0 0 -3 ) ) * * 2 $) * - - 1 0 · 2014-08-18 · 0 0 -3 ) ) * * 2 $) * - - 1 0 ... .,5

ACTA ACUSTICA UNITED WITH ACUSTICAVol 95 (2009) 356 ndash 372

DOI 103813AAA918158

The Sound Quality of Car Horns Designing NewRepresentative Sounds

Guillaume Lemaitre Patrick Susini Suzanne Winsberg Stephen McAdamslowast

STMS-IRCAM-CNRS 1 place Igor Stravinsky 75004 Paris France GuillaumeLemaitreircamfr

Boris LetinturierSCE-Klaxon 106 ZI du Clos de la Reine 78410 Aubergenville France

We would like to dedicate this article to the memory of our colleague Suzanne Winsberg

SummaryThis paper is the second of a two-part study of the quality of car horn sounds It aims to provide insights intothe design of new sounds It is based on the assumption that hearing a car horn sound warns road users becausethey recognize the sound of a car horn ie they know what this sound means and what they have to do as aconsequence The three experiments reported in this paper are grounded in a psychoacoustical framework Theyseek to provide car horn builders with recommendations allowing them to create new soundsIn the first part [1] we studied the perception of the timbre of existing car horn sounds We found that from theirperception of the sounds listeners were able to make inferences concerning the different mechanisms causing thesound and that the perceived differences between the sounds were based on the integration of three elementarysensations correlated with three acoustical descriptors In this second part we focus on the agreement among lis-teners in categorizing sounds as being members of the car horn category Membership agreement is operationallydefined as the result of a two-alternative forced-choice task We first study recordings of existing sounds The re-sults allow us to define relationships that predict membership agreement from a set of acoustical descriptors Toextend these results we create a new set of sounds in a second step which we submit to a timbre study similar tothe one reported in [1] We finally study membership agreement for these synthesized sounds The results allowus to define a methodology to create new car horn sounds

PACS no 4350-x 4366-x

1 Introduction Designing sounds thatwarn road users

Car horns are wildly used and widely down-cried How-ever they have the capital (and legal) function of warningroad users against potential danger1 At the same time carhorn builders wish to tune their sounds to match them tocar categories and brand identities Designing the sound ofcar horns thus involves a compromise between the need tocustomize the sounds and the necessity of providing effi-cient warning signals To fulfill these constraints car hornbuilders wish to create new sounds by means of a new de-

Received 3 January 2008accepted 3 November 2008lowast now at CIRMMT Schulich School of Music McGill University 555Sherbrooke St W Montreacuteal Queacutebec Canada H3A 1E31 For instance according to the French ldquoCode de la routerdquo (articles R313-33 R416-1 httpwwwlegifrancefr) i) every motorized vehicle must beprovided with an acoustic warning device ii) any other acoustic signalis forbidden iii) using the acoustic warning signal is only allowed towarn road users against an immediate danger Even the United Nationsrequires a mandatory minimum acoustic level [2]

vice made of an electronic synthesizer and a loudspeaker(see [3] for a description of the device) In this context thegoal of this study is to identify the acoustical properties ofcar horn sounds that allow them to be recognized as suchie that allow the sounds to convey information concern-ing danger to the listener This will allow car horn buildersto design new sounds still perceived as car warning sig-nals

11 Timbre of current car horn sounds

The first step of this study consisted in investigating ex-isting car horn sounds [1] A car horn is a self-oscillatingelectro-acoustical device Two main categories exist Thefirst kind (horn-like devices) is based on an electro-dynamical driver and horn The second kind (plate-likedevices) is also made of an electro-dynamical driver butthere is a metal plate attached to the membrane and nohorn The devices are usually mounted alone (monophonicsounds) or in twos or threes resulting in chords (poly-phonic sounds)

We studied their timbre by using the psychoacousti-cal definition of timbre (ANSI definition [4] as summa-rized by Krumhansl [5] p 44) ldquothe way in which musical

356 copy S Hirzel Verlag middot EAA

Lemaitre et al Sound quality of car horns ACTA ACUSTICA UNITED WITH ACUSTICAVol 95 (2009)

sounds differ once they have been equated for pitch loud-ness and durationrdquo The results of the experiments showedthat the perceived dissimilarities between the sounds of carhorns are subtended by a small number of independent au-ditory attributes (or perceptual dimensions) defining thetimbre space of the car horn sounds Moreover there areacoustical descriptors correlated with these attributes

bull the first attribute is correlated with modulations of thetemporal envelope (roughness)

bull the second attribute is correlated with the spectral dis-tribution of energy (spectral centroid) This descriptorhas been found in other studies to correspond to the at-tribute of brightness and

bull the third attribute is correlated with fine variations ofthe spectral envelope (spectral deviation)

Further a sorting task showed that listeners group togetherthe sounds in categories corresponding closely to the dif-ferent sound-production mechanisms (type of excitationtype of resonator number of devices) These categoriesare defined by the auditory attributes shared by the soundsand thus by the acoustical descriptors

12 Warning signals

There have been different approaches to the perception anddesign of warning signals (see [6] for an overview) Thefirst idea was that a warning signal has to be audible tothe listeners for whom it is intended [7] The issue canbe addressed by ensuring a given acoustic level in a spe-cific area [8] or by designing signals with a low detectionthreshold (considering the design issue as a signal detec-tion paradigm) [9 10 11] However as noted in [12] thedetection threshold is not only modulated by the signal tonoise ratio but also by the perceived relevance of the sig-nal with respect to the perceived dangerousness of the sit-uation

Indeed in many contexts (military aircraft surgeryrooms) many different warning signals occur incessantlyand concurrently And it may happen that users becomeunable to decide whether a warning is really urgent ornot and therefore if they have to respond or not to thissignal Thus Edworthy et al [13] [14] have proposed todesign warning signals with different levels of perceivedurgency relevant to the actual urgency of the situationUrgency is in this case conceived as an auditory attributeof the sound Signals made of bursts of harmonic pulses(after [7]) and speech signals [15 16] were studied Theresults of such experiments are mathematical relationshipsbetween perceived urgency and acoustical properties (themore urgent signals are high in pitch with short tran-sients partials with random frequencies and an irregularrhythm) However some results also show that if the per-ceived urgency can be manipulated by changing the acous-tical properties of the sounds the meaning of the speechsignals also has a great influence on the perceived urgency

So the idea has moved from designing alarms with dif-ferent urgency levels to warning signals that also informthe listener of the reason for this warning [17] Therefore

three requirements for a warning signal were identifiedthe signal must be recognized as a warning the listenermust know what it represents and what should be done inconsequence The same idea is found in the analysis ofthe warning process made by Rogers et al [18] for visualsignals the user must (a) notice the warning (b) encodethe warning (c) comprehend the warning and (d) complywith the warning In order to design warning signals thatlisteners may comprehend Edworthy et al [19] designedsounds informing helicopter pilots about the critical evolu-tion of flight parameters (torque high and low rotor speedetc) precursors to warnings trendsons (trend monitoringsounds) To evaluate the comprehension of the trendsonsthey asked listeners to rate them along semantic scales

Rather than or in addition to having listeners ratesounds on scales (urgency scale or other semantic scales)several authors have directly studied how listeners com-ply with the warning signals Some [20 21 22 23]have measured reaction times of plane pilots performinga tracking task and at the same time having to respondto warning signals (based on Edworthy et al [13]) Simi-larly Suied et al [24] measured reactions times of listen-ers performing a tracking task for signals used to warn cardrivers against too short of a distance between vehiclesUsing a driving simulator Belz et al [25] measured timesto brake when drivers were presented with different au-ditory and visual warning signals Interestingly the reac-tion times were different according to the relationship be-tween the signal and its meaning symbolic (arbitrary rela-tionship) or iconic (representational relationship eg thesound of breaking glass used to signal danger of a poten-tial accident) For acoustic signals displayed alone shortertimes to brake were obtained for iconic signals Graham[26] used a similar paradigm but participants had to de-cide how to react to the warning The results showed thatauditory icons2 lead to more false alarms than symbolicsounds because the sounds could have several (natural)meanings whereas symbolic sounds have only one (arbi-trary) meaning Among these symbolic sounds car hornslead to faster reaction times and fewer false alarms becauseof their lack of ambiguity

This indicates that the meaning of a warning compre-hended by the listener can play an important part in thewarning process This idea appears also in the results ofGuillaume et al [27] They replicated the experiments ofEdworthy et al [13] both with sounds synthesized accord-ing to Edworthyrsquos specifications and with recordings ofalarms occurring in a plane cockpit Overall their resultsfit Edworthyrsquos prediction of perceived urgency on the basisof the acoustical properties of the sounds However theyalso report interesting exceptions for instance one soundidentified as a bicycle bell was rated as less urgent than itshould have been according to Edworthyrsquos prediction Theauthors suggest that this sound was judged as non-urgent

2 Auditory icons are sounds used to convey a meaning when there is arelation of similarity between the sound and the meaning (eg the soundof an analogical camera meaning that a digital photo has been shot)

357

ACTA ACUSTICA UNITED WITH ACUSTICA Lemaitre et al Sound quality of car hornsVol 95 (2009)

because it ldquois often associated with a low level of threatand may even be associated with relaxation and pleasurerdquo(p 207)

An original approach is to be found in Vogelrsquos PhD the-sis [28] He studied several warning signals (priority ve-hicle sirens car horns bicycle bells whistles) in a psy-cholinguistic framework This approach is based on thestudy of the participantsrsquo free descriptions of the signalsmixed into two different background sequences (trafficnoise and public garden) Two experiments are reportedIn a first experiment (also reported in [29]) the partici-pants had to describe the sounds at identification thresh-old In a second experiment the signals were played wellabove threshold Participants had to describe triads of sim-ilar signals and to compare triads of different signals Theanalyses of both these experiments show that the contextimproves the univocal nature of some signals They alsoshow that the participants use different types of items todescribe or compare the sounds the name of the sourcesthe acoustical properties of the sounds and value judge-ments While the participants use mostly the names of thesources when they had to compare sounds described asdifferent sources they describe mostly acoustical proper-ties and value judgements when they describe signals sim-ilarly identified Finally by comparing the descriptions ofthe significant categories of warning signals and acousticalrepresentations of prototypes of these categories he infershypotheses on relationships between acoustical propertiesof the sounds and the sense given to them

13 Perception of environmental sounds

As indicated by some of the results reported above what alistener identifies as the cause of a sound might influencea lot what he or she associates with this sound Several ex-perimental studies have investigated the ability of listen-ers to spontaneously identify the cause of environmentalsounds

Sound events and environmental soundsA first notable result is that there are sounds that are noteven distinguished from their source For instance a seriesof psycholinguistic experiments on soundscape3 percep-tion [31 32 33 34 35 36 37 38] showed that listenersperceive differently amorphous sound sequences (ldquoback-ground noisesrdquo) and sound sequences in which listenersare able to identify emerging sound events (see [39 40]for a discussion)

Conversely to the former sort of sequences listenersdescribe sequences with sound events by referring to theidentified sound sources and with reference to how thesource affects them in their everyday lives Sound eventsare not distinguished from their sources and listeners ap-praise the source and the values they associate with the

