Haskins Laboratories Status Report on Speech Research1990, SR-103/104, 111·124

Stimulus Order Effects in Vowel Discrimination*

Bruno H. Repp and Robert G. Crowdert

In same-different discrimination tasks employing isolated vowel sounds subjects oftengive significantly more "different" responses to one order of two stimuli than to the otherorder. Cowan and Morse (1986) proposed a neutralization hypothesis to account for suche~ect~: The first vowel in a pair is assumed to change its quality in memory in thedirection of the neutral vowel, schwa. We conducted three experiments using a variety ofvowels a~d. obtai~ed som~ initial supp~rt for the hypothesis, using a large stimulus set,but conflicting eVIdence Wlth smaller stimulus sets. Rather than becoming more similar toschwa, th~ first ,vowel in a pair. seems to drift tow~rds the interior of the stimulus rangeempl~yed 10 .a glVen test. W~ discuss several possIble explanations for this tendency andnote Its relation to presentation order effects obtained in other psychophysical paradigms.

INTRODUCTIONThe perception of vowels has long been of

central interest to speech researchers (see Nearey,1989, for a recent review). Isolated vowels, beingthe simplest instantiation of speech, provide acommon testing ground for theories of auditorypsychophysics and of speech perception. For theformer, they offer the challenge of complexity, forthe latter the advantage of simplicity. Eventhough these sounds are far removed from theconnected speech that speech perception theoriesultimately need to be concerned with, they are notentirely unnatural: Some isolated vowels occur asexclamations, fillers, or even as real words; all arereadily elicited as pronunciation "prototypes"from native speakers; and their presentationin a perceptual test usually engages thelisteners' mechanisms of phonetic categorizationunless stimulus uncertainty is minimized (seeMacmillan, Goldberg, & Braida, 1988).

To study the identification and discrimination ofisolated vowels, many researchers have usedstimuli drawn from an acoustic continuum

This research was supported by NICHD Grant HD01994 toHaskins Laboratories and by NSF Grant GB86 08344 toRobert G. Crowder. We are grateful to William Flack forassistance in running experiments and tabulating data, and toNelson Cowan for helpful comments.

111

spanning two or three phonetic categories.Although the perception of isolated vowels is notstrongly categorical (Le., discrimination perfor­mance within phonetic categories is well abovechance), there is usually a contribution of phoneticcategorization to discrimination performance (Le.,discrimination is most accurate in the categoryboundary regions). This was demonstrated, for ex­ample, by Pisoni (1973, 1975) in several discrimi­nation paradigms, including a "same-different"task. This simple task has been employed in anumber of later studies concerned with the role ofauditory memory in vowel discrimination.

In one of these studies, we (Repp, Healy, &Crowder, 1979) presented subjects with pairs ofstimuli from a 13-member synthetic liI-IrJ-/elcontinuum, obtained by stepwise linear inter­polation between the formant frequencies of Iiiand lei. Our most important finding was that asubstantial part of the discrimination performancecould be accounted for by contrast effects betweenthe members of stimulus pairs, as revealed in alabeling task, though it remained unclear whetherthese contrast effects were the cause or theconsequence of heightened discriminability. (Seealso Healy & Repp, 1982.) We also observed, inagreement with earlier results of Shigeno andFujisaki (1980), that retroactive contrast (theeffect of the second vowel in a pair on the labeling

112 .Repp and Crowder

of the first) was larger than proactive contrast (theconverse) when both vowels in a pair had to beclassified, presumably because the first vowel in apair had to be held longer in memory and thuswas less stable when the second vowel arrived. Inaddition to these contrast effects, however, thesubjects' responses revealed an unexpected effectof stimulus order: For pairs of nonidentical stimulifrom the liI-III region of the continuum, a higherpercentage of correct "different" responses wasobtained when the more Iii-like stimulus camesecond in a pair than when it came first.! At theklend of the continuum, however, this order effectwas absent or even reversed.

Earlier authors employing the same-differentparadigm had not paid any attention to such ordereffects and had simply combined the responses forthe two orders of each pair. The order effectsattracted our attention because they were quitelarge in some stimulus pairs and, apparently,different in nature from the contrast effects,whose occurrence among vowel stimuli has beenknown for a long time (e.g., Eimas, 1963;Thompson & Hollien, 1970). Whereas contrasteffects extended across the whole vowelcontinuum, order effects were most pronounced atthe Iii-end. More importantly, contrast effectsvirtually disappeared when the interstimulusinterval was lengthened and filled with anintervening irrelevant vowel, but stimulus ordereffects survived such interference. Thus theyseemed to be caused by a different mechanism. Atthe time, we did not pursue the explanation of thissecondary finding any further.

