An articulatory study of fricative consonants using magnetic resonance imaging ShrikanthS. Narayananand Abeer A. Alwan•/ SpeechProcessing and Auditoo,Perception Laborator3; Department of ElectricalEngineering, UCLA, 405 Hilgard Avenue,Los Angeles,Cal!fornia 90024 Katherine Hakerb) ImagingMedical Group, Cedars-Sinai Medical Center,8700 BeverlyBoulevard, /.os Angeles, California 90048 (Received6 September1994; accepted for publication3 April 1995) Magneticresonance images of thevocal tract during sustained production of thefricatives/s, •, f, 0, z, 3, v, 6/by foursubjects are analyzed. Measurements of vocal-tractlengths and area functions, and morphologicalanalyses of the vocal tract and tongueshapes for these soundsare presented. Interspeakerdifferences in areafunctions are found to be greaterin thepharyngeal cavity than in the buccalcavity with the nonstriden:fricatives exhibiting greater differences than the stridentones. The anteriortongue body of the alveolarstridents exhibit concave cross-sectional shapes while that of the postalveolarsshow a relativelyraised tongue body with fiat or slightlyconvex cross-sectional shapes.The concavetongue shapes of the alveolarsresult in a moreabrupt area function behind the constrictionwhen compared to that of the postalveolars.Laminality or apicalityof articulationis foundto be speakerdependent. Moreover, a greaterdegree of anteriorroedial grooving and lateral lingua-palatalcontact is foundin apicalalveolar fricatives than in laminalones. The posterior tongue body of all fricatives shows concavecross-sectional shapes. Voiced fricatives are characterizedby larger pharyngealvolumes than the unvoicedfricatives due to tongue-root advancement.Tongue-shape asyr•metries arc found to be subjectand, in some cases,sound dependent.¸ 1995 AcousticalSociety of America. PACS numbers: 43.70.Aj, 43.70.Bk, 43.70.Jt

INTRODUCTION channelturbulence (jet emergingfrom a constriction)and waketurbulence (jet impingingon an obstacle).In addition Knowledgeof the 3-D geometry of the human vocal to turbulence,the vocalfolds may vibrate,at leastfor part of tract is an importantfactor in modelingspeech production. the frication period, as in the case of the voiced fricatives. Magneticresonance imaging (MRI) is a powerfultool in The eight ti'icativeconsonants in English,specified in terms obtainingthe vocal-tractgeometry and doesnot involveany of their placeof articulationin unvoiced-voicedpairs, are: knownradiation risks. The imageshave good signal-to-noise The labiodentals/f/ and /v/, interdentals/0/and/6/, alveolars ratio(SNR) and are amenable to computerized3-D modeling /s/ and /z/, and postalveolars/,[/ and /3/. The aerodynamic of the vocal tract. The low image samplingrate (i.e, high behaviorin the alveolarand postalveolarfricatives manifest acquisitiontime), however,has restrictedMRI use to the more wake turbulence characteristics,and are often referred studyof sustainedspeech sounds, correspcnding to "static" to as sibilant or strident fricatives. The labiodentals and in- tractshapes. Previous MRI studieshave been mostly limited terdentals, on the other hand, exhibit more channel turbu- to vowels(Baer et al., 1991; Moore, 1'992;Greenwood et al., lence, and are often referred to as nonsibilant or nonstrident 1992)and nasals (Dang et al., 1993).In thispaper, a detailed fricatives.Turbulence generation, or sourcemechanisms, for analysisof the vocal tract geometryobtaired from MR im- fricatives are, however, not completelyunderstood. An ar- aging in axial, coronaland sagittalplanes. of the fricatives ticulatoryand acousticdescription of fricativesin different /s, ,[, f, 0, z, 3 v, 6/in Englishis reported. languagesis foundin Ladefogedand Maddieson(1986). A. Fricative mechanisms In pre',iousstudies, information regarding the vocaltract geometryof fricative consonantshas been mainly derived Fricatives are produced by the lbrmation of a narrow from lateral x-ray data (Perkell, 1969; Subtelnyet al., 1972; supraglottalconstriction in the vocaltract and the generation Badin, 1'991).Midsagittal profiles of sibilantswere found to of turbulencein theregion downstream from theconstriction exhibitgreater consistency across varying phonetic contexts when air flows throughthe vocal tracl:(Fant, 1960; Stevens, in mandibleand tongue positions when compared to pharyn- 1971).The generationof turbulence,occurs near the vocal- geal shapingand lip opening(Subtelny et al., 1972). A ten- tract walls and/or the teeth which may act as an obstacleto dencytoward increased pharyngeal volumes was observedin the airflow. Catford (1977) classifiesthese mechanismsas the voiced sibilant /z/ when comparedto the unvoiced /s/ (Perkell,1969). The mainli•nitations of x raysinclude radia- :øAuthorto whomcorrespondence should be addressed. tion risks and difficulty in accuratelydeducing the cross- b)Also,Department of Radiology, UCLA School of Medicine. sectionalmorphology from mid-sagittalprofiles. Other tech-

1325 J. Acoust.Soc. Am. 98 (3), SeptemberlC•95 0001-4966/95/98(3)/1325/23/$6.00 ¸ 1995 AcousticalSociety of America 1325 niques such as ultrasound(Stone etal., 1992), static FRONT REGION palatography(Ladefoged, 1957), and dynamic electro- palatography(Fletcher et al., 1989; Hardcastleand Clark, hardpalate ='. 1981;Hoole et al., 1989) can alsoprovide important articu- latory information. Palatographicstudies have revealed a moreposterior and wider constriction for/•/when compared to/s/. Lateralx rays,palatographic data, and ultrasound data BACK have suggestedcross-sectional tongue grooving in English REGION alveolar fricatives. Other studies have documented intersub- mandible ject variabilitiesin the apical versuslaminal manner of ar- ticulationfor stridentfricatives (Dart, 1991). Articulatory asymmetrieshave also been observed in sibilantproductions (Hamletet al., 1986;Stone et al., 1992).Lingual asymme- tries in normal consonant articulations are, in fact, wide- spreadacross different sounds and languages (Marchal et al., 1988).None of thesestudies, however, provides a descrip- FIG. 1. Tracingof the mid-sagittalprofile of the vocal tract duringthe tion of the entire vocal tract duringthe productionof frica- productionof thevowel /a/ (subjectMI) highlightinglandmarks used for tives. MR imaging, on the other hand, could be usedto ob- lengthmeasurements. tain a detailed descriptionof the vocal tract; such 3-D geometryis crucialfor modelingfricative production mecha- respectively,based on a mid-sagittallocalizer image for each nisms(Shadle, 1991). subject.Similarly, the scanningregion for the sagittalplane was basedon axial and/orcoronal 1ocalizer images. The data I. METHOD set comprised28 to 35 images/sound/subjectin the sagittal plane, and 40 to 45 •images/sound/subjectin the axial and A. Subjects coronalplanes. During scanning,the speakerssustained each Four phoneticallytrained, native American English consonantfor about 13-16 s enablingfour to five image speakers[two males(MI, SC) and two females(AK, PK)] slicesto be recorded(about 3.2 s/image).The consonants servedas subjects.Subjects AK and MI, both in their twen- were producedin a VC contextwith the neutralvowel/o/. ties, were raisedin NorthernCalifornia and have spentthe The subjectsrepeated each soundsix to nine times, with a past sevenyears in SouthernCalifornia. Subject SC, in his pauseof 3 to 10 s betweenrepetitions, to enablethe entire thirties,spent the firstten yearsof his life in Indianaand has vocal tract to be scanned. Verbal communication was main- since been in California. Subject PK, in her early forties, tainedwith the subjectsthrough an intercomsystem. lived in New Jerseyand Ohio duringher first threeyears, and The dataused in this studywere collectedover a period in the Bostonarea throughher thirties.Since then she has of six months using the same scannerat the Cedars-Sinai beenin the Los Angelesarea. MedicalCenter (Los Angeles). Pilot studiesalong with cali- brationexperiments preceded the actualdata collectionby B. Image acquisition threemonths. Each subjectwas scannedin differentsessions ondifferent days. Scanning o'f.e'hch subject inany one par- Magneticresonance (MR) imageswere collected using a ticularplane (axial, for example),was completed within the GE 1.5 T SIGNA machinewith a fastSPGR (radio frequency samesession, which typically lasted for about1-} to 2 h. spoiled GRASS) protocol (TE--4.0 ms, TR--12.6 ms, Scanningin eachof the threeorthogonal planes was carried NEX=2, FOV=20 or 24 cm).The imageslice thickness was out in different sessions.In addition to fricatives, the data set 3 mm with no interscanspacing. Each image was repre- also containedvowels and liquids. sentedby a 256X256 pixel matrix, yielding a resolutionof 0.0088cm 2 per square-pixelfor an FOV=24 cm (about C. Image analysis 0.094cm/pixel-edge). During. scanning, subjects assumed a supineposition. A specialhead-neck coil, by Medical Ad- The data were processedon a special-purposeMR im- vances,which helped maintain the subjects'heads in a fixed age processingworkstation (ISG-Allegro, Silicon Graphics position,was used to enhancethe SNR of the images.In basedsystem). All dataprocessing was performed by thefirst order to provide a convenientreference to key anatomical author in a dark radiologyreading room. Image processing landmarksin the vocal tract region, tracing of a sample mid- was performed in several steps.The first step involved seg- sagittalMR imageof the vocaltract for the vowel/aJ spoken -menting the regions of interest in the image. Preliminary by a male subjectis shownin Fig. l. segmentationof the vocal tract airway was achievedusing an The scanningregion for the coronal and axial planes automaticthresholding procedure that relies on the contrast includedthe region betweenthe lips and the posteriorpha- betweenthe airway and the surroundingtissues (the airway ryngeal wall along the antero-posterioraxis and the region appearingrelatively darker than most of the surroundingsoft betweenthe top of the hardpalate and just belowthe eighth tissue).Due to variabilitiesand complexitiesin the vocal- vertebraalong the infero-superioraxis. Coronal and axial tractmorphology (for example,the lower-pharyngealand la- scanswere alignedto be approximatelyperpendicular to the ryngealregions) and nonuniformitiesin the imagecontrast vocal-tractmidlines, in the buccal and pharyngealregions, (for example,regions near the teeth) automatic thresholding