3 The term ldquosoundscaperdquo was introduced in the late 70rsquos by the Canadiancomposer R Murray Schafer [30] who defined soundscape as the audi-tory equivalent to landscape Beside Schaferrsquos project the term sound-scape perception is used in a scientific context to characterize how inhab-itants perceive experience and appraise their sonic environment

source These findings coincide with Gaverrsquos ecologicaldistinction between musical listening (when listeners fo-cus on qualities of the acoustic stimulus) and everyday lis-tening (when listeners identify the properties of the eventscausing the sound interaction material shape) [41 42]

This idea of sound event is close to the definition ofenvironmental sound proposed by Vanderveer [43 pp16ndash17]

ldquo any possible audible acoustic event which iscaused by motions in the ordinary human environ-ment ( ) Besides 1) having real events as theirsources ( ) 2) [they] are usually more ldquocomplexrdquothan laboratory sinusoids ( ) 3) [they] are mean-ingful in the sense that they specify events in theenvironment ( ) 4) The sounds to be consideredare not part of a communication system or commu-nication sounds they are taken in their literal ratherthan signal or symbolic interpretationrdquo

In the same study Vanderveer investigated how listenersidentify and describe environmental sounds The resultsshowed that they mostly described 1) the action 2) theobject of the action or 3) the place where the action tookplace

Perception of the cause of the soundsThereafter many publications have studied the perceptionof environmental sounds and have reported the listenerrsquosability to recover auditorily the properties of the eventscausing the sounds Some of these properties were relatedto the objects causing the sound the length of woodenrods dropped on the floor [44] the thickness of struck barsmade of wood or metal [45] the shape (square rectangu-lar or circular) and the materials of struck hung plates [46]the shape of a ball dropped on a plate [47] the categoriesof materials (metal and glass vs wood and Plexiglas) ofrecorded struck plates [48] Others were related to the ac-tion discrimination between bouncing or breaking events(glass objects falling) [49] or the ability of blindfoldedparticipants to fill a vessel to a normal drinking level orto the brim [50]

One important question raised by these results (espe-cially when it comes to design) is to identify the acousticinformation used (or needed) by the listeners to recoverthese properties Synthesized sounds (physical modelingsimulating the physics of the events) of impacted barsof different materials and multidimensional techniques al-lowed McAdams et al [51] to identify perceptual dimen-sions correlated with physical parameters

However another series of experiments using synthe-sized sounds (struck bars) [52 53 54 55] showed thatlisteners do not optimally use the available acoustic infor-mation to decide upon the material or the hollowness ofstruck bars Using recorded sounds it is sometimes diffi-cult to identify a clear correlation between acoustic prop-erties and the perceived event properties [56] or to revealstereotypical relationships between acoustical propertiesand listenersrsquo responses (eg slow loud and low frequencysounds systematically associated male hand-clappers [57]or walkers [58])

358


Factors in environmental sound identificationTherefore it can be assumed that both the acoustic prop-erties of the sound (ie the information present within thesound) and the context and the knowledge of the listenerare responsible for the recognition of a sound This ques-tion has been explored thoroughly in a series of paperspublished by Ballas and Howard The main idea of theseauthors is that the perception of environmental soundsshares similarities with the perception of language (thoughthe parallels having to be considered carefully [59])

Identification of sounds results from both a bottom-upprocess (recovering of the information available in thesound and in the context) and a top-down process (us-ing previous knowledge and expectations) ldquoIt is not onlywhat we hear that tells us what we know what we knowtells us what we hear [60]4 In [60] they showed that thesyntax and the semantics of sound sequences influencetheir memorization (organized and meaningful sound se-quences are better memorized)

In [61] they reported homonym-type sounds soundsbeing discriminated but confused when listeners have toidentify their cause In this case the context helped listen-ers to choose among the alternative causes of the sounds[59]

An imposing series of experiments reported in [62]showed that the identification performance is influencedby several factors including acoustic variables ecologicalfrequency (the frequency with which a listener encoun-ters a specific sound in everyday life) causal uncertainty(measured as the amount of reported alternative causes fora sound) and sound typicality Actually acoustic variablesaccounted for only about half of the variance in identifica-tion time and accuracy Therefore their results suggestedthat sound identifiability is related to many other factorsthan acoustical ones Some of these factors (context inde-pendence - when the sound can be identified easily withoutcontext - the ease of using words to describe the sound)have been studied thoroughly in [28]

14 Approach used in this paper

The function of a car horn sound is to warn people againsta danger Therefore this review leads us to analyze thecase of car horn sounds in light of the different approachesto warning signal design Car horn sounds are warningsignals used by car drivers to warn other drivers about adanger There is no automatic reaction to be undertakenwhen hearing a car horn sound road users have to local-ize the potential danger and to decide how to react (thesame analysis may be found as well in [27]) The first re-quirement is therefore that these sounds must be audiblein a road traffic background noise This requirement is ac-tually already addressed by the law car horn sounds mustbe very loud broadband sounds [2] Since at the time ofthis study the manufacturers are not allowed to supply cars

4 cited from RA Cole and J Jakimik Understanding speech how wordsare heard in G Underwood (Ed) Strategies of information processingNew York Academic Press 1978 p 113

with several different sounds it is not possible to imaginefor the time being a system that would allow the driverto choose between different sounds with different urgencylevels

Rather to warn road users against a danger requires thatwhen hearing the sound of a car horn they must under-stand immediately the meaning ldquodangerrdquo This is the sec-ond requirement

Even when heard out of any context car horn sounds arethe sounds identified the most rapidly [62 63] and almostperfectly [62 64] they have a low causal uncertainty [62]a low ambiguity of meaning [28] and lead to the shortestreaction times [26] (among the sounds studied in the pa-per) These results indicate that the association of meaningwith car horn sounds is very strong But to be associatedwith this meaning a sound must first be recogn ized as acar horn

In practical terms we base our study on the assump-tion that hearing a car horn sound warns road users be-cause they recognize the sound of a car horn because theyknow what this sound means and what they have to do asa consequence This leads us to reformulate our problemdesigning new sounds that still warn road users is equiv-alent to designing sounds that are still recognized as carhornrsquos This is also suggested by Vogel [28] when intro-ducing new warning signals care must be taken that th esesounds are not too different from the already-existing onesthe more the new signals are different from the already-existing ones the more the road users will need time tolearn their meaning

The third requirement is that these sounds must not beconfounded with other sounds [65]

This study situates itself within a psychoacousticalframework Because our aim is to provide car horn buil-ders with acoustical specifications we will base our studyon the perceptual dimensions of timbre and pitch and therelated acoustical descriptors revealed in the previous ar-ticle Indeed these descriptors provide us with a tool al-lowing us to study the acoustical properties of the soundsand are based on what listeners perceive Specifically thesedescriptors have shown their ability to account for the per-ception of the mechanical causes of car horn sounds

We will first seek to identify among existing car hornsounds which are the best items of this category by mea-suring the agreement of the participants on the member-ship of each sound in the category of car horn sounds(Section 2) By observing the gradient of this membershipagreement within the acoustical descriptor space we willbe able to relate it to the acoustical descriptors and thusto provide specifications for the design of new sounds Wewill then test the generality of these relationships and theirrelevance for the creation of new sounds by studying syn-thesized sounds (sections 3 and 4)

2 Experiment 1 Agreement on the mem-bership of current car horn sounds

The timbre of existing car horn sounds was studied in aprevious paper [1] In the first experiment of the current

359


study we will ask listeners whether they perceive thesesounds as coming from car horns (even though all thesounds tested are real car horn sounds heard on a dailybasis by road users) We will therefore measure the agree-ment among the listeners on the membership of each ofthese sounds in the category ldquocar hornrdquo

21 Experimental setupMethodParticipants Twenty-nine participants (14 men and 15women) volunteered as listeners and were paid for theirparticipation They were aged from 18 to 34 years oldMost of them were students from the various universitiesof Paris Thirteen were musicians (from amateur to nearlyprofessional level) and the other sixteen had no musicaleducation Three were audio specialists All reported hav-ing normal hearingStimuli Twenty-two car horn sounds were chosen so asto sample the nine classes obtained from the sorting taskin the previous study (see the lower panel of Figure 2)The car horns were recorded in an anechoic chamber (see[1] for the details of the recordings) All sounds lasted ap-proximately 550 ms They had been previously equalizedin loudness in a preliminary experiment Listeners wereasked to adjust the level of each sound so that they per-ceived it at the same loudness as a reference sound (1 kHzpure tone at 83 dB SPL) Their loudness is therefore 83phonsApparatus The test took place in the IAC sound-atten-uated rooms at IRCAM The experiment was run on aPersonal Computer under Linux and the graphical inter-face was implemented under Matlab The sounds wereamplified through a Yamaha P2075 amplifier and sent toSennheiser HD 520 II headphonesProcedure For each sound we studied the agreement ofthe participants on the membership in the car horn cate-gory by means of a two-alternative forced-choice (2AFC)procedure participants listened to each sound and had toanswer the question ldquoDo you recognize a car horn soundYes or nordquo

The sounds were played in random order The listenershad to answer by clicking one of two icons labeled ldquoyesrdquoand ldquonordquo (see Appendix A for the verbatim of the instruc-tions) The 2AFC task actually amounts to a binary cate-gorization in which the participants categorize each soundin one of two categories ldquocar hornrdquo or ldquonot car hornrdquoCoding results Two variables are derived from this exper-iment For each sound we count how many participantsgave a positive answer (ldquoyesrdquo) We call this variable themembership agreement We also count for each partici-pant how many sounds were rated as a car horn We callthis variable the positive answer rate This latter variableis only used to compare the participant strategies

22 ResultsParticipant strategiesOne way to study the participantrsquos response strategy is tocompare how they divided the 22 sounds into two cate-gories As all the sounds are genuine car horn sounds we

Positive answer rate

Figure 1 Cumulative distribution (in ) of the positive answerrate across participants

expect that most of them will be categorized as car hornsThe positive answer rates range from 39 to 100 witha median of 61 The cumulative distribution (Figure 1)shows that 83 of the participants have categorized morethan 50 of the sounds as car horn sounds

A single-sample t test was performed [66] to test thenull hypothesis that ldquothe average positive answer rate is50rdquo The test result is that the null hypothesis can be re-jected (t(28)=43 plt 001) This leads us to conclude thatparticipants did not make two equal partitions of soundsThe participants thus neither balanced their answers noranswered randomly Furthermore since not all the soundswere categorized as car horns we can expect to observedifferences among the values of membership agreementfor the different sounds

Agreement on the membership of the soundsObserving the agreement among the participants to cate-gorize a sound as a car horn or not three cases can behighlighted

bull There is consensus among the participants to catego-rize the sound as produced by a car horn The member-ship agreement τ is close to 100 The sound is repre-sentative of the category of car horn sounds

bull There is consensus among the participants to catego-rize the sound as not produced by a car horn The mem-bership agreement τ is close to 0 The sound is notrepresentative of the category of car horn sounds

bull There is no consensus among the participants Themembership agreement τ is around 50

Hence thresholds of membership agreement have to beset to decide whether each sound falls into one of thesecases This is done by means of an exact binomial test[66] which tests the null hypothesis for each sound ldquothemembership agreement is 50rdquo If the null hypothesiscan be rejected the sound is either representative or non-representative Otherwise there is no consensus among theparticipants to categorize the sound as a car horn or not A

