda noises (e.g. toilet flush, do or closing, scissors, car passing), ab out third of the sounds consisted of

pro duct sounds of electrical devices (e.g. hairdry er, kitc hen mixer, razor) recorded in their ypical use. The sounds aried in duration from 0.7 to s, and had erall sound-pressure lev els et een 30 and 80 dB SPL. The 40 recorded sounds ere pro cessed using the algorithm prop osed astl [3] in order to obtain 40 “neutralized sounds ha ving iden tical loudness-time functions. 2.3 Pr ocedure All participan ts erformed three tasks in the follo wing order: (1) scaling exp erimen (loudness or anno ance), (2) an iden tification task, and (3) seman tic-differen tial rating of all sounds.

or the anno ance scaling task rep orted in this pap er, category sub division pro cedure (CS, see [6]) as used: Sub jects ere ask ed to judge eac sound on com bined erbal-n umerical category 2/6
Page 3
scale that consisted of fiv erbal categories whic ere further sub divided in to ten steps and lab elled with the Danish equiv alen ts of “v ery sligh tly anno ying (1-10), “sligh tly anno ying (11-20), “medium (21-30), “strongly anno ying (31-40) and “v ery strongly anno ying (41- 50). The endp oin ts of the resulting 50-p oin scale ere erbally anc hored to denote “not at all

anno ying (0) and “un earably anno ying (b ey ond 50). After short practice run, eac sub ject judged the sounds once in random order. In the subsequen iden tification exp erimen t, the 40 recorded (resp. neutralized) sounds ere pla ed again in random sequence, and the sub ject as ask ed to iden tify the source pro viding oth noun and erb (e.g. “motor idling”). During second session, sub jects judged the same sounds using seman tic differen tial consisting of 12 bip olar adjectiv scales. RESUL TS In the presen rep ort will fo cus on the anno ance scaling data, and their relation to

instru- men tal psyc hoacoustic metrics. Preliminary results on loudness scaling, and on the outcome of the seman tic differen tial measures ha een rep orted elsewhere [4 7]. 3.1 Anno ance scaling 10 15 20 25 30 35 40 SOUNDS [ordered scale alues in the ’neutralized condition] 10 20 30 40 50 ANNO ANCE JUDGMENTS [CS scale] coffee mak er coin do or toilet flush tap glass ell Figure 1: nnoyanc sc ale values (plus/minus standar err ors of the me ans) of the 40 test sounds as judge by 25 articip ants in their neutr alize (fil le cir cles) and by differ ent sample of 25

articip ants in their original version (op en diamonds). Statistic al ly signific ant discr ep ancies ar marke by vertic al arr ows. The anno ance scaling data ere eraged across the 25 sub jects in eac group, and are displa ed in Figure with the sound samples (along the abscissa) eing arranged in ascending 3/6
Page 4
order according to the mean anno ance pro duced the neutralized sounds (filled circles), so that the judgmen ts of the original, iden tifiable sounds (op en diamonds) app ear as deviations from the ’neutral curv e. It is eviden that these deviations are

substan tial, whic is confirmed o-factor (sound pro cessing), mixed analysis of ariance: In addition to the highly significan (but trivial) main effect of the 40 sounds, there is significan main effect of pro- cessing (original vs. neutral), indicating that anno ance ratings of the neutralized sounds ere higher on the erage than those of the originals: (1 48) 99; 03. urthermore, there is highly significan (sound pro cessing) in teraction, sho wing that the effects of the neutralization differ significan tly et een sounds: (16 25 780 01) 92;

001. In order to determine, whic of the discrepancies et een judgmen ts of the original and neutralized sounds ere pro ducing these effects, ost-ho tests ere erformed using Bonfer- roni correction to adjust for hance outcomes due to ultiple testing. the test-wise -lev el of 00128 th us obtained, the sev en sounds mark ed arro ws in Figure ere iden tified as pro ducing significan differences in anno ance in the exp erimen tal conditions in estigated: coffee mak er neutr or ig 36 scale units), do or lo king (13 08 scale units), ouncing coin (11 52), toilet

flushing (13 6), ater running from tap (8 88), bicycle ell (12 04), hampaign glass (13 6). Note that for all of these sounds anno ance judgmen ts of the originals ere lo er than than those of the neutralized ersion. 3.2 Instrumental anal yses In the follo wing, instrumen tal analyses of the 40 sounds and their neutralized coun terparts ere erformed in order to explore whether hanges in psyc hoacoustic metrics can accoun for the outcome of the anno ance scaling exp erimen t. All analyses ere erformed using commercially ailable psyc hoacoustic analysis program (Br uel Kjaer Sound Qualit yp

7698, ersion 3.4.0). Sound-quality changes due to the ’neutr alization Since the parallel loudness scaling study of original and neutralized sounds [4 had sho wn no erall loudness differences due to pro cessing or ig 28 62; neutr 28 38), in estigated whether other sound-qualit metrics migh affected. As ma seen in able 1, on the erage the sounds sligh tly increased in fluctuation strength, and roughness, and decreased somewhat in sharpness. These observ ations are in close agreemen with astl’s [5 earlier analysis of long-duration traffic noise recording. The most striking

effect, ho ev er, is that due to the sp ectral broadening the sounds lo ose almost all of their tonal con ten (as indicated the prominence ratio statistics in the last line of able 1). In fact, for most of the neutralized sounds, no tonal comp onen ts could detected. Analysis of the diff erences betw een or iginal, and neutr aliz ed sounds In second step, psyc hoacoustic metrics ere explored with resp ect to their oten tial in ac- coun ting for the differences et een neutralized and original sounds (eviden in Figure 1). that effect, fiv sound qualit metrics (the