1326 J. Acoust. Soc. Am., Vol. 98, No. 3, September 1995 Narayanan et al.: MRI study of fricativeconsonants 1326 proceduresmay frequentlybe insufficientand/or result in I. Coronal measurements erroneoussegmentation. Hence, automatic thresholding was Analysisof thefront region yields important information followedby a carefulverification of the selectedregions in regardingthe fricatives'constriction geometry and cross- eachimage, and appropriate boundary corrections were made sectionaltongue shapes. Coronal sections in the buccalcav- manually.Various aids suchas radiology/anatomyatlases ity beginningat thelips and ending near the posterior edge of (Schnitzleinand Murtagh, 1990), and dental casts of thesub- the back of the tonguebody, showing no pharyngealsec- jectswere usedto ensureaccurate segmentation. Following tions, were consideredfor area-functionmeasurements in the segmentation,three dimensionalreconstruction of the entire front region.The numberof sectionsconsidered for areacal- vocal tract, or specificregions such as sublingualcavities culationsdiffered slightly dependingon the subject's andpirifonn sinuses,were madeby computer-aidedconcat- anatomyand soundproduced. enationof the selectedregions of interest.Length, area, and Estimatesof the minimumsupraglottal constriction area volumemeasurements were made directly fiom anyspecified (A•.) were obtaineddirectly from raw coronal scans.Al- regionof interestin the raw and/orreconstructed 3-D images thoughthe image slice corresponding to A c givesan approxi- usinga pixel countingalgorithm. Images scanned from a mate indication(within -+3 mm) of the locationof the con- particularplane (coronal, for example)could be usedto ob- striction(x,:), a more accuratex• estimatecould be made tain cross-sectionalinformation along ary other desired from the mid-sagittalprofiles, as is describedin Sec. II D 3. plane(arbitrary oblique sections, for example)through com- Areas and volumesof the sublingualcavities, such as those puterized image reformatting.Such a procedure,however, observedduring the productionsof/j7 and/3/, werealso mea- resultsin thedegradation of imagequality when compared to sured.Airway areasbelow the tonguesurface that were dis- the original scans. tinctly separatedfrom thoseabove the tonguesurface, ob- In this study,coronal and axial scans•3xe used to mea- served in the coronal sectionsof the front region, were sureand analyze the cross-sectional areas and morphology of identifiedas sublingualcavity components.Volumes were the "front"region (buccal cavity extendin• from the lipsto obtainedby 3-D reconstructionof thesesublingual compo- the posteriorpharyngeal wall) and "back"region (pharyn- nents. gealcavity extending from the uvulato thebeginning of the Two problemswere faced when calculatingthe cross- trachea),respectively. sectionalareas in the front region: (1) Therewere difficulties in specifyingthe boundaries at the regionssideways to the lips, typically in the first sec- D. Area function and length measurem•nts tion throughthe lips. In most cases,however, approximate Cross-sectionalareas were directl[ymeasured from the boundariescould be specifiedbased on the upperand lower coronaland axial scansto provideinfbrmation on the front lip boundaries.Frames from a video tape of the front and andback regions, respectively. Due to the slightcurvature of sidesof the mouth,recorded from the subjectson a later the vocal tractin the front and back regions,the areavalues date, were also used to aid in the segmentationof the lip obtained from the raw coronal and axial slices differed from areas. Video recording was done while subjectsassumed a thosemeasured along the planesperpendicular to the midline supineposition similar to that assumedinside the MRI scan- of the vocal tract by approximatelya cosne factor of the ner. anglebetween the two planes.Nevertheless. since the curva- (2) Unclearteeth boundaries with respectto the airway ture is small,the raw coronaland axial scaasstill providea also posedproblems, particularly in the anteriorpart of the fairly accuraterepresentation of the frontand backregions, frontregion. Several measures were undertaken in an attempt respectively.Area-function estimates in previousstudies to enhancethe contrastbetween the teeth and the airway. haverelied on a similarassumption (Ladefoged et al., 1971; Thesemeasures included smearing the teethwith mineraloil, Mermelstein,1973; Baer et al., 1991).Raw ,zoronaland axial with a thin coatof paraffinwax and with a pasteof of about scans,however, do not yield usefularea informationalong 1 cc of GD (gadolinium)in about30 cc of a BariumSulfate the vocal tract'sbend and imagereformatting becomes un- Esophagealcream (EZ-paste).None of these measures avoidable.Since mid-sagittal profiles provide the mostcon- yielded satisfactoryresults. Hence, we relied on measuring venientreference for specifyinggrid locationsfor perform- the dimensions of the oral structure, such as teeth sizes and ing area calculations,sagittal scans are chosenfor area the distancefrom the incisorsto the alveolar ridge using calculationsfrom reformattedimages, particularly along the precision callipers from each subject'sstone dental casts vocaltract bend.Mid-sagittal data are',alsc. used for length (Baeret al., 1991).Dental impressions, made nsing Alginate measurements. Plus dental material from casts, were used to obtain cross- Our calibrationexperiments indicated [hat, on average, sectionalslices in both coronaland sagittalplanes, at ap- overestimatederrors in areameasurements :?rom raw images proximately 3-mm intervals. Outlines of thesesections were ranged between 2%-8%; errors in volun•e measurements thencarefully traced on paper.Separate tracings of the upper rangedbetween 1.3%-8.0%, andthe averag• error for length and lower teeth,together with clearly visible structuressuch was 3.4%. In general, measurementof smaller dimensions as the teethroots, mandible, and palate were also madefrom resulted in larger errors. Errors in the area measurements coronal MRI scanswhere the tongue was kept braced against from reformattedsections ranged between 3 %-10%. Details the inner teeth boundaries,thereby enabling a fairly easy regardingthe calibrationexperiments are given in the Ap- segmentationfrom the airway.As a first cut, segmentationat pendix. the teeth boundarieswas carefullymade by followingthe

1327 J. Acoust.Soc. Am., Vol.98, No. 3, September1995 Narayananet aL: MRI studyof fricativeconsonants 1327 /f/N[ /th/ HI NgV(•) HSV(b)

HSV my (0

FIG. 2. Mid-sagittalprofiles of thevocal tract during the produclionof (a) If/, (b)/0/, (c) Is/, and(d) /•/ (subjectMI). Notethat in thisand subsequent figures, the scannerlabels use ARPABET notationsuch that hh/refers to/0/, and Ishl refersto/•/. gingivaloutline and employingknowledge of eachsubject's regionpose problems due to the appearanceof severalcavi- oral structure and measurements obtained from dental casts ties that result fiom various tissue structures such as the and dentalimpressions. The next stepinvolved the useof the uvula, epiglottis,and vocal folds, along the vocal tract air- templatesobtained from the teeth tracings to improve the way. [u an attemptto make the areacalculations in a system- segmentationaccuracy. Measurements obtained from dental atic manner.the back region was divided into several zones: castsand impressionswere particularlyuseful in the front 1. Uvu/ar regio.: This region is defined from the first teeth region while the tracings obtained from MR[ scans effective axial section up to the section where the were requiredin the regionbehind the alveo]at ridge. uvula disappears. Three-dimensionaltongue shapes were constructedfi'om 2. Upper-pha•3'.geal regio.: This region is defined coronal scans:Tongue outlines for each sound were seg- from the end of the uvular regionto the sectionwhere mentedfrom the correspondingcoronal scan set and usedfor the presenceof the epiglottisis markedby the ap- computerized3-D reconstruction.Fairly accuratetongue- pearanceof its crescent-shapedfree margin.This re- shaperepresentations were possible due to the availabilityof giou showsno cavity divisiousand area calculations closelysampled, contiguous image sections (3 mmthickness, are straightforward. no interscanspacing). 3. Lower-pharyngeal regio.: This legion is defined from the end of the upper pharyngealregion to just 2. Axial measurements before the sectionwhere the side pirilb•m sinusesare Areas in the back region were calculatedfi-om the first completely separatedby the interarytenoideminence effective slice in the uvular region, the slice appearingdis- into distinctcavities with respectto the centrallaryn- tinct from the buccalcavity, to the beginningof the trachea geal vestibule.In the initial sectionsof this region, (typicallyat the mid-levelof the interve]tebraldisk between crosssections may reveal a small anterior epiglottic- the fifth and sixth vertebra).The areacalculations in the back vallecularcontribution separated from the relatively

1328 J. Acoust. Soc. Am., Vol. 98, No. 3, September 1995 Narayanan et al.: MRI study of fricativeconsonants 1328 /v/ #! / h/HI

lt/#I /•h/'#l NSV(c) N3V(d)

FIG.3. Mid-sagittalprofiles of thevocal tract during the production of (a) Iv/, (b)/6/, (c)/z/, and(d) /3/ (subjectMI). Notethat in thisand subsequent figures, the scannerlabels use ARPABET notationsuch that Idhl refers to 161,and I•hl refers to/3/.

larger pharyngealairway by the medial glossoepig- suredseparately. In the lower-pharyngealregion, the bound- lottic folds.The pharyngealairway may showfurther aries of the pitiform sinuseswhich remain connectedto the cavity divisionsdue to the presenceof the phatyn- glottalcavity, were approximated by drawinga line alongthe goepiglotticfolds. Toward the end of this region,the aryepigiotticl•)ld endingat the centerof the posterioredge vallecular contributionsdisappear and the laryngeal of the pharynx.The volumesof the left and right piriform inlet with the recessesof the piriform sinuses,still sinuses were made fi'om the corresponding three- connectedto the centralcavity, appear.The aryepig- dimensional reconstructions. lottic foldsprogressively get thickeras the glottisis To enable comparative graphical analyses across the approached,thus increasing the separation of thepiri- various soundsand subjects,a simplified representationof form sinusesfrom the central lm'yngeal vestibule. the areafunction is presented.Areas up to the !atyngealinlet, Measurementsof theseareas are not straightforward defined by the sectionshowing the completeseparation of due to the connected-cavitystructures. the pitiform sinusesby the interawtenoideminence, are in- Lao,ngealregion: This regionis definedfrom the end cluded. Furthermore,the "eflbctive" area of the airway is of the lower-pharyngealregion to the sectionwhere obtainedby a simplificationof the morphology:Subtracting thepitiform sinuses disappear. In thisregion, the piri- tissue areas, such as the uvula, and the various epiglottal form sinusesappear distinct from the centralglottal folds, from the total pharyngealcavity areas. airway. Subglottalregion: This region is definedfrom theend 3. Sagittal measurements of the laryngealregion to the beginningof the tra- Sagittalscans were usedfor the following: chea,typically up to the mid-levelof the interverte- (1) Mid-sagittal profiles were used for various length bral disk between C5 and C6. measurements;a satnplemid-sagittal profile tracingshowing In general,the areaof eachindividual cavity was mea- relevant anatomical landmarks and the details of the various