360


0

20731

50

69793

100

Membershipagreement

(in)

Sounds

Class 1 standard polyphonic plate-like

Class 8 standard monophonic plate-like

Class 3 special plate-like

Class 5 prototype

Class 6 standard monophonic horn-like

Class 9 ship horn


Class 2 special horns

Class 4 very low-pitched horns

Sound Classes

Not representative No consensus Representative

Figure 2 Membership agreement of each of the 22 recordedsounds in the category ldquocar hornsrdquo Symbols refer to the nineclasses of sounds highlighted in [1] reported in the lower panelof the figure Gray scale codes the results of the exact binomialtest White and dark grey p lt 00522 light grey p gt 00522

simple exact binomial test with an α value of 005 leadsto two thresholds τ1=34 (11 among 29) and τ2=66(19 among 29 τ le τ1 rarr non-representative τ ge τ2 rarrtypical) However as several tests are compared a moreconservative significance criterion has to be chosen Thisis done by performing a Bonferroni procedure [66] Thesignificance criterion becomes αadjusted = 00522 and themembership agreement thresholds τ1adjusted = 207 andτ2adjusted = 793

Membership agreement is represented for each sound inFigure 2

The representative familiesFigure 2 shows that every sound in category 1 (standardpolyphonic plate-like) and almost all of the sounds in cat-egory 6 (standard polyphonic horn-like) and category 8(standard monophonic plate-like) are representative Thisindicates that the polyphony and the spectral characteris-tics due to the plate act as a kind of signature of car hornsounds Listening to them reveals that they indeed soundlike a caricature of car horns

For these three categories the membership agreement israther homogeneous (category 1 93-96 category 6 76-90 category 8 72-90) whereas the values are muchmore spread over the categories for which there is no con-sensus (eg category 2 21-52) or the categories of non-representative sounds

For instance within category 7 (standard monophonichorn-like) one sound was categorized as a car horn byonly 30 of the participants whereas another one wascategorized by more than 65 of the same participantsalthough they were judged to be perceptually close to one

another in the dissimilarity rating task The lack of con-sensus thus reveals that the listeners were actually unableto decide whether the sounds did or did not belong in thecategory of car horn sounds and could even give differentresponses for sounds rated as being similar They are am-biguous as the post-experimental interviews revealed par-ticipants declared that they did not know what to answerfor some sounds They could possibly have been car hornsounds but they could also have been emitted by othersound sources such as trumpets car alarms ambulancesirens etc

23 Relation to the acoustical descriptors

We now seek to link up membership agreement and acous-tical properties of the sounds The original LCREG (La-tent Class REGression with spline transformations) algo-rithm developed by Winsberg [67 68] is used This tech-nique aims to build a predictive additive model y of a de-pendent variable y based on spline functions of the in-dependent variables xi y = n

i=1 Smioi (xi) where S

mioi is

a spline function of order oi defined for mi knots Eachspline function is a linear combination of the B-spline ba-sis for the given order and set of knots In our case for eachsound the dependent variable y is the membership agree-ment and xi are acoustical descriptors For a given modelLCREG maximizes the likelihood to get the parametersThe best model is then chosen among competing mod-els using the Bayesian Information Criterion (BIC) [69] alog-likelihood measure of model fit that takes into accountthe number of degrees of freedom in the model

We test several acoustical descriptors in addition tothose revealed in the previous study [1] (we assume indeedthat identification may be based on other perceptual at-tributes than those used to rate dissimilarities between thesounds) The best model found by the algorithm includesas independent variables roughness spectral deviationand fundamental frequency5 Fundamental frequency is re-lated to pitch perception and is therefore not a dimensionof timbre (according to the definition used in this study)The model predictions are significantly correlated with themeasured membership agreement (r(20) = 09) Figure 3represents the three additive functions of the model

It indicates that the sounds leading to higher member-ship agreement are those with high roughness values lowspectral deviation values and a fundamental frequency ofaround 480 Hz The first condition corresponds to poly-phonic sounds The second condition corresponds to thesounds of the plate-like horns The third condition can berelated to the fact that most of the horns sold in Europe aretuned to a fundamental frequency around 440 Hz (this canbe related to the concept of ecological frequency devel-oped by James Ballas [62]) Listeners tend to favor soundsthey are used to listening to This can be further visualizedin Figure 4 which represents the positions of the category

5 For polyphonic sounds we take the fundamental frequency of the low-est note

361


05 1 15 2 25

20

25

30

35

40

45

Roughness (asper)

6 7 8 9 10 11

minus20

minus20

0

0

20

20

40

40

Spectral deviation (sones)

Partialpredictedvalue

300 350 400 450 500 550 600 650

Fundamental frequency (Hz)

Figure 3 Additive predictive model of membership agreement

0

1

2

3

456789101112

300

350

400

450

500

550

600

650

700

Roughness(asper)


F0(Hz)

Figure 4 Representation of the sounds in the descriptor spaceroughness spectral deviation and fundamental frequency Thecolor of the symbols codes the results of the exact binomial testblack = representative sounds gray = ambiguous sounds white= non-representative sounds

of sounds in the space formed by the three descriptors andthe results of the exact binomial test

In this figure the two categories of polyphonic soundsare represented by star symbols They correspond to theupper values of roughness The monophonic plate-likesounds represented by diamond symbols are located atlower values of spectral deviation and roughness Thehorn-like sounds represented by triangles or circles arelocated at higher values of spectral deviation Representa-tive sounds are roughly spread over a hyperplane around

480 Hz Interestingly they are located at the centre of thespace

24 DiscussionWe can draw several conclusions from this experimentQualitatively speaking it is clear that the standard poly-phonic sounds are representative of the category of carhorns whether they are horn-like or plate-like The stan-dard plate-like sounds are representative whether theyare monophonic or polyphonic Two main criteria henceemerge to characterize what makes a sound representativeof car horns polyphony and plate resonance The stan-dard monophonic horn-like sounds are ambiguous Theycan be confused with other sound sources such as trum-pets or alarms (according to informal interviews with thesubjects) This is quite an astonishing result inasmuch asmost of the high-end cars are provided with horn-like hoot-ers This indicates that for most people car horns are stillassociated with the old rough plate-like sounds which ac-tually define a caricature of car horn sounds mainly be-cause they are unlikely to be confused with another soundsource

This conclusion leads us to qualify our first assumptionsome car horn sounds are indeed almost always identi-fied But some others although they are regular car hornsounds are likely to be confounded with other soundsources when heard in a context-free situation One soundwas even judged as non-representative of the category Al-though this sound is currently mounted on cars it has avery high pitch quite different from the sounds usuallyheard on the street

Quantitatively speaking the membership agreement canbe specified and predicted by means of three descriptorsroughness spectral deviation and fundamental frequencyIt must be noted that fundamental frequency is not relatedto timbre It is rather related to the sensation of pitch Thisexperiment therefore shows the importance of the pitchof the car horn sounds This perceptual dimension didnot appear in the timbre study of the car horns becausewe explicitly asked listeners to rate the dissimilarities be-tween the sounds without taking pitch into account (whichhas been demonstrated to be feasible [70]) The predictivemodel does not rely on spectral centroid This may indicatethat although the car horn sounds are perceived with dif-ferent brightnesses these differences do not change theirbelongingness to the car horn category

To test the validity of this model we repeated the sameexperiment with synthesized sounds in the next section

3 Experiment 2 Synthesizing new sounds

The previous results were obtained using only recorded carhorn sounds This means that the relationships between themembership agreement and the acoustical descriptors areonly tested for the range of the acoustical descriptors cov-ered by these sounds To extend these results and to gen-eralize the relationships to descriptor values outside therange of the current recorded sounds we synthesized new

362


sounds These new sounds had to respect two constraintsFirst they had to share the same perceptual dimensionsas the recorded ones and to have a more extended rangeof values than the previously tested sounds in order toinvestigate the relationships between membership agree-ment and acoustical descriptors Second they had to beperceptually close to the categories of existing car hornsounds in order to not be set apart from the existing onesBecause the results of the previous study [1] showed thatlisteners perceive the mechanism causing the sound wetook care to preserve these aspects

Synthesizing sounds fulfilling these constraints alsohelped us to test a possible methodology for the designof new car horn sounds

31 Creation of a new set of sounds

The three descriptors correlated with the perceptual di-mensions shared by the car horns are related to the spec-tral properties of the sounds (spectral centroid and spectraldeviation) and to short-term temporal properties (rough-ness) Yet even this latter property can be seen as spectralsince it may result from the mistuning of the harmonic par-tials of the spectrum Thus we can assume that the per-ceived dissimilarities between the current car horn soundsare based only on spectral and harmonic differences Andwe can further assume that these sounds share identi-cal temporal properties and particularly have an identicaltemporal envelope Therefore to create new sounds closeto the current ones we have to create sounds with the sametemporal envelope Extending the range of the timbre di-mensions can thereby be achieved by modulating the spec-tral and harmonic properties

Synthesis methodCar horns like most of musical instruments can bethought of as an excitation mechanism (the membrane setinto vibration by an electromechanical or pneumatic sys-tem) and a resonator (the plate or the horn) Three phe-nomena are worth considering First it may be assumedthat the resonator acting like a filter has a strong influ-ence on the spectral envelope and thereby on the spec-tral centroid and spectral deviation of the sounds Secondthe particular excitation mechanism of the car horns forcesthe membrane to vibrate with a nearly square wave move-ment Thus the sounds produced have a very rich and har-monic spectrum in steady state However when car hornsare not firmly fixed to the body of the car (which occursafter the car has been driven for a while) the device is notfree to vibrate in the proper way which causes the funda-mental frequency to move slightly (therefore detuning thechords when horns are mounted in twos or threes) Thisslightly shifts the frequencies of the partials from a perfectharmonic series This phenomenon is assumed to be re-sponsible for the roughness of the sounds Third listeningcarefully to the car horn sounds reveals that the harmonicsteady state takes time to become established and to re-lease Furthermore we might suspect that these transientparts of the sounds and particularly the non-harmonic

noise are very important for the recognition of car hornsBecause the perceived dissimilarities between the currentcar horn sounds do not rely on any temporal property wecan therefore assume that these properties are identical forall the sounds of the category They must be kept identicalfor the synthesized sounds if we want them to be percep-tually close to the recorded ones

This analysis of the sound production mechanism of thecar horns led us to propose a synthesis model in four parts

bull A A nearly harmonic excitation source made of asum ofN normalized sinusoids the frequency of whichare integer multiples i of a fundamental frequency ω0

added to an inharmonicity term i The excitation isthen N

i=1 sin(iω0 + i)tbull B A temporal envelope Ti(t) defining the temporalevolution of each of the sinusoids

bull C A non-harmonic noise n(t)bull D A filter defining the amplitude of each of the sinu-soids Ai

The synthesis model is therefore defined by the followingmodel

S(t) =N

i=1

Ai sin (iω0 + i)t Ti(t) + αn(t) (1)

The different parameters of the model are modulated tocreate sounds with different descriptor values

bull The fundamental frequency is adjusted by varying ω0bull The roughness is adjusted by varying each inhar-monicity term i

bull The spectral deviation and the spectral centroid are ad-justed by varying the number of sinusoids N and theiramplitudes Ai