fifth ercen tile of non-stationary loudness and the four parameters listed in able 1) ere en tered in to ultiple-regression equation with the difference in anno ance ratings et een neutralized and original sounds eing the criterion (the alue to predicted). As result, all sound-qualit metrics, except for the prominence ratio dropp ed out as non-significan t, and the latter accoun ted for less than 12 ercen of the ariance ad 115) in the differences to predicted. 4/6
Page 5
able 1: Changes in psycho ac oustic metrics due to neutr alization. Stim uli original

neutralized Roughness [asp er] 0.715 0.795 Fluctuation Str. [v acil] 2.070 2.315 Sharpness [acum] 3.090 2.875 Prominence ratio [dB] 1.985 -5.120 Note. Standard measures of roughness, fluctution strength, and the prominence ratio (substituting ’threshold alue of -6 dB when no tonal comp onen as detected) ere computed for eac sound, as ell as the 50th ercen tile of Aures sharpness. The table en tries are median alues of these metrics across the 40 sounds analysed. Modelling er all anno ance If psyc hoacoustic metrics do not do ell in accoun ting for the differ enc es in anno ance

ratings due to pro cessing, they migh nev ertheless ell suited in predicting erall anno ance in oth kinds of sounds (originals, and neutralized ones). explore this, the measured sound- qualit indices of all 40 neutralized sounds ere en tered in to ultiple linear regression. It turned out, that com bination of loudness ), sharpness 50 ), and roughness predicted the erall anno ance ratings (category scaling of anno ance, CSA) fairly ell ad 856), accoun ting for nearly 86 ercen of the ariance: 07 563 022 50 175 (1) Using the same three metrics (while allo wing for differen co efficien

ts) to predict the an- no ance of the original sounds, reduced the ariance accoun ted for 15%: ad 704. But ev en allo wing for differen metrics in the regression equation for the original sounds did not impro the situation: The est mo del (obtained when substituting the prominence ratio for sharpness) did only sligh tly etter, ad 731. CONCLUSIONS Direct scaling of the anno ance of ell-recognizable sounds, and of their neutralized [3, 5] coun terparts yields substan tial differences. explore, whether these differences are due to the acoustical hanges inheren in the

neutralization pro cedure, um er of analyses using instrumen tal sound qualit metrics ere erformed, yielding three ma jor conclusions: (1) The hanges in psyc hoacoustical metrics due to the pro cedure are minor, except for reduction in tonalness for those sounds ha ving tonal comp onen ts. (2) The differences in the directly scaled anno ance et een original and neutralized sounds cannot accoun ted for an of the instrumen tal metrics explored. (3) Anno ance ratings of oth the neutralized, and the original sounds, can accoun ted for fairly similar ’com bination metrics, but the mo del for

the original sounds accoun ts for substan tially less of the ariance. The fact that the original sounds are less ell predicted psyc hoacoustic metrics suggests that the additional ariance is caused non-acoustic factors related to their iden tifiabilit These results suggest that the differences observ ed in the scaling of original vs. neutralized sounds are indeed due to differences in ’meaning, i.e. to non-acoustical factors. This conclusion is also supp orted preliminary analyses of the Seman tic Differen tial ratings [7 ], supp osedly 5/6
Page 6

the ’connotativ e asso ciations elicited the sounds. These ratings exhibit uc higher correlations with the difference scores than do the instrumen tal metrics. urther analyses comparing oth sets of predictors are under CKNO WLEDGEMENTS This in estigation as carried out while the first authors ere with the Sound Qualit Researc Unit (SQR U) at the Departmen of Acoustics, Aalb org Univ ersit This unit receiv es financial supp ort from Bang Olufsen, Br uel Kjær, and Delta Acoustics Vibration, as ell as from the Danish National Agency for Industry and rade (EFS), and the Danish ec

hnical Researc Council (STVF). The authors ould lik to thank Karin Zimmer for erforming the regression analyses, and Signe Kristensen and elle Coltau for their conscien tious assistance in helping with collecting and co ding the data. Alfred Zeitler is no with the BMW sound engineering departmen t: alfred.zeitler@bm REFERENCES [1] E. Zwic er, H. astl, Psyc hoacoustics. acts and mo dels, 2nd edition, Springer, Berlin 1999. [2] U. Jek osc h, Meaning in the con text of sound qualit assessmen t, custic acta acustic 85(5) pp. 681–684, (1999). [3] H. astl, “Neutralizing the meaning of sound

for sound qualit ev aluations, in Pr dings 17th Interna- tional Congr ess of oustics (ICA 2001) Rome, Italy 2001, (CD-R OM). [4] W. Ellermeier, A. Zeitler and H. astl, “Impact of source iden tifiabilit on erceiv ed loudness, in Pr dings of the 18th International Congr ess on oustics (ICA 2004) Ky oto, Japan, 2004, ol. I, pp. 1491-1494. [5] H. astl, “F eatures of neutralized sounds for long term ev aluation, in Pr dings orum custicum 2002 Sevilla, Spain, 2002, (CD-R OM). [6] J. Hellbr uc k, “Category sub division scaling erful to ol in audiometry and in noise assessmen t, In H. astl et

al. (Eds.), ent tr ends in he aring ese ar ch. estschrift for Seiichir Namb BIS, Olden burg 1996, pp. 317-336. [7] A. Zeitler, W. Ellermeier and H. astl, “Significance of meaning in sound qualit ev aluation, in Pr dings SF A/D GA 2004 Strassb ourg, rance, 2004. 6/6