1329 J. Acoust.Soc. Am., Vol. 98, No. 3, September1995 Narayananet al.: MRI studyof fricativeconsonants 1329 /s/ M! ½ • /s/ M! (c) JD VT (AV) 3D VT ½RLV)

lip openlng lip opening constrict Ion

pt r• form sinuses trachea

/s/ M! (b) / / aT 30 VT (PV) 3D • CLLV)

pharynx cons trl etlon

p I rl t:orm sinuses

trachea

FIG. 4. A 3-D modelof thevocal tract during the production of/s/(subject MI) reconshucted fi'om conreal scans. (a) AV anle•io•view. (b) PV--poste•ior view. (c) RLV--dght lateralview. (d) LLV left lateralview. length measurementsdescribed in this sectionis shown in from the midpointo1' the linc joining U and L to the Fig. 1. Let L and U mark the anterior-mostpoints of the point of minimum supraglottalconstriction. upperand lower lips, respectively.Let S mark the superior (2) Sagittalscans were t'efo•matted to obtaincross sec- point on the rear pharyngealwall, just behindthe uvula. tions along planes orthogonal to the midline of the vocal (a) The lengthof the vocal tract (Ivr) is definedto be tract bend and used for arezt function calculations in that the distancemeasured along the midline of the vocal region. Mid-liue tracing on the mid-sagittalprofiles corre- tract,beginning at the midpointof the line joining U spondingto each soundof each speakerwas first obtained and L to the midpoint of the perpendicularline in- basedon computer-aidedvisual estitnates:A numberof ap- tersectingthe trachea at the mid-level of the fifth proxi•natelyorthogonal sections were marked along the en- vertebra (C5). tire lengthof the vocal tract. and their respectivemid-points (b) Lengthof the front region(IF/O is definedto be the were identifiedbased on the evaluationof the corresponding horizontal distance between U and S. (tnid-sagittal)widths. The desired,slice locations tbr refor- (c) Lengthof the backregion (IBR) is definedto be the matting were then marked on these mid-sagimtlretirerice verticaldistance between the tangentat the superlin- profiles,approximately at about 3 to 4-ram interval, for the mostpoint on the hardpalate to the line perpendicu- whole lengthof the vocal tract,yielding a total of about45 to larly intersectingthe midline of the trachea at the 50 slices.Cross sections ahmg thesenew locationswere then level of C5. obtainedby retbrmattingthe raw sagittal image slice set cor- (d) Verticallip opening(lvo) is definedto be the length respondingto eachsound. The optionof i,•teractivelyobtain- of the line joining U and L. ing crosssections at any specifiedlocation, and along any (e) Locationof the supragiottalconstriction (x•) is de- arbitrary plane, served as a valuable tool in the analysis of fined to be the distancemeasured along the midline cross-sectionalmorphology.

1330 d. Acoust. Soc. Am., Vol. 98, No. 3, September 1995 Narayanan et aL: MRI study o[ fricativeconsonants 1330 (a) 30 Vr CRLV) 3D VT (R[V) f (•K) th (PK)

1 ip openln•l / Iip opening

(c) 30 VT (RI V) 30 VT (R[V) sh CPK) s (PK)

11p openinq l•p open1 ng

FIG. 5. A 3-D modeIof thevocal tract (right lateral view) fi)r theunvoiced fricatives (subject PK): (a)/f/, (b)/St, (c) Isl, (d) I•l.

(3) Mid-sagittalMRI datacould be usedfor comparative A. Front region analyseswith publishedmid-sagittal tracings from x-ray data. Analysis of the front region is basedon coronal scans and midsagittalprofiles. Sample coronal cross sections of the II. MORPHOLOGICAL ANALYSIS vocal tractduring the productionof/s/and/•[/(subject MI) Morphologicalanalysis of the vocal-tractshapes was are shown in Figs. 6 and 7, respectively.Area functions of penformed in severalsteps. First, overall tract shapeswere the front region, measuredfrom raw coronal images,of the analyzedusing mid-sagittal profiles and complete3-D mod- unvoiced and voiced fricatives for the four subjects are els were reconstructedfrom appropriatelysegmented raw shownin Figs. 8 and 9, respectively.The areascalculated in scans.All the 3-D reconstructionsreported in this study,ex- the final one or two sectionsof the front region, close to the cept thoseof the piriformsinuses, were constructedusing vocal tract bend, inay reveal artificially large values due to coronalscans; the piriform sinuseswere reconstructedusing the curving of the back of the tongue. Interpretationof the axial scans.Sample midsagittal profiles for the voicelessand area valuesin suchcases is made possibleby crossreferenc- voicedfricatives by subjectMI are shownin Figs. 2 and 3, ing with the correspondingmid-sagittal slices. Values of x c respectively.A sample3-D vocaltract showing anterior, pos- andA c are listedin TablesI and 11,respectively. terior,and lateralviews, is shownin Fig. 4 for/s/ (subject The front region tract shapeswere, in general, similar MI). Right lateralviews of the 3-D vocaltracts for the four voicelessfricatives (subject PK) areshown in Fig. 5. Further for the voiceless and voiced fricatives that share the same detailedanalysis of the cross-sectionalmorphology of the placeof articulation.The voicedfricatives, however, tend to front and backregions of the vocaltract are madewith coro- have slightly larger area values in the region immediately nal and axial scans,respectively. behind the constriction when compared to their voiceless

1331 J. Acoust.Soc. Am., Vol. 98, No. 3, September 1995 Narayanan et aL: MRI study of fricativeconsonants 1331 FIG. 6. Coronalprofiles of the vocaltract during tile ploduclionol /q (subjectMI) takenalong dillelent section.,,ill the fr,c)ntlegion. Distance,;. are measured with re,.,peclto the lip opening.(a) SectionthroLtgb Ille lips itt 0.6 Clll.(b) Seetk)nIlut•Ltgll the frontcadty at 1.5cm. (c) Ctm',tricti(mregion at 1.8 cm. (d) (b) Regionbehind the constrictionrevealing conca• e tongueshapes with distinctroedial groin ing: secti•n.ssh()wn are t,•kenat 2.4 to 4.8 cm, respectively.in steps of 0.6 cm. (i)-(k) Posteriorregion, near the vehlm,showing decreasing concavity m cms•-secti•naltongue ,,hape,,--•,ections ale takenat 5.4, 6.0, and 6.6 cm, respectively.(I) Sectionat 7.2 c,n •howingincreasing airway area as the prateriotpharyngeal wall is approached. counterparts.The posteriortongue region for the voiced fi'i- Smaller constriction widths. together with marked central catives is raised slightly higher than it is lot the wficeless tonguegrooving at Ihe constriction,however, resulted in a fricatives suggestingthe influence of tongue-rootadvance- more circular orificelike constriction, as was found in the meritobserved in the voicedcases (Figs. 2-3). The x• values and17/of subjectPK. The constrictionlbr/J7 and/3/occurred Ibr the voiced and unvoiced sounds ate more similar than the in the postem-alveoladantem-palatalregion, with the mini- A c values.Strident fi'icatives. in general.a•e chatacteri7edby mum constrictionoccurring aboul 5 to 10 trim behind that of s,nallerA c valuesthan the non-stridents;variability in theA, /s/and/z/. The constrictiou for/s/and/• was made with a values implies variability in the aerodynamics,such as the raisedtongue tip at the alveolarridge by MI and SC indicat- degreeof turbulence,and varying degreesof couplingbe- ingan apical t articulation: the tongue tip raising for PKwas tween the back and fi'ont cavities. not as distinct as that of MI and SC while the tongue blade was raised to titan the constrictionfi•r AK suggestinga laint- 1. Strident fricatives rial a•ticulation.For/•/and/31, the constrictioafor all sub- jects was characlcrizedby a raisedtongue blade, ratherthan a. Com'trictionregion. The cross-sectionalshape at the tongue tip. The constrictionhad a slitlike appearanceand constriction,in general,resembled a slit, ratherthan a circn- tendedto be slightly wider than that of/s/and /z/. The A c lar. o•ifice due to a flat tongue surfaceat the constriction. values,however. do not sl•ow a contrastingpattern between

1332 J. Acoust. Soc. Am., Vol. 98, No. 3, September 1995 Narayanan et al.: MRI study of fricative consonants 1332 FIG. 7. Coronalprofiles of the vocaltract during the prodnctiollof I•l (subjectMll takenalong d•ffe•cnl sccli•ns iu the froulmgiou. Distances are measmed with respectto the lip opening.(a) Sectionthrough the lips at 0.6 c,n showingdistiuct hp rimriding.i'bl Sectionthrough the horntax ity at 1.5 cm. (c)-(d) Sectionsat 2.1 and 2.4 cm showingdistinct sublingual areas. (el Secuonin the xicinity ol the constrictionat 2.3'cm. (l}-[h) Regionbehind the consniclion revealingflat/slighdy convex tongue shapes: sections •hown are takenat 3.3, 3.9. and 4.5 cm. respectively(i)-(I) Po•te,ior ,cgion.sho•ing increasing concavityin cross-sectionaltongue shapes--sections are takenat $.1 to 6.6 cm. respecuvely.in stepsof 0.6 cm. the alveolarand the postalveolarfi'icatives (Table Ill. For/•7 greaterthan the verticalopening). The correspondingsec- and /3/, significant right-left asymmetry was visible in the tionscould be approximatedby elliptical shapes,if the open constrictionsonly for AK and SC, with a larger right-side side bcmndariesare al•proxinuttedby 'rounded" .q•apes.In opening.The tonguecontour in the constrictionregion was the posteriorlip region,where the upperand lower lips were flat or slightly convex. in contact,the shapesappemed elliptical. b. Regio. i. front of the lingual constriction. Cross- The lenglh and the meas in the region in fi'ont oœthe sectionalshapes/areas in the region in front of the lingt,al constrictitmfi)r the postalveolarfricalives were gtcate• than constrictionare influencedby the individual's lip sizes and thoseof the alveolarsresulting in relativelylarge front cavity palatal morphologyin additionto any lip "shaping"behav- volumes.Although no systematicdifferences in shapesand ior. The front teeth act as an obstacleto the airflow. thereby playing a c•ucial role in the turbt,lence soundgeneration in areasof the anteriorlip-region sections were found,twtt.di.g strident fricatives. eflkcts(i.e., moreci•culm than laterallyelongated elliptical The cross-sectionalshapes neat' the lips were approxi- shapes)wele noticeablein the postel'k•rlip regionof/•,3 / [fi)r tnately elliptical. The sectionsin the anterior lip region, example,compare Figs. 6(a) and 7(a)]. Such elliptical where the "boundary"region sidewaysto the lip was open, shapes,with a tendencytoward smaller lateral elongations appeared"rectangular" (with the lateralwidth considerably (due to largervmtical airway openings)continued well into

1333 d. Acoust. Soc. Am., Vol. 98, No. 3, September 1995 Narayanan et al.: MRI study of fricative consonants 1333 FRONT AREAS BACK AREAS

(a)

• 2I '"'.,, . '-. ...-./. -•2 I • oI "-':• , .g0 / 0 2 4 6 8 -8 -6 -4 -2 0 distancefrom lips,cm distance to laryngeal inlet, cm

(b) (f)

/el

0 2 4 6 8 distancefrom lips,cm

(c) (g)

6-8 - /s/

•2(• [ .'...... ' ' ...... /..,:."--•"' •4,•2 ½o / go 0 2 4 6 8 -8 -6 -4 -2 0 distancefrom lips,cm distanceto laryngealinlet, cm

(d) (h)

... •6 -.•. ,. ,.", ß..'::.,,..,, .... ,. .•.• ....:' •0 / 2 4 6 8 -8 -6 -4 -2 0 distancefrom lips, cm distanceto laryngealinlet, cm