The temporal envelope Ti(t) of each partial and the non-harmonic noise n(t) are kept constant for all the soundsThey are actually computed from two recordings of carhorn sounds (one horn-like and one plate-like both beingrated as representative in the first experiment)

The temporal envelopes of each partial and non-harmo-nic noise are extracted using the ADDITIVE algorithm[71] The signal to noise ratio α is kept constant and ad-justed to 18 dB as measured from the recorded soundsThe principle of the synthesis method is summarized inFigure 5

Synthesized soundsForty-six new sounds were synthesized We chose 19sounds from among them They were all 550 ms in du-ration and were equalized in loudness in a preliminary ex-periment We report in Table I the range of the descriptorsfor these 19 sounds as well as for the 22 sounds previouslytested There are synthesized sounds with higher values ofroughness lower and higher values of spectral centroidlower values of spectral deviation and lower fundamentalfrequencies than the recorded sounds We assume that thetimbre of these sounds is defined by the same perceptualdimensions as the recorded sounds To test the assumptionwe subjected these new sounds to the same experimentalprocedure as in [1] (see section 32)

363


Time

Frequency

Amplitude

A Excitation

ω

ω +ε0

0

i

B Temporal envelope

C Non harmonic noise

Time

Frequency

Frequency

D Filter

Figure 5 Principle of the synthesis method

Table I Range of the acoustical descriptor values for the set of22 sounds tested in [1] and for the set of 19 synthesized soundstested in Experiments 2 and 3 r Roughness (asper) c Spec-tral centroid (Hz) d Spectral deviation (sones) f Fundamentalfrequency (Hz)

Recorded sounds Synthesized soundsMin Max Min Max

r 037 281 053 291c 1380 3790 1180 4560d 530 116 285 116f 262 683 198 500

32 The perception of the timbre of the new sounds

We perform a dissimilarity rating experiment to investi-gate the timbre of the synthesized sounds According to thepsychoacoustical definition of timbre used in this studytimbre is what allows a listener to differentiate two soundsthat have been equalized in duration loudness and pitchFollowing the multidimensional scaling approach (see [1]for a rationale of the method) we first collect dissimilarityjudgments

Dissimilarity judgmentsParticipants Thirty participants (15 men and 15 women)volunteered as listeners and were paid for their participa-tion They were aged from 22 to 43 years old All reportedhaving normal hearing The majority of the participantswere students from the various universities in Paris Thir-teen were musicians (from amateur to nearly professionallevel) and the other 17 had no musical education None ofthem was considered to be an audio specialistStimuli Nineteen sounds were chosen from among the 46synthesized sounds Four recorded sounds that had alreadybeen tested in the previous experiments were also includedto make sure that recorded sounds and synthesized sounds

would not be set apart They were played at the same levelas in the previous experiment (83 phons)Apparatus Same as in previous experimentProcedure Participants all received written instructionsexplaining the task (see Appendix B) They were told thatthey were to make judgments on the timbre The meaningof the word timbre (neither pitch nor perceived durationnor loudness) was explained to them Particular emphasiswas placed on ignoring pitch [70]

All 253 different pairs (AB or BA pairs are consideredas equivalent) among the 23 sounds were presented At thebeginning of the session the participant listened to all ofthe samples in random order to get a sense of the range ofvariation possible Next six training trials were presentedto familiarize the participant with the rating task On eachtrial a pair of sounds was presented separated by a 500-ms silence The participant saw a horizontal slider on thecomputer screen with a cursor that could be moved withthe computer mouse The scale was labeled ldquoVery Simi-larrdquo at the left end and ldquoVery Dissimilarrdquo at the right endA rating was made by moving the cursor to the desired po-sition along the scale and clicking on a button to record itin the computerCoherence of the responses The correlations between theresponses of the participants ranged from 011 to 072One participant was removed from subsequent analysesbecause of the poor correlation of his judgments with theother participants (the correlation between his judgementsand the other participants ranged from r(21) = 011plt001 to r(21) = 036 plt001)

CLASCAL analysisCLASCAL a multidimensional scaling (MDS) techniqueis described in detail in [72] Here we only give a shortdescription In the CLASCAL model dissimilarities aremodeled as distances in an extended Euclidean space of Rdimensions In the spatial representation of the N stimulia large dissimilarity is represented by a large distance The

364


CLASCAL model for the distance between stimuli i and jpostulates common dimensions shared by all stimuli spe-cific attributes or specificities particular to each stimulusand latent classes of subjects These classes have differentsaliences or weights for each of the common dimensionsand for the whole set of specificities For the tthlatent classthe distance between two sounds i and j within the percep-tual space is thus computed according to

dijt =R

r=1

wtr(xir minus xjr)2 + vt(Si + Sj) (2)

In this equation dijt is the distance between sound i andsound j t is the index of the T latent classes xir is thecoordinate of sound i along the rth dimension wtr is theweighting of dimension r for class t R is the total numberof dimensions vt is the weighting of the specificities forclass t and Si is the specificity of sound i

The class structure is latent ie there is no a priori as-sumption concerning the latent class to which a given sub-ject belongs The CLASCAL analysis yields a spatial rep-resentation of the N stimuli on the R dimensions the speci-ficity of each stimulus the probability that each subjectbelongs to each latent class and the weights or saliencesof each perceptual dimension for each class We found aspatial model with two dimensions specificities and twolatent classes (see Figure 6)

We chose the model configuration by comparing BIC[69] across models as well as by performing Hopersquos(Monte Carlo) test [73]

The two classes of participantsTable II displays the weights of the two latent classes ofparticipants over the two dimensions of the spatial modelThe most noticeable difference between the two classes isthat participants in class 2 weight the two dimensions moreoverall than do participants in class 1 This indicates thatthey used a larger range of the slider to rate the dissimi-larities The second difference between the two classes isthat participants in class 2 weight the specificities morethan the dimensions conversely to participants in class 1ie they placed more emphasis on the particularity of eachsound than on the shared properties of all the sounds Wedid not find any relation between the biographical data ofthe participants (gender age musical skills) and the be-longingness to the latent classes

Perceptual dimensions and acoustical descriptorsThe first dimension of the spatial model is correlated withroughness (r(21) = minus08 plt001) and the second dimen-sion is correlated with spectral centroid (r(21) = minus08plt001) No dimension is correlated with spectral devia-tion even in non-optimal (according to BIC values) modelswith higher dimensionality The dissimilarities betweenthe recorded sounds are consistent with the previous dataThe four recorded sounds are not set apart from the syn-thesized ones which indicates that synthesized sounds areperceptually close to the recorded ones (see Figure 6) Thetwo sounds with the highest specificity values have exces-sively audible noise transients

-04 -03 -02 -01 0 01 02 03 04 05

-03

-02

-01

0

01

02

03

04

05

Dimension 1

Dimension2

Figure 6 Representation of the perceptual space obtained bythe Clascal analysis in Experiment 2 Circles represent syntheticsounds Gray squares represent recorded sounds

Table II Weights of the two latent classes of participants overthe two perceptual dimensions Class1 23 participants class 26 participants

Dimension 1 Dimension 2 Specificity

Class 1 083 092 071Class 2 117 108 129Total 2 2 2

DiscussionThe assumption that the timbre of synthesized sounds isdefined by the same perceptual dimensions as the recordedsounds is only partially supported because none of thedimensions of the synthesized sounds is correlated withspectral deviation One hypothesis could be that thesesounds really do not differ according to a perceptual di-mension related to spectral deviation But they were cre-ated such that the range of this descriptor is wider than forthe real sounds Listeners should have been able to hearthe differences

In [1] we assumed that combinations of the sensationscorrelated with spectral centroid and spectral deviationwere used by the listeners as cues that help them to distin-guish between horn-like and plate-like sounds As our syn-thesis method did vary both descriptors arbitrarily combi-nations of these descriptors were no longer related to res-onating phenomenon and it may have become difficult forthe participants to hear dissimilarities due to variations ofspectral deviation alone (see Caclin et al [74] for a discus-sion on a similar phenomenon) They may have focused onmore obvious differences due to roughness or spectral cen-troid variations Moreover listening to the stimuli revealsthat the sounds with the highest specificity values havemore audible transient noises These particularities mayalso have pushed listeners to concentrate only on strongdissimilarities

However it should be stressed that the absence of a per-ceptual dimension related to spectral deviation does notnegate the fact that the first two perceptual dimensions arestill present

365


4 Experiment 3 Agreement on the mem-bership of synthesized sounds

These new sounds are presented in a 2AFC experimentsimilar to the one reported in Section 2 In order to test theconsistency of these measures we include in the test the22 recorded sounds tested in the first experiment

41 Experimental setup

The experiment took place during the same session of ex-periment 2 The participants6 began either with experi-ment 2 or with experiment 3 The order (Exp 2 Exp 3vs Exp 3 Exp 2) was counterbalanced across the partic-ipants None of them had taken part in the previous ex-periment 1 The apparatus and the procedure were exactlythe same as in section 2 The 19 synthesized sounds weretested as well as the 22 recorded sounds

42 Results

Participant strategies and consistency of the measureThe set included roughly as many recorded as syntheticsounds The latter were designed to explore the limits ofthe perceptual space so we expected that a lot of themwould not be categorized as car horn sounds The positiveanswer rates range from 32 to 80 with a median of51

A single-sample t test (testing the null hypothesis ldquotheaverage positive answer rate is 50rdquo) confirmed that par-ticipants partition the set equally (t(30)=097 pgt005)

To determine the consistency of the results the mea-sures of membership agreement for the recorded soundsin this experiment were compared to those found in theexperiment described in section 2 Figure 7 represents theregression of the membership agreement for the previousexperiment (homogeneous set of recorded sounds) ontothe membership agreement measured in this experiment(mixed set of both recorded and synthetic sounds)

The correlation coefficient is 09 (df=20 plt001) andthe measures of membership agreement are smoothlyspread over the regression line (slope 09 origin ordinate52)

The perfect consistency regression line (slope 1 originordinate 0) falls within the 95 confidence interval ofthe regression This allows us to conclude that the measureof membership agreement is not influenced by the kindof set tested The indicator can be compared for the twoexperiments

The representative soundsThe measures of membership agreement are representedin Figure 8 They range from 32 to 968 For eachsound the result of an exact binomial test is representedby the gray scale

There is no consensus among the participants for cat-egorizing most of the synthetic sounds in one of the two

6 There was one more participant in Experiment 3 there are therefore 31participants in this experiment

Representative

Non-representative

Absence ofconsensus

Absence of consensus Representative

Figure 7 Linear regression between the membership agreementmeasured for the recorded sounds when they were part of anhomogeneous set of recorded sounds and when they are part ofa mixed set of both recorded and synthetic sounds The dashedlines define the 95 confidence interval around the predictionThe horizontal and vertical dashed lines correspond to the thresh-olds fixed by the exact binomial test τ1ajust and τ2ajust

0

193

323

50

677

806

100

Non-representative RepresentativeNo consensus

Membershipagreement(in)