FIG. 8. Frontand back region area functions during the production of unvoicedfricatives by thefour subjects: AK (solid),PK (dashed),MI (dot-dashed),and SC (dotted).(a) and(e) Ifl, (b) and(O/0/, (c) and(g)/sl and,(d) and(h) I$l.The left-side panels [(a)-(d)] representfront region areas and the right side panels [(e)-(h)] representback region areas. the alveolarregion, resulting in the relativelylarger areaval- and, to a lesserextent of PK, however,were characterizedby ues found in thoseregions. a more laminal articulationand with no significm•tlowering c. Region behind the lingual constriction. As notedear- of the tonguebody behind the constriction.For all subjects, lier, the mid-sagittalprofiles for Is/ and /z/ of MI and SC the cross-sectionalshapes have flat to slightly convex con- showeda moreapical articulation with a markedlowering of toursin the constrictionregion which changesignificantly to the anteriortongue behind the constriction,with respectto concaveas the posteriorregion of the tonguebody is ap- the tip and back of the tonguebody. The Is/aim/z/of AK proached.Three-dimensional tongue images demonstrating

1334 J. Acoust.Sec. Am., VoL 98, No. 3, September1995 Narayanan et al.: MRI studyof fricativeconsonants 1334 FRONT ARE. iS BACK AREAS

(a)

6-8 •-6 -"-'• /v/ • 4• / ...... '• ..-.. ..' (• / '.."-.'--' '-"~-' -. -,•' • 21 ...'•",•....•-•,•>,.',•" •0/• , •,-'<-","--' ,, 0 2 4 6 8 -8 -6 -4 -2 0 distancefrom lips, cm distanceto laryngealinlet, cm

(b)

/a/

0 2 4 6 8 -8 -6 -4 -2 0 distancefrom lips, c'n distanceto laryngealinlet, cm

(g) ß-t /z/ ''",, 0 2 4 6 8 -6 -4 -2 0 distancefrom lips, cm distance to laryngeal inlet, cm

(d) (h) E

'"- . /3/

0 2 4 6 8 -8 -6 -4 -2 0 distancefrom lips,(:m distanceto laryngealinlet, cm

FIG. 9. Frontand back region area functions during he productionof the voicedfricatives by the four subjects:AK (solid),PK (dashed),MI (dot-dashed). andSC (dotted).(a) and(e)/v/. (b) and(13 161, (c) and(g)/z/and, (d} and (h)/3/. Theleft side panels [(a)-(d)] representfrout region areas and the right side panels[(e)-(h)] representback region areas. concavepostconstriction tongue shapes are shownin Fig. constrictiontongue concavily of MI was found to be the 10(g)-(i) (subjectMI) andFig. 11(g)-(i) isubjectPK). The moststri 'king while that of PK, the least striking.The greater concaveshape, which contributesto the large back cavity "degree" of postconstrictiontongue concavity is attributed volume, probably results from the bracing of the tongue to the presenceof additionalroedial tongue grooving. In par- againstthe hard palate along its sides ant a relatively de- ticular, the tongueshapes of subjectsMI and SC exhibited pressedtongue center. The degree of co •cavity, however, significantmediai groovingof lhe front tonguebody, about was speakerdependent. Among the four: ubjects,the post- 2-3 cm behind the constriction for MI and 1-2.2 cm behind

1335 J. Acoust.Soc. Am., Vol.98, No. 3, Sept.,mber1995 Narayananet al.: MRI studyo! fricativeconsonants 1335 TABLE I. Locationof minimumsupraglottal constriction (x c , cm) mea- raising,however, was speakerdependent, and was found to suredfrom lips. be most strikingfor AK.

Subject No significantdifferences were found between the cross- sectionaltongue shapes of Is/and Izl for any of the subjects. Fricafive AK PK MI SC The tonguebody behindthe constrictionfor I$l and 13/ /•/ 2.6 2.08 2.78 3.37 risesslightly along its midline,before it startssloping toward /3/ 2.62 2.24 2.34 3.24 /s/ 1.97 1.4 1.96 2.54 its posteriorend [for example,Figs. 2(d), 3(d), and 5(d)]. /z/ 1.97 1.75 2.0 2.22 This results in a relatively gradual increasein the corre- /0/ 0.95 O.89 1.01 1.51 spondingarea values behind the constrictionwhen compared 161 0.91 1.01 0.89 1.5 to Is/and/z/. The degreeof "palatality"(relative height of If/ 0.87 0.55 0.83 0.79 lvl 0.89 0.66 0.91 0.69 the front of the tonguewith respectto the back) was speaker dependent;palatality exhibited by SC was the moststriking while that of AK was not prominent.Palatality contributes to the constrictionfor SC, with the maximumroedial groove graduallyincreasing post-constriction area valuesin the di- depthsfor Isl and lzl reaching11.5 and 10.6 mm, respec- rectiontoward the posteriorpharyngeal wall [referto panel tively, for Ml and 8.6 and 10.4 ram, respectively,for SC (d) of Figs.8-9]. A higherdegree of palatalityis reflectedin [Fig. 12(a)-(b)]. This grooved-concavityeffect results in an a greateramount of increasein the postconstrictionarea val- abrupt increasein the area values behind the constriction. ues.The I•l and/3/of PK were differentfrom thoseof the SubjectsAK and PK, on the other hand, did not reveal no- othersubjects with the tongueraising occurring at the middle ticeablemedial tonguegrooving when comparedto MI or rather than at the anterior tongue body; the corresponding SC. The tonguecontour is ahnostflat for about 1 cm behind area functionsexhibited a plateauregion following a small the constrictionfor AK, with a more (anatomical)left- increase in values behind the constriction. favoredlingua-palatal contact. For AK, the rapidincrease in the area values behind the constriction is attributed to the The /.[,3/tongueshapes did not exhibit any noticeable subject'shighly domed palate, in additionto the concave concavityin the regionimmediately posterior to theconstric- tongueshape. The front tonguebody regionbehind the con- tion, unlike that observedin Isl and Izl. The 1•,31cross- stdcfionfor PK is concavebut with no grooving,and the sectionaltongue shapes behind the constrictionwere, how- maximumroedial depth reaches only 3.9 mm; the areafunc- ever, variableacross subjects. The 3-D tongueshapes for/•/ tion behindthe constrictiondid not show an abruptincrease are shownin Fig. 10(j)-(l) andFig. 11(j)-(1), for subjectsMI in values. and PK, respectively.Typical medial "groove" dimensions The following phenomenawere observedin the postcon- are shownin Fig. 12(c)-(d) for M1 andSC, respectively.For strictionregion, in the directiontoward the posteriorpharyn- MI and AK, a slight convex shape was observedwhich geal wall. graduallyturned concavetoward the posteriorregion. Both (1) For all subjects,a decreasingtrend in the degreeof PK andSC, on the otherhand, did not domethe tonguebody postconstrictionconcavity could be observed,for example, as illustratedin the posteriorview of the 3-D tongueshapes behind the constriction.Their tonguecontours were almost shownin Figs. 10(i) and 11(i) for Isl (subjectsMI and PK, flat and asymmetricalwith right-favoredlingua-palatal con- respectively).Asymmetry in the posteriorregion tongue tact.Previous studies have indicated that tongue asymmetries shapeswas found to be subjectdependent: Among the four in normalsibilant productions are not uncommon(Hamlet subjects,significant asymmetry is observedonly in PK [Fig. et al., 1986; Stone et ai., 1992). It is unclear whether the 11(i)]. observedasymmetries are typical or whetherthey are influ- (2) Raisingof the tongueback, which also contributes to enced by the supineposition assumedinside the scanner.It the decreasedarea values near the velar region,was observed shouldbe notedthat the subjects'dental casts did not reveal in all subjects[for example,refer to Figs.2(c) and3(c), Fig. any noticeableplatal asymmetries. 5(c) and,Figs. 10(i) and 11(i)]. The amountof tongueback In summary,the tongueshapes for Isl and Izl suggest that the apical or laminal nature of their production is TABLEII. Areaof minimumsupraglottal constriction (Ac , cm•-). speakerdependent. Moreover, the anteriortongue behavior is found to influence the tongue body shapesbehind the con- striction. The anterior tongue body behind the constriction Fricative AK PK MI SC reveals a relatively deeper grooving for the apical alveolars /$/ 0.262 0.098 O.112 O.174 when comparedto the laminal ones, resultingin a relatively 13/ 0.299 O.125 O.124 0.252 /s/ O. 146 0.098 O. 142 0.296 rapid increasein area valuesbehind the constriction.The /z/ O. 117 O. 136 0.159 0.247 pressuredrop due to lossesat the contractionand expansion /0/ 0.208 O. 175 O. 155 0.203 in the constrictionregion, on which the SPL of the turbu- 10/ 0.24 0.151 0.207 0.252 lence sourcedepends (Stevens, 1971), is predictedto be Ifl 0.259 0.2 0.142 0.193 Ivl 0.214 0.16 0.13 0.236 smaller for smooth transitions when compared to more abruptones. The postalveolarfricatives, however, do not ex-

1336 d. Acoust.Sec. Am., Vol. 98, No. 3, September1995 Narayananet al.: MRI study of fricativeconsonants 1336 j'Df/tonqueI•! (AV) (a) If/3l) tongue ! (RIV) (•) /3•/tongueMI

/

posterior tongue antertor tongue ante½tor tongue

/th/ M! Cd) /th/ M! (f) 3Dton,ue (AV) RDtongue ½RIV) /th/3D tonclue M!

tip posterior tungu: ttp

Cg) (h) (1) •y tongueM! (AV) /s/3D tongtie M[ (RLV) /s/3D tongue

groove

medtal groove

[•oO•terlor ngue tip anterior tongue

C1) •sh/ M! (j) /sh/ M! (k) /sh/ M! D tongue (AV) 3D tonclue(RLV) 3D tongue (PV)

blade blade po•t erlol tonoue

F'IG.I0. Three-dimensionaltongue shapes (subject MI) dunngthe production of unvoicedfricafives (AV anteriorview. RLV--right lateral view, PV-- posteriorview). (a)-(c) Ifl, (d)-(f)/0/, (g)-(i) Is/, and(j)-(I) I•l. hibitany concavity immediately behind the constriction. The the posteriorregion, the amountof which is subjectdepen- postalveolarfricatives of ore'subjects also show a moreiami- dent, resultingin smaller areasin that region. hal, than apical, articuhttion.As a result,postconstriction area functions for the alveolar fricatives are more abrupt 2. Nonstrident fricatives whencompared to thoseof the postalveolars.In theposterio• tongueregion, significant palatality effects result in relatively a. h•terdentalfricatives. The anteriortongue body for largerareas in thepostalveolar fricatives, with the degree of /0/and/6/was locatedbetween the teeth for all subjects.The palatalitybeing subject dependent. The alveolarfricatives, locationof the tonguetip (with respectto the teeth),how- on the otherhand, reveal slight raising of the tongueback in ever, showedsignificant interspeaker variability. The tongue

1337 d. Acoust.Soc. Am., Vol. 98, No. 3, September1995 Narayananet aL: MRI studyof fricativeconsonants 1337 (a) /f! PK /f/ PK 3D tongue(RLV) /f/ PK 3D tongue (AV) 3D tongue

posterior tongue

antertor tongue

/th/ PK (d) /th/ PK ½e) /th/ PK Cf) tongue(AV) 3Dtongue (RLV) 3Dtongue (PV)

jlosterlor toncue ttp t•p •

/s/ PK (g) /s/ PK (h) /s/ PK •D tongue(AV) •D tongue(RLV) •1) tongue(PV)

ß posterior ttp tongue

/sh/ PK (J) /sh/ PK (k) /sh/ PK 3D tongue CAV) 3D tonuue (RLV) 3D tongu, (PV)

posterior tongue

blade

FIG. 11. Threedimensional tongue shapes (subject PK) duringthe productionof unvoicedfricalives (AV--anterior view, RLV--right lateridview, PV-- posteriorview). (a)-(c) If/, (d)-(0/0/, (g)-(i) Is/, and (j)-(l)/.•/. blade of AK and M[ was betweenthe upperand lower teeth, thoseof AK and MI, on the other hand,had their tonguetips with the slightlyupward-pointing tongue tip well in front of restingmore or less on the lower incisors.For all subjects, the incisors(5-6 ram), a productionpattern prevalent in the minimum constrictionarea was formed by the upper Calilbrnian dialects.Illustrative mid-sagittalprofiles for/0/ teethand the anteriortongue body. In all cases,except PK's and/6/of subjectM[ are shownin Figs. 2(b) and 3(b), re- /c3/,the anteriortongue appeared either level or slightlyslop- spectively;3-D tongueshapes of/0/for subjectMI and PK ing (downwards)behind the tip, before it startedto show a are shownin Fig. 10(d)-(f) and Fig. 11(d)-(0, •espectively. significantraising of the tongue body behind the anterior SubjectsSC and PK, whosebackgrounds are differentfrom region,with the dotsurnat a higherlevel thanthe middle and

1338 J. Acoust. Soc. Am., Vol. 98, No. 3, September 1995 Narayanan et al.: MRI study of fricative consonants 1338 (a) (c)

2 3 4 5 6 3 4 5 6 distancefrom lips,cm distance from lips, cm (b) (d)

-10

2 3 4 5 6 2 • 4 5 6 distancefrom lips,cm distancefrom lips, cm

FIG.12. Medial tongue groove dimensions behind th •' constriction (unvoiced--solid, voiced4ashed). (a)-(b)/s,z/of MI andSC, respectively. (c)-(d) l•,3/ of MI andSC, respectively. Negative values imply g• ooving wh/ile positive values imply doming of thetongue with respect to its sides. posteriorparts. For example,see Figs. 2•b) and 3(b). For raising.The overallcross-sectional tongue shapes were either PK's/6/, however,the tongueraising began fight at the tip fiat or slightlyconcave. The cross-sectionalareas in thepala- and continuedto the dorsum,without exh biting any fiat or tal region appearedsemielliptical to semicircularin shape downward-slopinganterior tongue portion. It hasto be noted dependingon the subject'spalatal morphology. The cross- that the normalproductions of PK's/6/varied betweenstop- sectionalareas in the lip and alveolar regions may be ap- like and fricativelike articulations. The tol•en of PK's/6/in proximatedby elliptical shapes.Lingua-palatal contact was the micl-sagittalprofile indicated a stoplillebehavior while noticedbeginning around the middleregion of the tongue the tokencorresponding to the coronalscar used in themea- body.The tongueshape in the middleregion shows slight surementof A c clearlyshowed a finite,nonzero constriction concavitywhich graduallybecomes fiat in the directionto- area.It appearsthat the raising of thetongue body behind the ward the posteriorregion. Significantasymmetry in the constrictionmay be importantin directingthe air jet between tongueshapes of PK and SC was observedin the dorsal the tongueand the upperteeth. In addition,the shapeof the regionwith the left lateraltongue body at a higherlevel with anteriortongue body is downwardsloping only if the inter- respectto the fight side. dentallYicative is articulatedwith the tongte tip well in front b. Labiodentalfricatives. The tonguebody along the of the incisors(subjects MI, AK). mid-sagittalline for /f/ and /v/ was characterizedby a The locationand degree of the tongue•ody raisingwere "bunched"position for all subjects,with a raised tongue speaker'dependent. In addition,the tongueroot adw[ncement dorsum,lowered anterior and posteriorregions, and a down- observedin 16l possiblyinfluences the post erior tongue body ward pointing tip. The minimum constrictionarea was formed betweenthe upper teeth and the lower lips. Mid- TABLEIII. Volumeof sublingualcavity (mm3). sagittalprofiles of/f/and/v/of subjectMI are shownin Figs. 2(a) and 3(a); 3-D tongueshapes for/f/ of MI and PK are Subject shownin Figs. 10(a)-(c) and ll(a)-(c), respectively.The

Fricative AK PK MI SC cross-sectionalareas in the lip region may be approximated by ellipticalsections. The tonguetip appearedat 1.5 to 2 cm /.[/ --- 143.35 2 2,61 37.25 behindthe lip openingwith distinctsublingual cavities. The /3/ 175.42 208.02 2 !4.35 231.86 /f/ 73.79 445.37 5:•3.77 664.00 free anterior tongue body exhibited, in general, an asym- /v/ 49.35 498.95 41)2.09 22.65 metrical convex contour,which tums concaveonce lingua- palatalcontact is established(at about 1.5 cm from the lip

1339 d. Acoust.Soc. Am., Vol. 98, No. 3, Sept..,mber19•,5 Narayananet al.: MRI studyof fricativeconsonants 1339 FIG. 13. Axial profilesof pharyngealand laryngeal regions of the vocaltract during the productionof Is/(subjectMI). All sectionsare orienled with thetop co]]espon(lingto the front sideof the hum0nbody (teethand/or mandible can be •een on Ihe top while the vertebraecan be ,,eenon the bottom,in the first six sectionsshown). Distances are measuredwith respectto the glottis.(a) Sectionin the uvularregion (7.5 cm). (b) (e) Section,,in the upperpharsngeal region,at 6.6, 6.0, 5.4, and4.8 cm. respectively.ff}-(g) Sectionsin the lower-pharyngealregion, showing increasing presence of the epiglotti,,,taken at 3.6 and3.0 cm, respectively.(h) Sectionin the lower-pharyngealregion showing di,,tinct cpiglottic vallcculae (3 cm). (i) Sectionat the laryngealinlet ghowing the appearanceof the pitiformrecesses (1.8 cm). (j)-(k) Sectionsin the laryngealregion showing the ifidl'ormsinuses di,;tinct from the laryngealvestibule (I.2) andglottis (0.6). •espectively.(I) Sectionthrough the glottis.

TABLEIV. Pirifi•rmsinus volumes (L--left, R-•right)from 3-D reconstructions(mm•).

Subject

Fricative AK (L) AK I,R) PK (L) PK (R) MI (L) MI IR) SC (L) SC (R)

/•/ 1096.84 880.84 1043.74 727.30 2733.95 9649.55 954.27 863.49 /,3/ 323.95 459.31 1495.72 1525.82 4173.70 3974.20 1563.13 1774.55 /• 925.66 1060.61 1156.81 793.03 3304.78 3452.18 1554.73 1242.04 /• 858.78 1128.91 1202.34 1137.91 4787.09 4041.85 1243.35 1182.14 /0/ 741.54 809.79 1330.97 1349.57 2991.97 3025.87 995.87 761.33 /•/ 651.96 766.16 1162.59 844.72 3560.84 3366.95 1460.14 1593.77 /ff 380.75 659.59 1055.60 682.71 3235.(-,4 3144.09 1000.46 883.94 /v/ 468.19 661.67 1012.27 763.48 3598.63 3272.35 1952.41 18(18.46

1340 J. Acoust. Soc. Am., Vol. 98, No. 3, September 1995 Narayanan et al: MRI study of fricative consonants 1340 I'IG. 14.Cross-sectional profiles of theregion along the vocal tract bend during the production of lsl (snbjeclMI) obtainedb) [eft,matting sagiual ,,cans. To providedirectional orientation. the region toward the lower mandible ("bottom") is markedwith a "W' andthe region away t'rom the mandible ("top") is markedwith a "T" m thefigures. The sections shown are taken at approximately3-ram intervals, along the plane perpendicular to the vocal in,el midline. staUingatabout 7.8 cm from the lip opening. The uvular presence can be noticed in thesections appearing iu (c) (e). openingfor AK andPK, 2.4 cmfor SC and4.5 cmfor MO. an inwardlydrawn tongue body logetherwith the fortnation The cross-sectionalareas in the front region appeared of sublingualcavities. semielliptical/sendcircular.Among the eight fricatives,the As shownin Figs. 8(a), 8(e), 9(a), and 9(e), lhe labio- area function was the most variable for If/and/v/. There was dental area ftmctionsof MI appeardrastically different fi'om a trendto createa relativelylarge volume behind the lips by thoseof the othersubjects: The MR imagesshowed a tongue

1341 d. Acoust.Soc. Am., Vol. 98, No. 3, September1995 Narayananet al.: MRI studyof fricativeconsonants 1341 (a) (c) 6

7

(]