Sounds

Figure 8 Membership agreement measured for the mixed set ofboth recorded and synthetic sounds Symbols refer to Figure 2Stars represent the synthetic sounds Results of the binomial testp lt 00541 for white and dark grey bars p gt 00541 for lightgrey bars

categories Only one of the synthetic sounds was represen-tative of car horns For recorded sounds among the eightsounds that were representative when they were tested inthe homogeneous set (not counting the reference sound)seven are again representative when tested in this mixedset The qualitative results described in section 2 remainunchanged

43 Relation to acoustical descriptors

Following the method described in section 2 we relate themeasures of membership agreement to the acoustical de-scriptors by means of the LCREG algorithm The best pre-dictive additive model is again based on roughness spec-tral deviation and fundamental frequency Correlation be-tween predicted and measured membership agreement is(r(39) = 07 plt 001) Figure 9 represents the splinefunctions defining this model

366


The first function indicating the contribution of rough-ness to the membership agreement is qualitatively identi-cal to the first function of Figure 3 Sounds the most of-ten associated with the category ldquocar hornsrdquo are those withthe highest roughness values The second function predictsthat sounds with a spectral deviation value around 7 soneslead to the highest membership agreement The model de-picted in Figure 3 predicted that the highest membershipagreement would have been obtained for sounds with aspectral deviation lower than 9 sones But none of thesounds tested had a spectral deviation lower than 5 sonesHence this experiment allows us to extend the model tolower values of spectral deviation The third function pre-dicts that sounds with a fundamental frequency of around350ndash400 Hz are those that are most often categorized ascar horns

To better visualize the localization of the representativesounds in the descriptor space this space is represented in3D in Figure 10

As in Figure 4 the representative sounds are locatedat the center of the space Two areas can be distin-guished one corresponding to the polyphonic sounds(plate- or horn-like) and one corresponding to the mono-phonic plate-like sounds The unique representative syn-thetic sound is located close to the area corresponding tothe monophonic plate-like sounds These two areas aredefined for a fundamental frequency around 350-400 Hzand for a spectral deviation value between 6 and 9 sonesRoughness in itself does not allow segregation betweenrepresentative and non-representative sounds Rather ithas to be combined with spectral deviation This is mainlydue to the monophonic plate-like sounds which are repre-sentative whereas monophonic horn-like sounds are am-biguous Hence to be representative a sound may possi-bly have a low roughness (ie monophonic) but only if ithas a high spectral deviation value (ie plate-like) In othercases the roughness value must be high

44 DiscussionThe addition of synthetic sounds to the set allows us togeneralize the conclusions drawn from the recorded soundset First of all duplicating the measure of membershipagreement demonstrates that this measure is stable Hereagain the representative car horn categories are the stan-dard polyphonic sounds (both plate- and horn-like) and thestandard monophonic plate-like sounds

We demonstrate again that fundamental frequency playsan important role in predicting the membership agreementin the car horn category Membership agreement is thusrelated to both the timbre and the pitch of the sounds

Finally we generalize the description of the representa-tive sounds The additive regression model applied to syn-thetic sounds and recorded sounds allows us to define moreprecisely the combinations of descriptors that describe thesounds categorized as car horn sounds without ambiguityRepresentative sounds are those with the largest values ofroughness a spectral deviation around 7 sones and a fun-damental frequency around 400 Hz These quantitative re-sults are important for the design of new sounds

05 1 15 2 25

10

20

30

Roughness (asper)

3 4 5 6 7 8 9 10 11

0

20

40



200 250 300 350 400 450 500 550 600 650

minus20

0

20


Figure 9 Predictive model of the membership agreement for theset of mixed recorded and synthetic sounds The crosses repre-sent the sound samples used

0

1

2

3

24

681012

200

300

400

500

600

700

Roughness(asper)

Spectral deviation (sones)Synthetic sounds

F0(Hz)

Figure 10 Localization of the sounds in the descriptor space(Roughness spectral deviation fundamental frequency) The re-sults of the exact binomial test are coded by the color of the sym-bols black = representative white = non-representative gray =no consensus Symbols refer to Figure2

However we shed a light on the importance of the am-biguity phenomenon A large number of synthetic soundsas well as some recorded sounds are not categorized ascar horn sounds not only because they are perceptuallydifferent from the sounds most often categorized as carhorns but also because they possess specific propertiesthat would lead listeners to identify them as other soundsources This is coherent with the CLASCAL analysesof dissimilarity ratings of both recorded and syntheticsounds which showed that sounds were compared accord-ing to specificities in addition to continuous dimensionsHowever we are not able from these experiments to pre-dict possible associations with other categories of soundsources

367


5 General discussion and conclusions

This paper concerns the design of new car horn soundsWarning is the main function of car horns This functionmust be preserved when the sounds are tuned accordingto the customerrsquos wishes The review of the literature onwarning signal design reported in Section 1 leads us tobase our approach on the following assumption hearing acar horn sound warns road users because they recognizethe sound of a car horn they know what this sound meansand they know what they have to do as a consequenceTherefore the experimental studies reported in this paperseek to identify acoustic properties that are responsible fora sound (among sounds sharing common dimensions withcurrent car horns) to be categorized as coming from a carhorn

Following a paradigm that was originally designed tostudy the timbre of musical sounds we have defined in[1] the timbre of car horn sounds as the integration ofthree continuous perceptual dimensions (shared by all thesounds) and specificities (particular to each sound) Thecontinuous perceptual dimensions were correlated withappropriate acoustical descriptors Latent class analysis re-vealed that different classes of participants weighted thedimensions and specificities differently The latent classeswere not related to any recorded biographical data con-cerning the participants (age gender musical skills)

The experiments reported in this paper were focused onmeasuring the agreement of some groups of listeners onthe membership of sounds in the car horn category Themembership agreement was operationally defined as theresults of a 2AFC task These measures were analyzed soas to highlight three types of agreement there might bean agreement among listeners to categorize a sound as acar horn In this case we called the sound representativeConversely there might be an agreement of the listenersto categorize a sound as not coming from a car horn Inthis case we called the sound non-representative Whenthere is no agreement among the participants analyses ofthe participantsrsquo comments suggested that the sounds wereambiguous

We first measured in Section 2 the membership agree-ment of the 22 recorded sounds tested in [1] Whereasmost of the sounds were representative we observed agradient of membership agreement By means of a mul-tispline regression technique we were able to relate themembership agreement to three acoustical descriptors ofthe sounds roughness spectral deviation and fundamentalfrequency Furthermore the results showed that some cat-egories of car horn sounds were systematically categorizedas car horns

To generalize these conclusions to descriptor values out-side the range of the recorded sounds (thereby testing amethodology for the design of new sounds) we createdin Section 3 a set of new sounds The synthesis methodwas designed so as to preserve the temporal properties ofthe sounds (particularly the transient parts) suspected tounderlie the recognition of the car horns and so as to cre-ate sounds sharing the same perceptual dimensions as the

recorded ones To test this last assumption we subjectedthe created sounds to a dissimilarity rating task The anal-ysis revealed that these sounds shared only two of the threedimensions of the recorded sounds

Finally in Section 4 we performed a 2AFC task thataimed to measure car horn membership agreement for amixed set of recorded and synthesized sounds These re-sults confirm those of Section 2 and lead to a generaliza-tion of the predictive model

Several conclusions are to be drawn from these resultsFrom a general standpoint a major result of the first partof the study [1] was the importance of the perception ofthe sound-producing mechanism The categorization ex-periment indeed revealed that categories of car horns builtby listeners closely correspond to the different kinds ofdevices These categories were preserved both in dissimi-larity judgments and in 2AFC tasks

These experiments have suggested that when sounds arenot categorized as car horns it is not only because theyare dissimilar to the sounds most often categorized as carhorns but also because they may be identified as anothersound source This has shed light on the problem of iden-tification ambiguity Ambiguity occurs when a sound maybe associated with distinct categories of sound sources Inour case ambiguity must be avoided if a sound is con-fused with another sound source it may fail to convey thewarning message

However our results only allow us to predict howldquocloserdquo a sound is to what we may call a shared repre-sentation of what a car hornrsquos sound is It does not predictif the sound may be identified as another sound sourceAs our experimental tasks were based on sound compar-isons we have emphasized the properties shared by thesounds The acoustical descriptors that we used are cor-related with the common perceptual dimensions they de-scribe the shared properties of the sounds A more effi-cient description should also include idiosyncratic proper-ties because these distinctive properties may explain whysome sounds can be confused with other sound sourcesThe CLASCAL analyses performed on both synthetic andrecorded sounds have included not only common dimen-sions but also specificities which are individual propertiesof the sounds (this is similar to Tverkyrsquos contrast model[75] which aims to include both common and distinctivefeatures) These specificities may indicate possible mis-interpretations of the sound It is however difficult to de-termine to which acoustical properties these specificitiescorrespond and then to make any a priori predictions

The question of sound source identification is howeveronly partially addressed by this study Indeed all the re-ported experiments were done in a laboratory without anyacoustic visual or situational context The question is stillopen as to how these sounds would be categorized if theywere heard on the street However we can assume that thecontext would play two roles Firstly the acoustic back-ground noise would raise the detection threshold of thesounds The issue of detection of the sounds is howeveralready addressed by the very high level imposed by law

368


Secondly it can be assumed that a road traffic situationwould lower the ambiguity of some sounds Indeed if asound heard in a laboratory can be confused with a musi-cal instrument sound and the more so when this labora-tory is located in a institution devoted to music it can beassumed that the same sound heard in a road traffic situa-tion would have less chance to be confused with a trumpetcall As noted by Vogel [28] listening to warning signalsin context can improve the univocal nature of these sig-nals We can therefore assume that our results are moreconservative than the real situation

Another interesting finding of our experiments is thatlisteners do not use the same perceptual dimensions whenthey have to judge the dissimilarity between the soundsand when they have to categorize a sound as a car hornIndeed the timbre study reported in [1] revealed three per-ceptual dimensions correlated to roughness spectral cen-troid and spectral deviation When listeners had to cate-gorize the same sounds our analyses concluded that theyhad based their judgments on roughness spectral devia-tion and fundamental frequency It is easy to explain whyfundamental frequency did not appear in the timbre studyin this experiment listeners were specifically asked to notbase their judgment on pitch and other studies have shownthat they are able to do so [70] Because the wide major-ity of car horn sounds that can be heard nowadays have afundamental frequency in the region of 440 Hz introduc-ing sounds with lower or higher fundamental frequenciesmay lead listeners to judge them as unusual and makesthis descriptor a good predictor of the membership agree-ment More puzzling is the fact that spectral centroid didnot seem to have been used to categorize the sounds ascar horns even when we introduced sounds with moreextreme values of this descriptor The brightness of thesounds (a sensation correlated to spectral centroid) ap-pears very often in the multidimensional study of timbre(see [1] for a review) but seems therefore to be of minorimportance to categorize car horn sounds with respect tothe other sensations related to the modulations of funda-mental frequency roughness and spectral deviation Spec-tral deviation on the other hand did not appear as a per-ceptual dimension of the timbre of the synthesized soundscertainly because some synthesis artifacts overwhelm themore subtle differences due to variations of spectral devi-ation Yet the analyses of the third experiment predictedthat listeners use a sensation correlated to this descriptorto categorize the sounds as car horns or not This again in-dicates that listeners can weigh their sensations differentlyaccording to what they have to judge

Going back to the framework of sound design theseresults are useful despite the reservations expressed inthe above paragraphs Car horn builders will continue todesign broadband loud harmonic sounds Hence tuningnew sounds may be conceived as choosing values of thedescriptors of the car horn sounds Our synthesis methodeasily allows a car horn builder to design a new sound andto compute the descriptor values With the results of thestudies of the agreement on the membership of the sounds

the sound designer is thus able to predict whether such asound will be close to the sounds best recognized as carhorns

Appendix Experimental instructions pro-vided to the participants

A 2AFC experiment (Experiments 1 and 3)

Goal of the study

The goal of this experiment is to answer the question ldquoDoyou recognize a car horn sound rdquo for each sound of a set

Procedure

You will sit in front of a computer screen You will heara set of sounds played one after the other For each soundyou will have to answer the question ldquodo you recognize acar horn sound rdquo Two buttons are displayed on the inter-face labeled with ldquoyesrdquo and ldquonordquo To indicate your answeryou will have to click on one of these buttons The soundsare only played once When you have entered your answerthe next sound will be played after a pause Try to answerspontaneously

Note

The sounds may originate from different sources We areinterested in your opinion so there is no ldquocorrect answerrdquoDo not try to balance the amount of ldquoyesrdquo and ldquonordquo an-swers You can even answer ldquoyesrdquo for every sound or ldquonordquofor every sound if this is what you hear

B Dissimilarity rating experiment (Experiment 2)

Goal of the study

The goal of this experiment is to study the perception ofthe timbre of a set of sounds Your task is to judge thedissimilarity that you perceive between two sounds

Procedure

You will sit in front of a computer screenThere are 23 sounds in the test They all last about halfa second At the beginning of the test you will be pro-vided with 23 buttons which allow you to listen to the 23sounds and to familiarize yourself with them Then youwill be provided with each one of the 23 possible pairsof sounds among the 23 sounds For each pair of soundsthe interface looks similar there are two buttons labeledldquolisten againrdquo and ldquovalidaterdquo above a cursor with the la-bels ldquovery differentrdquo and ldquovery similarrdquo at each extremityWhen you click on the ldquolisten againrdquo button you can hearthe two sounds You can listen to the pair of sounds asmany times as you wish The cursor allows you to rate thedissimilarity between the sounds When your are sure ofyour rating click on the ldquovalidaterdquo button This moves tothe next pair of sounds Before the real test you will beprovided with six pairs of sounds to familiarize yourselfwith the interface in the presence of the experimenter

369


Remark on the notion of timbre

You have to group together sounds with similar timbreTimbre is what allows you to distinguish between twosounds having the same duration the same intensity andthe same pitch For instance two musical instrumentsplaying the same note with the same intensity and of thesame duration do not sound identical What distinguishesthem is referred to here as ldquotimbrerdquo Timbre may alsobe called the ldquocolorrdquo ldquotexturerdquo of the sound Thesesounds are supposed to have the same intensity You mayhowever feel that certain sounds are louder than othersWe ask you to not take into account intensity in your judg-mentsSimilarly the sounds do not all have the same pitch Theyldquoplay different notesrdquo Here again we ask you to not in-clude these differences of pitch in your judgments butrather to focus on timbre

Acknowledgements

This study has been founded by SCE Klaxon and by theFrench Ministry of Industry

References

[1] G Lemaitre P Susini S Winsberg B Letinturier SMcAdams The sound quality of car horns a psychoacous-tical study of timbre Acta Acustica united with Acustica93 (2007) 457ndash468

[2] United Nation Organization Economic and Social Coun-cil Uniform provisions concerning the approval of audi-ble warning devices and of motor vehicles with regard totheir audible signals Regulation 28 to the TRANSWP29343Rev13 agreementhttpwwwuneceorgtransmainwp29wp29wgs

wp29genwp29fdocTRANS-WP29-343r13efrpdfhttpwwwuneceorg trans 2000

[3] G Lemaitre B Letinturier B Gazengel Electroacousti-cal model of an electrodynamic horn loudspeaker used as anew car horn Applied Acoustics 69 (2008) 47ndash59

[4] USA acoustical terminology S11-160 American StandardAssociation 1960

[5] C Krumhansl Why is musical timbre so hard to under-stand ndash In Structure and perception of electroacousticsound and music S Nielzen O Olsson (eds) Elsevier1989 43ndash53

[6] N A Stanton e J Edworthy Human factors in auditorywarnings Ashgate Publishing 1999

[7] R D Patterson Guidelines for auditory warning systemson civil aircraft Civil Aviation Authority 1982 Paper82017

[8] C Laroche L Lefebvre Determination of acoustic fea-tures for reverse alarms field measurements and the de-sign of a sound propagation model Ergonomics 41 (1998)1203ndash1221

[9] H Fleischer J Blauert Audibility of some specific public-adress warning signals in typical environmental noise situ-ations Applied Acoustics 27 (1989) 305ndash319

[10] J D Binnington A W McCombe M Harris Warning sig-nal detection and the acoustic environment of the motorcy-clist British Journal of Audiology 27 (1993) 415ndash422

[11] M Ebata H Baba Detection of the ambulance siren whiledriving Proceedings of Inter-noise 1997 Budapest Hun-gary 1997 1009ndash1014

[12] M R Lehto J D Papastavrou A signal-detection theorybased perspective on design of warnings Perceptual andMotor Skills 86 (1998) 720ndash722

[13] J Edworthy S Loxley I Dennis Improving auditorywarning design Relationship between warning sound pa-rameters and perceived urgency Human Factors 33 (1991)205ndash231

[14] E J Hellier J Edworthy I Dennis Improving auditorywarning design Quantifying and predicting the effects ofdifferent warning parameters on perceived urgency HumanFactors 35 (1993) 693ndash706

[15] E Hellier B Weedon J Edworthy K Walters Using psy-chophysics to design speech warnings Proceedings of theIEA 2000HFES 2000 Congress 2000 698ndash701

[16] B Weedon E Hellier J Edworthy K Walters Perceivedurgency in speech warnings Proceedings of the IEA2000HFES 2000 Congress 2000 690ndash693

[17] J Edworthy N Stanton A user-centred approach to the de-sign and evaluation of auditory warning signals 1 Method-ology Ergonomics 38 (1995) 2262ndash2280

[18] W A Rogers N Lamson G K Rousseau Warning re-search An integrative perspective human factors 42 2000(102-139)

[19] J Edworthy E Hellier R Hards The semantic associa-tions of acoustic parameters commonly used in the designof auditory information and warning signals Ergonomics38 (1995) 2341ndash2361

[20] J L Burt D S Bartolome D W Burdette J R ComstockA psychophysiological evaluation of the perceived urgencyof auditory warning signals Ergonomics 38 (1995) 2327ndash2340

[21] E C Haas J C Casali Peceived urgency of and responsetime to multi-tone and frequency-modulated warning sig-nals in broadband noise Ergonomics 38 (1995) 2313ndash2326

[22] E C Haas J Edworthy Designing urgency into auditorywarnings using pitch speed and loudness Computing andControl Engineering Journal (1996) 193ndash198

[23] E C Haas J Edworthy The perceived urgency and de-tection time of multitone auditory signals ndash In HumanFactors in Auditory Warnings N A Stanton J Edworthy(eds) Ashgate Publishing Ltd 1999

[24] C Suied P Susini S McAdams Evaluating warningsound urgency with reaction times Journal of Experimen-tal Psychology Applied 14 (2008) 201ndash212

[25] S M Belz G S Robinson J G Casali A new class ofauditory warning signals for complex systems Auditoryicons Human Factors 41 (1999) 608ndash618

[26] R Graham Use of auditory icons as emergency warningsevaluation within a vehicle collision avoidance applicationErgonomics 42 (1999) 1233ndash1248

[27] A Guillaume L Pellieux V Chastres C Drake Judgingthe urgency of nonvocal auditory warning signals percep-tual and cognitive processes Journal of experimental psy-chology Applied 9 (2003) 196ndash212

[28] C Vogel Eacutetude seacutemiotique et acoustique de lrsquoidentificationdes signaux sonores drsquoavertissement en contexte urbainPhD thesis Universiteacute de Paris 6 1999

370


[29] C Vogel D Dubois J-D Polack M Castellengo Per-ception of warning signals in urban context Proceedingsof Inter-noise 98 Christchurch New-Zealand November1998 1203ndash1206

[30] R M Schafer The tuning of the world Random HouseInc 1977

[31] V Maffiolo D Dubois S David M Castellengo J-D Po-lack Loudness and pleasantness in structuration of urbansoundscapes Proceedings of Inter-noise 98 ChristchurchNew Zealand 1998

[32] C Guastavino D Dubois J-D Polack C Arras Low fre-quency perception in urban soundscapes a cognitive ap-proach Proceedings of the 17th International Conferenceon Acoustics ICA 2001 Roma Italy 2001

[33] M Raimbault C Lavandier Sound ambient environmentof urban places comparison of sound appraisal factors withacoustical parameters Proceedings of the 3rd EuropeanCongress on Acoustics ndash Forum Acusticum Sevilla SpainSeptember 2002 Reproduced in Acta Acustica united withAcustica 88 (2002) Suppl 1 S70

[34] M Raimbault D Dubois Perceptual judgements about ur-ban soundscapes Proceedings of Euronoise 2003 NaplesItaly 2003

[35] M Raimbault C Lavandier M Beacuterengier Ambient soundassessment of urban environments field studies in twofrench cities Applied Acoustics 64 (2003) 1241ndash1256

[36] D Dubois C Guastavino V Maffiolo A cognitive app-roach to soundscape Acoustic phenomena between ldquonoi-se(s)rdquo and ldquosound(s)rdquo Proceedings of the Joint CongressCFADAGArsquo04 Strasbourg France march 2004 347ndash348

[37] C Guastavino B F Katz J-D Polack D J Levitin DDubois Ecological validity of soundscape reproductionActa Acustica united with Acustica 91 (2005) 333ndash341

[38] B Schulte-Fortkamp D Dubois Recent advances insoundscape research Acta Acustica united with Acustica92 (2006) Preface to the special issue vndashviii

[39] D Dubois Categories as acts of meaning The case ofcategories in olfaction and audition Cognitive ScienceQuaterly 1 (2000) 35ndash68

[40] D Dubois Perception representation and knowledgeacoustic phenomena between noise and sounds Proceed-ings of the EAA Symposium and architectural and urbanacoustics TecniAcustica Bilbao Spain october 2003

[41] W W Gaver How do we hear in the world Explorationsin ecological acoustics Ecological Psychology 5 (1993)285ndash313

[42] W W Gaver What do we hear in the world An ecologicalapproach to auditory event perception Ecological Psychol-ogy 5 (1993) 1ndash29

[43] N J Vanderveer Ecological acoustics human perceptionof environmental sounds PhD thesis Cornell University1979

[44] C Carello K L Anderson A J Kunkler-Peck Perceptionof object length by sound Psychological science 9 (1998)211ndash214

[45] S Lakatos S McAdams R Causseacute The representation ofauditory source characteristics simple geometric sourcesPerception and psychophysics 59 (1997) 1180ndash1190

[46] A J Kunkler-Peck M T Turvey Hearing shape Journalof Experimental psychology human perception and perfor-mance 26 (2000) 279ndash294