18 2 4 6 8 10 12 14 16 18 Diatancefi'om I pa, cm D[stenc•from I p,s,cm

(b} (cO

3

2

1

0 o! 0 2 4 6 8 10 12 14 16 18 .0 • 4 6 8 10 12 18 DlatalX• from IIp•, crrl

FIG. 15, Completearea functions for unvoiced(solid) and voiced (dashed) fricative pairs for subjectMI: (a)/f-v/, (b)/0-6/. (c)/s-z/,and (d) t,[-3/. body that was drowninwards and maintainedclose to the B. Back region baseof the oralcavity (away from the palate,with no ob- Analysisof the back regionis basedon axial scansand servablelingua-palatal contact until the vicinity of the velar mid-sagittalprofiles. Sample axial sectionsin the pharyngeal region),perhaps, reflecting the coarticulatory influence of the and laryngealregions are shownin Fig. 13 for/sl (subject preceding initial neutral vowel. For the other subjects, MI). Simplifiedback regionarea functionsof the unvoiced lingua-palatalcontact is establishedwith a raisedtongue at a and voicedfricatives are shownin Figs. 8 and 9. more anteriorpalatal location thereby resulting in a smaller The maximummeasured area in the back regionis 9.02 airwayopening. Moreover, the oral cavity sizeof MI wasthe cm2 (/3/ofSC). Consistent patterns in the area functions are largestamongst the four subjects. observedalthough interspeaker variabilities are more marked The influenceof the neutral vowel/•/, which preceded whencompared to thoseof the front region.Among the fri- the sustainedfricative, on the observedlabiodental tongue shapescannot be delineatedwith the available MR[ data. catives,the upperpharyngeal areas of/•/and/3/are foundto Variabilityin frontcavity volumes (space between the labio- be the largestwhile thoseof/f/and iv/, the smallest.For/•/ dentalconstriction and the outerlip opening)may contribute and /3/ the area increaseis most significantin the upper to the variabilityusually observed in the acousticspectra of pharyngealregion due to palatality[panel (h) in Figs.8-9]. these sounds. Fricatives/0/and/6/exhibita similarpharyngeal behavior as /•/and/3/. The raisingof the posteriortongue body in the 3. Sublingual cavities case of/s,z/ and /f,v/, on the other hand, contributes to Distinct sublingualcavities were visible in the coronal smaller area values in the uvular and upper pharyngealre- sectionsfor the fricatives/.[,3,f,v/ofall subjectsexcept for/.[/ gions. by AK; it may be that AK's /.[/ cavity is smallerthan the In general,the tongueroot tendedto be more advanced spatial samplingused. Illustrative examplesof sublingual in the caseof the voicedfricatives when compared to their cavitiesare shownin Fig. 5(a) and(d) (PK's/f/and/•/3-D unvoicedcounterparts [for example,compare mid-sagittal vocaltracts), and in the coronalsections of/•/by subjectMI profilesin Figs.2 and3]. Tongueroot advancement resulted in Fig. 7(c)-(d). Sublingualcavity volumes, obtained from in greaterareas in the mid- and lower-pharyngealregions. 3-D reconstructions,are given in Table 1II. Acoustic studies The amountof tongue-rootadvancement also influencedthe have shownthat the presenceof sublingualcavities contrib- epiglottic-vailecularvolume to some extent. In addition, utes to a relatively low frequencyspectral peak due to an tongue-rootadvancement was found to affect the amount of increasein the effective volume of the cavity between the posteriortongue body raising, particularlyfor the alveolar constrictionand teeth (Perkell et al., 1979; Shadle, 1991). fricatives. Supraglottal cavity enlargement due to active

1342 J. Acoust. Soc. Am., Vol. 98, No. 3, September 1995 Narayanan et aL: MRI study of fricativeconsonants 1342 TABLE V. Lengthmeasurements (in ram)for diffennt subjects.

SubjectAK SubjectPK SubjectMI SubjectSC

Fficative IVT lr• IBR lvo lvr lr• 1• lvo lvr lr• lun lvo lvr Ire 1•

[•l 170.1 97.1 84.8 18.3 159.4 84.5 84 II 177.8 100.8 91 20.2 176 99.6 100.8 16 13/ 164.2 96.1 83.4 15 159.5 84 82.8 12.9 172.8 101.3 94.7 20.6 175.2 101.6 t00 14.9 Isl 166.4 95.2 84.4 17.4 158.5 85.2 86.4 10.9 175.5 102.2 92.8 24.8 178.3 98.9 98 13.3 /• 164.5 94.7 84.4 19.2 156.1 84 86 11.3 ]72.1 102.2 93.8 21.5 172.2 101.6 96.9 14.1 IOI 171.1 98 82.5 19.7 157.• 83.6 84.8 11.7 177.1 103.1 91.9 25.7 173.9 99.6 98.4 14.8 161 169.1 95.1 82.9 15.4 154.• 84.4 84.8 10.5 177.3 101.7 93.8 25.8 176.7 102 101.6 12.5 /ff 173.4 91.9 83.9 13.6 159 83.6 85.2 9 177.6 101.3 92.9 17.8 178.3 97.3 100.4 11.8 Iv/ 167.3 95.6 85.8 12.6 156.• 86 82.4 10.9 178.9 102.7 95.6 18.7 177 98.8 101.6 9.8

average 168.26 95.46 84.01 16.4 157.3 84.41 84.55 11.02 176.14 101.91 93.31 21.88 175.95 99.93 99.71 13.4 s.d. 3.24 1.81 1.07 2.63 1.38 0.83 1.42 1.09 2.47 0.78 1.48 3.15 2.12 1.66 1.75 1.99 max. 173.4 98 85.8 19.7 159.5 86 86.4 12.!) 178.9 103.1 95.6 25.8 178.3 102 101.6 16 min. 164.2 91.9 82.5 12.6 154.• 83.6 82.4 9 172.1 100.8 91 17.8 172.2 97.3 96.9 9.8

tongue-rootadvancement has also been observedin voiced C. Region along the vocal tract bend stopconsonants (Westbury, 1983). The increasedpharyngeal Reformattedsagittal images were used to calculateareas volumesobserved in the case of voiced fiicatives and stops alongthe vocal tractbend. Sample cross sections along the may be a possiblemechanism for sustainingthe transglottal bend are'.shown in Fig. 14 for/s/ of subjectMI. Complete flow during voicing. Another possiblem.:chanism for sus- area functions for the fricatives of MI are shown in Fig. 15. tainingflow in the presenceof increasedsupraglottal pres- The arez, functions include the bend areas as well as the areas sure is a decreasein the complianceof tl'e vocal tract wall of the front and back regionswhich were calculatedfrom (McGowan et al., 1995). The relatively ;mall tongueroot raw coronaland axial scans,respectively. The "boundaries" advancementfor the labiodentalssupports such a possibility. of the bendwere markedbased on the end of the front region and the beginningof the backregion marked from the coro- nal and axial scans,respectively. The cross-sectionalshapes along the bendare influenced 1. Pirfform sinuses by the positionand shapeof the posteriortongue body, the velumand the uvula.In the velarregion, the shapeof the top Volumesof the right andleft pitiform ;inuses,calculated portionof the vocal tract resemblesthe sidesof a triangle; from 3-D reconstructions, are listed in T•.ble IV. The table the bottom portion is affected by the position and the illustratesa wide range of values acrossthe subjects.No concawdconvexshape of the posteriortongue body. For ex- clearpattern based on placeof articulationis observed.Simi- ample,the raisedposterior tongue body observed in the velar larly, no distinctivepattern is seenwith :espectto voicing vicinity for/s/yielded decreasedconcavity when compared althoughalmost two-third of the voicedfScative tokens re- to that in /.•/. For /.[/, increasedconcavity, and hence,in- veal larger volumesthan the unvoicedones. The volumesof creasedareas, in the velar regionwas influencedby the low- the left and right sinuses,in general, are not the same. The eringof the posteriortongue body. The presenceof the uvula observedasymmetry is speakerdependenl: PK, M[, and SC in the post-velar region decreasesthe effective cross- have, in general, larger left-side volume:s,while AK has sectional area. The uvular effect was noticeable for about larger right-sidevolumes. 1-1.5 cm alongthe bendin the postvelarregion. The cross- The acousticsignificance of the pitif,)rm sinusesis not sectionalshapes in the postvelarregion may be approxi- clearly understood.In addition,we lack informationon the mated by elliptical sections.A straightforwardsolution to tissueproperties and the degreeof mecl'ano-acousticcou- accountfor the uvula in the cross-sectionalshapes is, how- plingwith the laryngealvestibule (Baer ct al., 1991).Con- ever, not immediatelyevident; investigation of the acoustic stantvolume shunt-cavity models (Fant, 1960; Lin, 1990)for significanceof the uvular presencein the vocal tract may the pitiform sinusesresult in a zero in the acousticspectrum offer insightsin this regards.The areasalong the bend are above5 kHz accompaniedby a sharpersp•:ctral cutoff at that also influencedby the degreeof tongueroot advancement/ frequency(Fant, 1960) andlowering of th,:formant frequen- retraction.The advancedtongue root in the voicedfricatives cies,the extentof whichdepends on the Piaceof articulation resultedin relativelylarger areas along the bend,particularly (Lin, 1990). For fricatives,the acoustic•,ignificance of the in the postvelarregion. piriforrnsinuses greatly depends on the • egreeof coupling betweenthe cavitiesanterior and posteriorof the supraglottal D. Length measurements constriction.In the presenceof appreci•ble coupling, the variability observedin the measuredpirifi•rm sinusvolumes Lentgthsof the entire vocal tract (lvT), front region is likely to introducevariability in the correspondingacous- Urn), back region(/eel and verticallip opening(lvo) are tic spectra. given in Table V. As shown in the table, the vocal-tract

1343 d. Acoust. Soc. Am., Vol. 98, No. 3, Sep ember 19!;)5 Narayanan et aL: MRI study of fricativeconsonants 1343 lengths(lvr), acrossall fricatives,vary within 3 mm for when comparedto the unvoicedcases. In addition,the pos- eachspeaker. The smallestlip openingvalues occur for the terior tongueregion for the voiced fricativeswas raised labiodentals.Across speakers, no contrastingpattern is ob- slightlyhigher than it was for the voicelessfricatives sug- servedin the outerlip openingvalues (lvo) of/s,.•,z,3/,nor gestingthe influenceof tongue-rootadvancement observed do the data differ for the voiced and voicelessfricatives. 2 in the voicedcases. The amountof tongueroot advancement The cross sections for /•,3/, however, appeared more variedacross the differentplaces of articulationand subjects; rounded,when comparedto/s,z/, in the inner lip region in the variations,however, do not occur in a systematicway. the directiontoward the teeth[for example,compare Figs. Voiced labiodentalfricatives are the only fricativesthat do 6(a) and7(a)]. The roundingeffect contributes to increased, not showmarked tongue-root advancement. Enlarged supra- and a more uniform distributionof, areasin the cavity in glottalvolumes in phoneticallyvoiced sounds have been ob- front of the supraglottallingual constriction in /S,3/when servedin previousstudies (Perkell, 1969; Westbury, 1983). comparedto/s,z/. The increasedpharyngeal volumes for the voiced fricatives and stopsmay be a possiblemechanism for sustainingthe III. SUMMARY AND DISCUSSION transglottalflow during voicing.Another possible mecha- In thispaper, a morphologicalanalysis of the vocaltract nism for sustainingflow in the presenceof increasedsupra- geometryduring sustainedproduction of the Englishfrica- glottalpressure is a decreasein the complianceof the vocal tires/s,S,f,0,z,3,v,6/obtainedby MRI is presented.Coronal tract wall (McGowan etal., 1995). The relatively small scanswere usedto obtainarea functions of the frontregion, tongueroot advancementfor the labiodentalssupports such a to construct 3-D models of the entire vocal tract, and to possibility.The resultssuggest that the tongue-rootadvance- measurevolumes of the sublingualcavities. Axial scanswere ment mechanismis more likely in the case of the lingual used to measurearea functionsin the back region, and for fricatives,where the tongueis activelyinvolved in dictating volumemeasurements of the pitiform sinuses.Sagittal scans the aerodynamicfeatures such as constrictionformation. were mainly usedfor length measurements,and for measur- Asymmetriesin tongueshapes and lingua-palatalcon- ing areafunctions along the vocaltract bend for reformatted tactswere found to be subjectdependent and providecon- images. vergingevidence supporting the resultsof previousarticula- The front region tract shapesacross subjects were, in tory studies.These asymmetriesmostly occurred in the general,similar for the voicelessand voiced fricativesthat posteriortongue region and near the alveolar region; this was sharethe sameplace of articulation.Actual cavity volumes, especiallytrue for the stridents.The asymmetriesobserved however,were affectedby both the type of articulation,api- nearthe alveolarregion of the stridentsmay be attributedto cal versuslaminal for example,and the oral morphologyof aerodynamicconstraints in directingthe air jet toward/along the subject.Interspeaker variabilities in the backregion were an obstaclesuch as the teeth(Shadle, 1990). The asymme- foundto be greaterthan those in the front region.The con- triesin the posteriortongue region, also a subject-dependent strictionlocation x c values for the voiced and unvoiced behavior,may be influencedby the supineposition assumed soundswere more similar than the constrictionarea A c val- duringthe scanning.No gender-relateddifferences were de- ues. Strident fricafives, in general, were characterizedby tectedin the articulationpatterns. smallerA½ values than the nonstridents.It is possiblethat the The tongueshapes for Is/and/z/suggest that apicalor A½values derived from the MR imagesare largerthan the laminalarticulations for thesesounds are speakerdependent. actualA c values;this is primarilydue to the limitationsim- The alveolarfricatives are characterizedby postconstriction posedby insufficientresolution in the spatialsampling (3 tongueconcavity, the degreeof whichis foundto be subject mm). Theseeffects may particularlybecome significant in dependent.The apical/laminalnature of the anteriortongue the caseof labiodentalsbecause of the relativelysmall front is foundto influencethe tonguebody shapebehind the con- cavityand constriction lengths. In sucha case,this limitation striction:The anteriortongue body behindthe constriction may be overcomeby usingother approaches such as video revealsa relatively deepergrooving for the apical alveolars imagingof the lip region. when comparedto the laminal ones,contributing to a rela- In general,among the fricatives,the labiodentalsexhib- tively rapid increasein area valuesbehind the constriction. ited the most variability acrossspeakers. The variationsin This observationregarding intersubject variability in apical- the labiodentalsare not surprising:The tongue,which is the ity and laminality in the alveolars'production agrees with principalarticulator for the otherfricatives, is relativelyun- Dart's (1991) study.We disagree,however, with Dart's restricted for the labiodentals. In fact, the acoustical charac- speculationon deducingcross-sectional tongue shapes from teristicsof labiodentalsare greatly influencedby the vocalic lateral x-ray data for French and English apical alveolars environment(Harris, 1954). Suchcoarticulatory effects are [Dart (1991,p. 111)].Dart usedx-ray data for theFrench and expected to play a significantrole in the overall tongue EnglishIs/ of Bothorelet aL (1986) and Subtelnyet al. shapesassumed in a labiodentalarticulation. (1972), respectively.Dart posits that the postconstriction The voiced fricatives exhibited a tendency toward tongueshape is front and convexin the Frenchapical alveo- slightly larger area valuesin the region immediatelybehind lar while for the English one, the tongue shapeis more con- the constrictionwhen comparedto their voicelesscounter- cave. While we agree with Dart that the postconstrictionre- parts.In the mid- and low-pharyngealareas of the backre- gionfor Englishapicals is concave,her associationof French gion,the voicedfricatives showed significant tongue-root ad- apicalswith a convexshape may not be true. In our study, vancementresulting in relativelylarger pharyngeal volumes the midsagittalprofile of Isl for subjectMI [Fig. 2(c)], who