[47] M Grassi Do we hear size or sound Balls dropped onplates Perception and Psychophysics 67 (2005) 274ndash284

[48] B L Giordano S McAdams Material identification ofreal impact sounds effect of size variation in steel glasswood and plexiglass plates Journal of the Acoustical Soci-ety of America 119 (2006) 1171ndash1881

[49] W H Warren R R Verbrugge Auditory perception ofbreaking and bouncing events a case study in ecologicalacoustics Journal of Experimental Psychology Humanperception and performance 10 (1984) 704ndash712

[50] P A Cabe J B Pittenger Human sensitivity to acousticinformation from vessel filling Journal of experimentalpsychology human perception and performance 26 (2000)313ndash324

[51] S McAdams A Chaigne V Roussarie The psychome-chanics of simulated sound sources material properties ofimpacted bars Journal of the Acoustical Society of Amer-ica 115 (2004) 1306ndash1320

[52] R A Lutfi Correlation coefficients and correlation ratiosas estimates of observer weights in multiple-observationtasks Journal of the Acoustical Society of America 97(1995) 1333ndash1334

[53] R A Lutfi E L Oh Auditory discrimination of materialchanges in a struck-clamped bar Journal of the AcousticalSociety of America 102 (1997) 3647ndash3656

[54] R A Lutfi Auditory detection of hollowness Journal ofthe Acoustical Society of America 110 (2001) 1010ndash1019

[55] R A Lutfi E Oh E Storm J M Alexander Classifi-cation and identification of recorded and synthesized im-pact sounds by practiced listeners musicians and non-musicians Journal of the Acoustical Society of America118 (2005) 393ndash404

[56] D J Freed Auditory correlates of perceived mallet hard-ness for a set of recorded percussive sound events Journalof the Acoustical Society of America 87 (1990) 311ndash322

[57] B H Repp The sound of two hands clapping an ex-ploratory study Journal of the Acoustical Society of Amer-ica 81 (1987) 1100ndash1109

[58] X Li R J Logan R E Pastore Perception of acous-tic source characteristics walking sounds Journal of theAcoustical Society of America 90 (1991) 3036ndash3049

[59] J A Ballas T Mullins Effect of context on the identifi-cation of everyday sounds Human Performance 4 (1991)199ndash219

[60] J H Howard J A Ballas Syntactic and semantic factorsin the classification of nonspeech transient patterns Per-ception and Psychophysics 28 (1980) 431ndash439

[61] J A Ballas J H Howard Interpreting the language of en-vironmental sounds Environment and Behavior 19 (1987)91ndash114

[62] J A Ballas Common factors in the identification of anassortment of brief everyday sounds Journal of Exper-imental Psychology Human Perception and Performance19 (1993) 250ndash267

[63] M M Marcell D Borella M Greene E Kerr S RogersConfrontation naming of environmental sounds Journal ofclinical and experimental neuropsychology 22 (2000) 830ndash864

[64] C-Y P Chiu D L Schacter Auditory priming for non-verbal information implicit and explicit memory for envi-ronmental sounds Consciousness and Cognition 4 (1995)440ndash458

[65] H Houmlge A Schick S Kuwano S Namba M Bock HLazarus Are there invariants of sound interpretation Thecase of danger signals Proceedings of the 5th internationalcongress on noise as a public health problem Stockholm1988 253ndash258

371


[66] D C Howell Statistical methods for psychology PWS-Kent 1992

[67] S Winsberg Multiple regression with tensor productsplines Presented at the European Meeting of the Psycho-metric Society University of Leuven Belgium July 1989

[68] S Winsberg Multiple regression with tensor productsplines Presented at the 1990 Meeting of the PsychometricSociety Princeton New Jersey June 1990

[69] G Schwarz Estimating the dimensions of a model Theannals of Statistics 6 (1978) 461ndash464

[70] J Marozeau A de Cheveigneacute S McAdams S WinsbergThe dependency of timbre on fundamental frequency Jour-nal of the Acoustical Society of America 114 (2003) 2946ndash2957

[71] Depalle Garcia Rodet Woehrmann Perry The additiveanalysis-synthesis package Ircam wwwircamfr equipesanalyse-synthese DOCUMENTATIONS additive

[72] S Winsberg G D Soete A latent class approach to fittingthe weighted Euclidian model CLASCAL Psychometrika58 (1993) 315ndash330

[73] A C A Hope A simplified Monte-Carlo significance testprocedure Journal of the Royal Statistical Society B 30(1968) 582ndash598

[74] A Caclin S McAdams B K Smith S Winsberg Acous-tic correlates of timbre space dimensions a confirmatorystudy using synthetic tones Journal of the Acoustical So-ciety of America 118 (2005) 471ndash482

[75] A Tversky Features of similarity Psychological review 84(1977) 327ndash352

372


sounds differ once they have been equated for pitch loud-ness and durationrdquo The results of the experiments showedthat the perceived dissimilarities between the sounds of carhorns are subtended by a small number of independent au-ditory attributes (or perceptual dimensions) defining thetimbre space of the car horn sounds Moreover there areacoustical descriptors correlated with these attributes

bull the first attribute is correlated with modulations of thetemporal envelope (roughness)

bull the second attribute is correlated with the spectral dis-tribution of energy (spectral centroid) This descriptorhas been found in other studies to correspond to the at-tribute of brightness and

bull the third attribute is correlated with fine variations ofthe spectral envelope (spectral deviation)

Further a sorting task showed that listeners group togetherthe sounds in categories corresponding closely to the dif-ferent sound-production mechanisms (type of excitationtype of resonator number of devices) These categoriesare defined by the auditory attributes shared by the soundsand thus by the acoustical descriptors

12 Warning signals

There have been different approaches to the perception anddesign of warning signals (see [6] for an overview) Thefirst idea was that a warning signal has to be audible tothe listeners for whom it is intended [7] The issue canbe addressed by ensuring a given acoustic level in a spe-cific area [8] or by designing signals with a low detectionthreshold (considering the design issue as a signal detec-tion paradigm) [9 10 11] However as noted in [12] thedetection threshold is not only modulated by the signal tonoise ratio but also by the perceived relevance of the sig-nal with respect to the perceived dangerousness of the sit-uation

Indeed in many contexts (military aircraft surgeryrooms) many different warning signals occur incessantlyand concurrently And it may happen that users becomeunable to decide whether a warning is really urgent ornot and therefore if they have to respond or not to thissignal Thus Edworthy et al [13] [14] have proposed todesign warning signals with different levels of perceivedurgency relevant to the actual urgency of the situationUrgency is in this case conceived as an auditory attributeof the sound Signals made of bursts of harmonic pulses(after [7]) and speech signals [15 16] were studied Theresults of such experiments are mathematical relationshipsbetween perceived urgency and acoustical properties (themore urgent signals are high in pitch with short tran-sients partials with random frequencies and an irregularrhythm) However some results also show that if the per-ceived urgency can be manipulated by changing the acous-tical properties of the sounds the meaning of the speechsignals also has a great influence on the perceived urgency

So the idea has moved from designing alarms with dif-ferent urgency levels to warning signals that also informthe listener of the reason for this warning [17] Therefore

three requirements for a warning signal were identifiedthe signal must be recognized as a warning the listenermust know what it represents and what should be done inconsequence The same idea is found in the analysis ofthe warning process made by Rogers et al [18] for visualsignals the user must (a) notice the warning (b) encodethe warning (c) comprehend the warning and (d) complywith the warning In order to design warning signals thatlisteners may comprehend Edworthy et al [19] designedsounds informing helicopter pilots about the critical evolu-tion of flight parameters (torque high and low rotor speedetc) precursors to warnings trendsons (trend monitoringsounds) To evaluate the comprehension of the trendsonsthey asked listeners to rate them along semantic scales

Rather than or in addition to having listeners ratesounds on scales (urgency scale or other semantic scales)several authors have directly studied how listeners com-ply with the warning signals Some [20 21 22 23]have measured reaction times of plane pilots performinga tracking task and at the same time having to respondto warning signals (based on Edworthy et al [13]) Simi-larly Suied et al [24] measured reactions times of listen-ers performing a tracking task for signals used to warn cardrivers against too short of a distance between vehiclesUsing a driving simulator Belz et al [25] measured timesto brake when drivers were presented with different au-ditory and visual warning signals Interestingly the reac-tion times were different according to the relationship be-tween the signal and its meaning symbolic (arbitrary rela-tionship) or iconic (representational relationship eg thesound of breaking glass used to signal danger of a poten-tial accident) For acoustic signals displayed alone shortertimes to brake were obtained for iconic signals Graham[26] used a similar paradigm but participants had to de-cide how to react to the warning The results showed thatauditory icons2 lead to more false alarms than symbolicsounds because the sounds could have several (natural)meanings whereas symbolic sounds have only one (arbi-trary) meaning Among these symbolic sounds car hornslead to faster reaction times and fewer false alarms becauseof their lack of ambiguity

This indicates that the meaning of a warning compre-hended by the listener can play an important part in thewarning process This idea appears also in the results ofGuillaume et al [27] They replicated the experiments ofEdworthy et al [13] both with sounds synthesized accord-ing to Edworthyrsquos specifications and with recordings ofalarms occurring in a plane cockpit Overall their resultsfit Edworthyrsquos prediction of perceived urgency on the basisof the acoustical properties of the sounds However theyalso report interesting exceptions for instance one soundidentified as a bicycle bell was rated as less urgent than itshould have been according to Edworthyrsquos prediction Theauthors suggest that this sound was judged as non-urgent

2 Auditory icons are sounds used to convey a meaning when there is arelation of similarity between the sound and the meaning (eg the soundof an analogical camera meaning that a digital photo has been shot)

357
















358


















359















360


0

20731

50

69793

100

Membershipagreement

(in)

Sounds




Class 5 prototype


Class 9 ship horn




Sound Classes












mioi is





361


05 1 15 2 25

20

25

30

35

40

45

Roughness (asper)

6 7 8 9 10 11

minus20

minus20

0

0

20

20

40

40



300 350 400 450 500 550 600 650



0

1

2

3

456789101112

300

350

400

450

500

550

600

650

700

Roughness(asper)


F0(Hz)











362














S(t) =N

i=1








363


Time

Frequency

Amplitude

A Excitation

ω

ω +ε0

0

i

B Temporal envelope


Time

Frequency

Frequency

D Filter




r 037 281 053 291c 1380 3790 1180 4560d 530 116 285 116f 262 683 198 500







364



dijt =R

r=1







-04 -03 -02 -01 0 01 02 03 04 05

-03

-02

-01

0

01

02

03

04

05

Dimension 1

Dimension2








365






42 Results









Representative

Non-representative

Absence ofconsensus



0

193

323

50

677

806

100



Sounds





366








05 1 15 2 25

10

20

30

Roughness (asper)

3 4 5 6 7 8 9 10 11

0

20

40



200 250 300 350 400 450 500 550 600 650

minus20

0

20



0

1

2

3

24

681012

200

300

400

500

600

700

Roughness(asper)


F0(Hz)



367














368








Goal of the study


Procedure


Note



Goal of the study


Procedure


369




Acknowledgements


References






























370







































371












372
















358


















359















360


0

20731

50

69793

100

Membershipagreement

(in)