1344 J. Acoust.Soc. Am., Vol. 98, No. 3, September1995 Narayananet al.: MRI studyof fricativeconsonants 1344 also hasthe highestdegree of concavityamong our subjects, postalveolarfricatives. The flat/slightly convex tongue is very similarto the mid-sagittaltracing of the Frenchapical shapes in the correspondingregion are attributed to alveolar with the postulatedconvex cross-sectionalshape. downward-directed bilateral force tensors, similar to those We also found that the midsagittalprofile ,)f/s/for speaker positedfor the postalveolarfricatives. The posteriortongue PK, whichwas articulatedwith lessapical!•ty and concavity bodycross sections are concavealthough the degreeof con- thanthe othersubjects, was most similar to the x-ray tracing cavity is somewhatsmaller when comparedto that foundin of theEnglish apical alveolar. Hence, mid-s •gittal x-ray trac- /.[,3/,suggesting weaker forces causing concavity in the pos- ingsare not sufficientto indicatethe convexityor concavity teriorregion of the interdentalfricatives. of the cross-sectionaltongue shapes in fric:•tiveproduction. Basedon the tongue-shapeobservations and the acoustic The'.postalveolar fricatives do not exhi!)it any concavity theory of speech production,we can speculateon the in the constrictionregion. As a result, a more gradualin- articulatory-to-acoustictransformations for fricatives.In the creasein postconstrictionarea function is observedin the caseof the alveolarconsonants, the concavity("grooving") postalve.olar fricatives when comparedto the alveolars.In helpsin directingthe jet toward the incisors,which serveas the posteriortongue region, significanteffects of palatality an obstacleto the impingingjet and is believed to play an resultint relatively larger areasin the postalveolarfricatives, importan•Irole in turbulencegeneration. Turbulence genera- with the degreeof palatalitybeing subject diependent. On the tion for the postalveolarsis probablymanifested along the otherhand, the alveolarfricatives reveal slightraisintg of the upperwall (palatalroof) of thevocal tract near the supraglot- tongueback in the posteriorregion, the amountof which is tal constrictionand at the teeth, at a location, in general, again subject-dependent,resulting in smaller areas in that higher tl•tanthat of the alveolars.Palatality together with region.These observations agree with otherarticulatory stud- "doming"of the tonguefront, perhaps,play a key role in ies basedon palatographicdata (Ladefo•,.ed,1957; Hard- guidingtire airflow upward along the tonguesurface close to castle and Clark, 1981) and ultrasoundd.•ta (Stone et al., the roof of the mouthand towardthe upperteeth. The ante- 1992). rior tongue region for the interdentalsalso reveal tongue Differencesin tongueshapes of alveohrsversus postal- shapesthat facilitate in directingthe air jet throughthe con- veolarsmay be attributedto directionalitydifferences in the strictionbetween the upper teeth and the tongue.The con- resultantlingual force tensors.The tongue-palateinteraction caveposterior tongue shapes found in all the fricativesis to observe,dfrom the MR imagessuggests tha: the bilaterallin- facilitatethe airflow from the pharyngealto the buccalcav- gualforce tensors in the anteriorregion are directed.laterally ity. upwardfor alveolar stridentsas againstlai'erally downward The vocal tract and tongueshapes for the labiodentals for postalveolarstridents. Consequently, th,: resultanttensor exhibitedwide variabilities.Hence, it is not possibleto posit along the mid-sagittalplane is directeddownward for the generalizedaerodynamic characteristics for the labiodentals alveolar stridentsand upwardsfor postalveolarones. Two with the currently available data. Recall that the tongue concurrently occurring tongue-shaping mechanismsare shapesdescribed in this papercorrespond to sustainedcon- speculated: sonantsproduced in a VC contextwith a neutralvowel. It (1) Intrinsicmuscular action is neces,•.:aryin achieving remainsto be seenhow tongueshapes are affectedby coar- thegrooving (primary effect):3 presence/abe. enceof grooving ticulatoryeffects in differentvocalic environments,espe- is attributedto the presence/absenceof c.)ntractionof the cially for labiodentalfricatives. Electropalatography and anteriorgenioglossus and the verticalis. fasterMRI machinescould provide better insights in these (2) "Extrinsic" lingua-palatal br',tcing facilitates dynamic scenarios. grooving/doming(secondary effect): although the palate pro- This paperprovides a detailedstudy of the articulatory vides an anchoringboundary for the lateral lingual forces, configurationsof fricativeconsonants. The significanceof an the precisenature of the palatal contact,[erhaps, is not of accuratedescription of the 3-I) geometryin understanding importancein achievingthe grooving/nongroovingdistinc- and modelingfricative source mechanisms has been detailed tion. Lingua-palatalcontact details (force and area of con- by Shadle(1991). The availabilityof fairly accuratedimen- tact) dictatethe degreeof grooving/doming.Greater lateral sionsand cross-sectionalshapes may be exploitedin speci- lingua-palatal contact was found to ac:ompany deeper fying more realisticproduction models. For example,acous- groovingin the alveolarstridents. Furthermore, it was found tic couplingbetween the front and back cavitiescan be modeledaccurately. The acousticsignificance of the sublin- that apical alveolars showed greater lat,•.ralcontact and deeper postconstrictiongrooving when compared to the gual cavitiesand pitiform sinusescould be studiedin greater detail. With the availability of data from four subjects,one laminal alveolar fficatives. Further convergingevidence for can begin to investigateintersubject variabilities in the ar- theseobservations was providedby the re•..ultsof a concur- ticulatorydomain. In the future, detailedacoustic modeling rent electropalatographicstudy involving the same subjects basedon the datareported in this paperwill be presented. (Narayanan,1995). The anteriortongue shapes for the int{:rdentalfricatives ACKNOWLEDGMENTS is influencedby the lingual anchoringat •he front teeth. A tendencytoward slight concavingis probally due to a slight The cooperationand supportof the Cedars-Sinaimedi- contraction of the anterior genioglossus.The manner of cal imaginggroup, the UCLA Departmentof Radiology,and lingua-palatalcontact in the region of the raised dorsumis the UCLA PhoneticsLaboratory are gratefully acknowl- similar to that observed in the domed tongue region of the edged.Vqe are gratefulto our subjectsAK, MI, PK, and SC