Sounds




Class 5 prototype


Class 9 ship horn




Sound Classes












mioi is





361


05 1 15 2 25

20

25

30

35

40

45

Roughness (asper)

6 7 8 9 10 11

minus20

minus20

0

0

20

20

40

40



300 350 400 450 500 550 600 650



0

1

2

3

456789101112

300

350

400

450

500

550

600

650

700

Roughness(asper)


F0(Hz)











362














S(t) =N

i=1








363


Time

Frequency

Amplitude

A Excitation

ω

ω +ε0

0

i

B Temporal envelope


Time

Frequency

Frequency

D Filter




r 037 281 053 291c 1380 3790 1180 4560d 530 116 285 116f 262 683 198 500







364



dijt =R

r=1







-04 -03 -02 -01 0 01 02 03 04 05

-03

-02

-01

0

01

02

03

04

05

Dimension 1

Dimension2








365






42 Results









Representative

Non-representative

Absence ofconsensus



0

193

323

50

677

806

100



Sounds





366








05 1 15 2 25

10

20

30

Roughness (asper)

3 4 5 6 7 8 9 10 11

0

20

40



200 250 300 350 400 450 500 550 600 650

minus20

0

20



0

1

2

3

24

681012

200

300

400

500

600

700

Roughness(asper)


F0(Hz)



367














368








Goal of the study


Procedure


Note



Goal of the study


Procedure


369




Acknowledgements


References






























370







































371












372


















359















360


0

20731

50

69793

100

Membershipagreement

(in)

Sounds




Class 5 prototype


Class 9 ship horn




Sound Classes












mioi is





361


05 1 15 2 25

20

25

30

35

40

45

Roughness (asper)

6 7 8 9 10 11

minus20

minus20

0

0

20

20

40

40



300 350 400 450 500 550 600 650



0

1

2

3

456789101112

300

350

400

450

500

550

600

650

700

Roughness(asper)


F0(Hz)











362














S(t) =N

i=1








363


Time

Frequency

Amplitude

A Excitation

ω

ω +ε0

0

i

B Temporal envelope


Time

Frequency

Frequency

D Filter




r 037 281 053 291c 1380 3790 1180 4560d 530 116 285 116f 262 683 198 500







364



dijt =R

r=1







-04 -03 -02 -01 0 01 02 03 04 05

-03

-02

-01

0

01

02

03

04

05

Dimension 1

Dimension2








365






42 Results









Representative

Non-representative

Absence ofconsensus



0

193

323

50

677

806

100



Sounds





366








05 1 15 2 25

10

20

30

Roughness (asper)

3 4 5 6 7 8 9 10 11

0

20

40



200 250 300 350 400 450 500 550 600 650

minus20

0

20



0

1

2

3

24

681012

200

300

400

500

600

700

Roughness(asper)


F0(Hz)



367














368








Goal of the study


Procedure


Note



Goal of the study


Procedure


369




Acknowledgements


References






























370







































371












372















360


0

20731

50

69793

100

Membershipagreement

(in)

Sounds




Class 5 prototype


Class 9 ship horn




Sound Classes












mioi is





361


05 1 15 2 25

20

25

30

35

40

45

Roughness (asper)

6 7 8 9 10 11

minus20

minus20

0

0

20

20

40

40



300 350 400 450 500 550 600 650



0

1

2

3

456789101112

300

350

400

450

500

550

600

650

700

Roughness(asper)


F0(Hz)











362














S(t) =N

i=1








363


Time

Frequency

Amplitude

A Excitation

ω

ω +ε0

0

i

B Temporal envelope


Time

Frequency

Frequency

D Filter




r 037 281 053 291c 1380 3790 1180 4560d 530 116 285 116f 262 683 198 500







364



dijt =R

r=1







-04 -03 -02 -01 0 01 02 03 04 05

-03

-02

-01

0

01

02

03

04

05

Dimension 1

Dimension2








365






42 Results









Representative

Non-representative

Absence ofconsensus



0

193

323

50

677

806

100



Sounds





366








05 1 15 2 25

10

20

30

Roughness (asper)

3 4 5 6 7 8 9 10 11

0

20

40



200 250 300 350 400 450 500 550 600 650

minus20

0

20



0

1

2

3

24

681012

200

300

400

500

600

700

Roughness(asper)


F0(Hz)



367














368








Goal of the study


Procedure


Note



Goal of the study


Procedure


369




Acknowledgements


References






























370







































371












372


0

20731

50

69793

100

Membershipagreement

(in)

Sounds




Class 5 prototype


Class 9 ship horn




Sound Classes












mioi is





361


05 1 15 2 25

20

25

30

35

40

45

Roughness (asper)

6 7 8 9 10 11

minus20

minus20

0

0

20

20

40

40



300 350 400 450 500 550 600 650



0

1

2

3

456789101112

300

350

400

450

500

550

600

650

700

Roughness(asper)


F0(Hz)











362














S(t) =N

i=1








363


Time

Frequency

Amplitude

A Excitation

ω

ω +ε0

0

i

B Temporal envelope


Time

Frequency

Frequency

D Filter




r 037 281 053 291c 1380 3790 1180 4560d 530 116 285 116f 262 683 198 500







364



dijt =R

r=1







-04 -03 -02 -01 0 01 02 03 04 05

-03

-02

-01

0

01

02

03

04

05

Dimension 1

Dimension2








365






42 Results









Representative

Non-representative

Absence ofconsensus



0

193

323

50

677

806

100



Sounds





366








05 1 15 2 25

10

20

30

Roughness (asper)

3 4 5 6 7 8 9 10 11

0

20

40



200 250 300 350 400 450 500 550 600 650

minus20

0

20



0

1

2

3

24

681012

200

300

400

500

600

700

Roughness(asper)


F0(Hz)



367














368








Goal of the study


Procedure


Note



Goal of the study


Procedure


369




Acknowledgements


References






























370







































371












372


05 1 15 2 25

20

25

30

35

40

45

Roughness (asper)

6 7 8 9 10 11

minus20

minus20

0

0

20

20

40

40



300 350 400 450 500 550 600 650



0

1

2

3

456789101112

300

350

400

450

500

550

600

650

700

Roughness(asper)


F0(Hz)











362














S(t) =N

i=1








363


Time

Frequency

Amplitude

A Excitation

ω

ω +ε0

0

i

B Temporal envelope


Time

Frequency

Frequency

D Filter




r 037 281 053 291c 1380 3790 1180 4560d 530 116 285 116f 262 683 198 500







364



dijt =R

r=1







-04 -03 -02 -01 0 01 02 03 04 05

-03

-02

-01

0

01

02

03

04

05

Dimension 1

Dimension2








365






42 Results









Representative

Non-representative

Absence ofconsensus



0

193

323

50

677

806

100



Sounds





366








05 1 15 2 25

10

20

30

Roughness (asper)

3 4 5 6 7 8 9 10 11

0

20

40



200 250 300 350 400 450 500 550 600 650

minus20

0

20



0

1

2

3

24

681012

200

300

400

500

600

700

Roughness(asper)


F0(Hz)



367














368








Goal of the study


Procedure


Note



Goal of the study


Procedure


369




Acknowledgements


References






























370







































371












372














S(t) =N

i=1








363


Time

Frequency

Amplitude

A Excitation

ω

ω +ε0

0

i

B Temporal envelope


Time

Frequency

Frequency

D Filter




r 037 281 053 291c 1380 3790 1180 4560d 530 116 285 116f 262 683 198 500







364



dijt =R

r=1







-04 -03 -02 -01 0 01 02 03 04 05

-03

-02

-01

0

01

02

03

04

05

Dimension 1

Dimension2








365






42 Results









Representative

Non-representative

Absence ofconsensus



0

193

323

50

677

806

100



Sounds





366








05 1 15 2 25

10

20

30

Roughness (asper)

3 4 5 6 7 8 9 10 11

0

20

40



200 250 300 350 400 450 500 550 600 650

minus20

0

20



0

1

2

3

24

681012

200

300

400

500

600

700

Roughness(asper)


F0(Hz)



367














368








Goal of the study


Procedure


Note



Goal of the study


Procedure


369




Acknowledgements


References






























370







































371












372


Time

Frequency

Amplitude

A Excitation

ω

ω +ε0

0

i

B Temporal envelope


Time

Frequency

Frequency

D Filter




r 037 281 053 291c 1380 3790 1180 4560d 530 116 285 116f 262 683 198 500







364



dijt =R

r=1







-04 -03 -02 -01 0 01 02 03 04 05

-03

-02

-01

0

01

02

03

04

05

Dimension 1

Dimension2








365






42 Results









Representative

Non-representative

Absence ofconsensus



0

193

323

50

677

806

100



Sounds





366








05 1 15 2 25

10

20

30

Roughness (asper)

3 4 5 6 7 8 9 10 11

0

20

40



200 250 300 350 400 450 500 550 600 650

minus20

0

20



0

1

2

3

24

681012

200

300

400

500

600

700

Roughness(asper)


F0(Hz)



367














368








Goal of the study


Procedure


Note



Goal of the study


Procedure


369




Acknowledgements


References






























370







































371












372



dijt =R

r=1







-04 -03 -02 -01 0 01 02 03 04 05

-03

-02

-01

0

01

02

03

04

05

Dimension 1

Dimension2








365






42 Results









Representative

Non-representative

Absence ofconsensus



0

193

323

50

677

806

100



Sounds





366








05 1 15 2 25

10

20

30

Roughness (asper)

3 4 5 6 7 8 9 10 11

0

20

40



200 250 300 350 400 450 500 550 600 650

minus20

0

20



0

1

2

3

24

681012

200

300

400

500

600

700

Roughness(asper)


F0(Hz)



367














368








Goal of the study


Procedure


Note



Goal of the study


Procedure


369




Acknowledgements


References






























370







































371












372






42 Results









Representative

Non-representative

Absence ofconsensus



0

193

323

50

677

806

100



Sounds





366








05 1 15 2 25

10

20

30

Roughness (asper)

3 4 5 6 7 8 9 10 11

0

20

40



200 250 300 350 400 450 500 550 600 650

minus20

0

20



0

1

2

3

24

681012

200

300

400

500

600

700

Roughness(asper)


F0(Hz)



367














368








Goal of the study


Procedure


Note



Goal of the study


Procedure


369




Acknowledgements


References






























370







































371












372








05 1 15 2 25

10

20

30

Roughness (asper)

3 4 5 6 7 8 9 10 11

0

20

40



200 250 300 350 400 450 500 550 600 650

minus20

0

20



0

1

2

3

24

681012

200

300

400

500

600

700

Roughness(asper)


F0(Hz)



367














368








Goal of the study


Procedure


Note



Goal of the study


Procedure


369




Acknowledgements


References






























370







































371












372














368








Goal of the study


Procedure


Note



Goal of the study


Procedure


369




Acknowledgements


References






























370







































371












372








Goal of the study


Procedure


Note



Goal of the study


Procedure


369




Acknowledgements


References






























370







































371












372




Acknowledgements


References






























370







































371












372







































371












372












372

Documents

0 0 -3 ) ) * * 2 $) * - - 1 0 · 2014-08-18 · 0 0 -3 ) ) * * 2 $) * - - 1 0 ... .,5