1345 J. Acoust. Soc. Am., Vol. 98, No. 3, Sept{,mber 1995 Narayanan et aL: MRI study of fricativeconsonants 1345 for their cooperation.We thank Patricia Keating and Peter For the pipettes,errors in the volumesof the bulbswere Ladefogedfor many insightfuldiscussions. We thankMoran within 5%-8%. Errors in the maximum cross-sectional areas Sondhi and JuergenSchroeter for helpful discussions,par- of the bulbs obtained from either raw coronal or reformatted ticularly during the initial stagesof the project. We also sagittal data were both around3%. Errors in the minimum thankPierre Badin, Anders L/3fquist, Maureen Stone and an cross-sectionalareas (neck region) were anywherebetween anonymousreviewer for their helpfulcomments and sugges- 20%-100%; this demonstratesthe dramatic decreasein reli- tions. The assistanceof Zareh Movsessianis appreciated. ability as the measureddimensions approach image resolu- This work was supportedin part by NSF and by UCLA. tion. (c) Uniformand nonuniformbent tubes:Three tubes APPENDIX with varying degreesof bendingwere used.Tube A was a Calibration experiments rigid-walledimpalene tubing with a uniformarea of 74 mm2 and was bent in a form of a circular sector, 18 cm in diam- Extensivecalibration experiments were performedwith eter, subtendingan arc of about240 ø. Tube B was a standard the objectiveof identifyingsources of errorsin the various glass U-tube, with cross-sectionalareas varying between stagesof dataacquisition and processingand to quantifythe 130-150mm 2 withthe bend along its inner side following a performanceand reliabilityof MRI data for modelingpur- semicirculararc of diameter2.4 cm. Dependingupon the poses.All calibrationdata were collectedusing the same orientationof the object insidethe scanner,these tubes were MRI scannerand the same scanningprotocols used for the scannedin either the axial or coronalplane to obtain cross- human subjectsin our study.Furthermore, data-processing sectionalareas. Tube C was a rigid plastictubing with areas techniquesemployed were identicalto thoseemployed for inthe range of 25-34mm 2. The tube comprised two straight MRI data of human vocal tracts. "leg" regions,which were orthogonalto one another,con- Several tubeswith differentgeometries (uniform and nectedby a gentle bend. The tube was scannedin both the nonuniformareas) and shapes (straight and bent) were used axial and coronalplanes to obtaincross-sectional areas of the in theseexperiments. The dimensionsof the calibrationtubes "leg" regions.Since tube C was set inside the scannerin a were chosento cover the typicalrange of dimensionsin the positionsimilar to a supinehuman subject, the "front" leg of differentregions of the humanvocal tract during speech pro- the tube was scannedin the coronalplane while the "back" duction.All tubeswere hollow and filled with water to pro- leg was scannedin the axial plane. Tubes B and C were vide boundarycontracts, in the images,against the surround- treated as nonuniform tubes with actual areas evaluated at ing air medium; mineral oil provided a similar contrast. severalpoints along their lengths,particularly around the Presenceof air bubblesin the waterdid not posea serious bends.All threetubes were also scanned in the sagittalplane. problembecause their presencein the images,if any, was Averageerrors in areasobtained directly from coronalor easily distinguishablefrom the surroundingwater and, axial scans were in the order of 11%(_+5.6%), 22.7%( hence, could be corrected for during the data-processing +5.6%), 4.5%(+2.8%), and 3.1%(+2.2%) for tubesA, B, C stage.Actual areasof the variousobjects were obtainedfrom (coronal),and C (axial), respectively.The largevalues of standardspecification sheets from the supplier,whenever errors and their variabilities for the first two cases are not possible.In addition, measurementswith precisiondigital surprisingdue to the extreme bending; the raw cross- calipers(Mitutoyo Digimatic Calipers, least count 0.01 mm) sectionalareas do not correspondto the actualareas in the were made. plane perpendicularto the midline of the tubes.The errors for tube C, however,are not appreciabledue to the fact that Results the "leg" regions were approximatelyorthogonal to the (a) Uniform-straighttubes: Three straighttubes (TY- scanningplanes. Interestingly, a simplecoordinate transfor- GON 3603 series) with uniform cross-sectionalareas of.' mationusing a cosinefactor, of the raw areasfrom the first 1.27,0.71, and0.32 cm2 wereused. Errors between actual two casesto reflectthe areasin a plane perpendicularto the and MRI-derived area values, from raw axial scans,were, on midline of the tubes,using the anglesmeasured from the average,2%-4%. respective3-D reconstructions,resulted in decreasederrors (b) Nonuniform-straighttubes: For theseexperiments, a of 5.4%(+_3.4%)and 7.6%(+4.8%) for tubesA and B, re- standard100-ml flat-bottomedflask, and two pipettesof vol- spectively.The highererrors for tube B may be attributedto umes 2 and 10 ml were used. For the flask, the minimum inaccuraciesin the actualarea and angle measurementsdue diameter(neck region) was 1.18 cm, the maximumdiameter to the highly nonuniform nature of the U-tube along the was 5.8 cm, and the length of the neck was 7.7 cm. For the bends.Finally, errorsin the areascalculated from reformat- pipettes,the maximum cross-sectionalareas of the bulbs ted sagittaldata were 10%(+_5%),3%(_+2%), and 4.6%( were0.9 and2.92 cm 2, for thesmaller and larger pipette, _+2.2%)for tubesA, B, and C, respectively.The areaswere respectively,and the minimumcross-sectional areas (neck calculateddirectly by employing a grid overlay specifying region)for both was 0.02 cm 2. oblique cuts along planesperpendicular to the tube midline For the flask, errors in volumes calculated from 3-D re- over the entire length.The relatively large errorsfor tube A constructions,from either axial or sagittalscans, were within may be attributedto our assumptionthat the area is uniform 1.3%-1.6%. Average errors in length measurements,ob- while there may be slight deformitiesin the tube. tained from the reconstructed3-D object or mid-sagittalim- Note that segmentationerrors in the reformattedimages ages,were 3.4% for the length of the neck. were not accountedfor in the calibrationexperiments since

1346 J. Acoust. Soc. Am., Vol. 98, No. 3, September 1995 Narayanan et al.: MRI study of fricativeconsonants 1346 the experimentsused tubes with well-cefinedregion-of- Hoole, P., Ziegler,W., Hartmann,E., and Hardcastle,W. (19891."Parallel interestboundaries. Segmentation of vocal-tractimages, on electropalatographicand acousticmeasures of fricatives,"Clin. Linguis- the other hand,is more complicated(fo; example,in the tics Phoo.3(11, 59-69. Ladefoged,P. (19571. "Use of palatography,"J. Speech Hear. Dis. 22, 764-- regionnear the teeth)and will resultin highermeasurement 774. errorsin the reformattedimages than thos,'.from raw scans. Ladefoged,P., Anthony, J. F. K., andRiley, C. (19711."Direct measurement of the vocal tract," UCLA WorkingPapers in Phonetics19, 4-13. •Theterm apical refers to articulationsmade with tte tonguetip upwhile Ladefoged,R andMaddieson, I. (19861."Some of thesounds of theworld's laminal refers to articulationsmade with the tonguetip down relative to a language:s,"UCLA WorkingPapers in Phonetics64. Also, "Soundsof the raisedtongue blade. world'slanguages," (in preparation). 2Atthe ,onset of ourstudy, we attemptedto measure the amount of lip Lin, Q. (1!990)."Speech production theory and aniculatory speech synthe- protrusionfor the variousfricatives. The data, howecer,were iuconclusive sis," Ph.D. thesis,Royal Institute of' Technology,Stockholm. due to Ihe limited spatialsampling of the MRI m•chine(scans obtained Marchal, A., Farnetani,E., Hardcastle,W. J., and Butcher,A. (19881}. every3 ram). "Cross-languageEPG data on lingual-asymmetry,"J. Acoust. Soc. Am. 3Theintrinsic muscular action, rather than linguapal:.tal bracing, is posited Suppl. I 84, S127. to be theprimary effect,in light of evidencefor anteJiortongue grooving in McGowan,R., Koenig,L., and L6fquist,A. (19951. "Vocal tract aerody- the absenceof any significantlingua-palatal bracing such as that observed namicsin /aCa/utterances:Simulations," Speech Commun. 16, 67-88. for the low vowel/a/. Mermelstein,I• (19731."Articulatory model for the studyof speechproduc- tion," J. Acoust. Soc. Am. 53, 1070-1082. Badin,P. (19911."Fricative consonants: acoustic and x-ray measurements," Moore,C. A. (19921."The correspondenceof vocaltract images with vol- J. Phon. 19, 397-408. umesobtained from magneticresonance images," J. SpeechHear. Res. 35, Baer,T, Gore,J. C., Gracco,L. C., and Nye, R W (19911."Analysis of 1009-1023. vocaltract shape and dimensions using magnetic resonance imaging: Vow- Narayanan,S.S. (19951."Fricative consonants: An articulatory,acoustic, els," J. Acoust.Soc. Am. 90(21,799-828. Bothorel,A., Simon,P.; Wioland, F., andZerling, J. 1:'(19861. "Cin•radiog- and systemsstudy," Ph.D. thesis,UCLA, Dept. of ElectricalEngineering, raphie des voyelleset consonnesdu francais,"TJavaux de I'institut de Los Angeles,CA. Phon&iquede Strasbourg. Perkell,J. S. (19691.Physiology of SpeechProduction: Results and Impli- Catford, J. C. (19771. FundamentalProblems in Ptionetics(Indiana U.P., cationscg' a QuantitativeCineradiographic Study (MIT, Cambridge,MA). Bloomington,IN). Perkell,J. S., Boyce,S., andStevens, K. (19791."Articulatory and acoustic Dang, J., Honda, K., and Suzuki, H. (19931. "MI•,i measurementsand correlatesof the [s-•] distinction,"io SpeechCommunication Papers' Pre- acousticinvestigation of the nasaland paranasalcw•ities," J. Acoust.Soc. sentedar the 97th Meetingof theAcoustical Society of America,edited by Am. 94, 1765 (A). J. J. Wolf andD. K. Klatt (AcousticalSociety of America,New York). Dart, S. (19911."Articulatory and acoustic properties. of apicaland laminal Schnitzlein,H. N., andMurtagh, F. R. (19901.Imaging Anatomy of theHead articulations,"UCLA WorkingPapers in Phonetics79. and Spine(Urban & Schwarzenberg,Baltimore-Munich), 2nd ed. Fant, G. (19601. AcousticTheory of SpeechProt/uction (Mouton, The Shadle,C. H. (19901."Articulatory-acoustic relationships in fricativecon- Hague),. sonants,"in SpeechProduction and SpeechModelling, edited by W. Hard- Fletcher,S. G., McCutcheon,M., andWolf, M. (1989•."Palatometric speci- castleand A. Marchal(Kluwer Academic, New York), pp. 187-209. ficationof stop,affricate and sibilantsounds," J. SpeechHear. Res. 32, Shadle,C,. H. (19911. "The effect of' geometryon sourcemechanisms of 736 -748. fricative consonants,"J. Phon. 19, 409 424. Greenwe,od,A. R., Goodyear,C. C., and Martin, lh A. (19921."Measure- Stevens,K. N. (19711."Airflow and turbulencenoise for fricativeand stop ment of vocal tract shapesusing magnetic resonance imaging," IEE Proc. consonants: Static considerations," J. Acoust. Soc. Am. 50, 1180-1192. I (Commun.,Speech Vision) (UK), 139(61,553-5½,0. Hamlet, S., Bunnell,H., and Struntz,B. (19861. "Articulatoryasymme- Stone,M, Faber,A., Raphael,L. J., and Shawker,T H. (19921. "Cross- tries," 1. Acoust. Soc. Am. 79, 1165-1169. sectionaltongue shapes and linguopalatal contact patterns in [s], [S] and Hardcastle,W. J. and Clark, J. E. (19811."Articula:ory, aerodynamic and [1]," J. Phon.20(21,253-270. acousticproperties of lingual fricatives," Tech. Re•. 5, SpeechResearch Subtelny,J. D., Oya, N., and Subtelny,J. D. (19721."Cineradiographic Laboratorywork in progress,University of Readin:;,pp. 51--7'8. studyo1: sibilants," Folia Phon. 24, 30-50. Harris, K. S. (19541."Cues for the identificationof the fricativesof Ameri- Westbury,J. R. (19831."Enlargement of the supraglottalcavity and its re- can English," J. Acoust. Soc. Am. 26, 952. lation to stopconsonant voicing," J Acoust.Soc. Am. 73, 1332-1336.

1347 d. Acoust. Soc. Am., Vol. 98, No. 3, September 1995 Narayanan et al.: MRI study of fricativeconsonants 1347