An article by Cowan and Morse (1986) drewrenewed attention to these order effects andsuggested that they reveal a hitherto unnoticedproperty of memory for vowels. These authorsused stimuli ranging from Iii to III, correspondingto one half of our earlier continuum. Again,discrimination accuracy was higher when themore Iii-like stimulus came second in a pair.However, the effect also interacted withinterstimulus interval: It grew larger as the(empty) interval was increased from 250 to 2000ms, due to a more rapid decrease indiscriminability for those stimulus pairs in whichthe more iii-like stimulus came first. On the basisof these results, Cowan and Morse proposed thatthe perceived quality of the first vowel in a pairchanges gradually while it is held in memory.2Specifically, they suggested that it changestowards a more neutral quality-that its internalrepresentation drifts toward the center of the

acoustic-phonetic vowel space (henceforth, theneutralization hypothesis). They furtherspeculated that this drift may be strongest forvowels such as Iii, which are near the periphery ofthe vowel space.

Thus, according to this hypothesis, an Iii-likevowel held in memory becomes more like I':JI andhence more similar to III (III being more centralthan Iii in the vowel space; see Figure 1 below),whereas an III-like vowel held in memory becomeseven more central and hence more dissimilar from1tJ. Therefore, an Iii-like vowel is difficult todiscriminate from a following, more III-like vowel,while the reverse order is easy to discriminate.The neutralization hypothesis also predicts areduction of the order effect at the Ie/-end of an liI­!J.I-Ie/ continuum, though not a reversal (asmistakenly claimed by Cowan and Morse), sincekI is somewhat more central than II! (Figure 1).

The neutralization hypothesis is interestingbecause it suggests a speech-specific memorymechanism. However, at this point its supportingevidence rests entirely on high front vowels.Cowan and Morse stressed the need for studies oforder effects in other regions of the vowel space.The purpose of the .present experiments was to fillthis gap, and thereby to assess the validity of theneutralization hypothesis.

I. EXPERIMENT 1

In this experiment we employed nine groups ofstimuli from all over the vowel space, arranged ina way that enabled us to make clear predictionsabout the direction and magnitude of order effects.

A. Methods

1. StimuliFigure 1 represents the stimuli schematically as

points in the two-dimensional acoustic spacedefined by the frequencies of the lowest twoformants (F1 and F2). Eight monophthongalvowels, li,I,e,re,a,;),u,uI, were selected from thePeterson and Barney (1952) norms for adult malespeakers of American English.3 In addition, theneutral vowel 1':11, which was not included in thestudy of Peterson and Barney and is less welldefined phonetically, was assumed to have F1 andF2 frequencies of 500 and 1500 Hz, respectively(Le., the resonance frequencies of a uniform tubehaving the length of the average male vocal tract;see Chiba & Kajiyama, 1941). These nineprototype vowels are located at the centers of thecrosses in Figure 1.

Stimulus Order Effects in Vowel Discrimination 113

2400 I

4~22200 3

1-

2000 4~2 I

4f2

3

18004

,~;3.....

N::t: 4 2-- 1600(IJ fIJ..

:3 I

14002

1200

'~:3

'*1000 3 4 32 I

8002~4 I ,~4

200 300 400 500 600 700 800FI (Hz)

Figure 1. Positions of the stimuli of Experiment 1 in FI-F2 space.

For each of the prototype vowels, four neighborsin vowel space were chosen as indicated by theendpoints of the cross arms in Figure 1. Eachcross was oriented so that one of its arms pointeddirectly to the neutral 1';)/ vowel. The fourneighbors of each prototype were numbered in aclockwise fashion starting with the vowel farthestfrom I':J! (cf. Figure 1). The neighbors of the I':J!prototype were arbitrarily determined by a crosswhose arms were parallel to the F1-F2coordinates, and they were numbered arbitrarily.The' fixed arm length of all crosses was chosen onthe basis of pilot observations, so as to make theneighbors fairly difficult to discriminate from theprototypes in a high-uncertainty task.

The formant frequencies of all the vowels arelisted in Table 1. The third formant of all stimuli

was fixed at 2440 Hz, except for Iii and itsneighbors, which had an F3 of 3010 Hz. Thestimuli were synthesized on the HaskinsLaboratories serial resonance software synthe­sizer with a duration of 250 ms and a linearlyfalling fundamental frequency contour (100-80Hz). An experimental tape was recordedcontaining four blocks (replications) of 117randomly ordered stimulus pairs each. The 117pairs resulted from each of the nine prototypesbeing paired with each of its four neighbors inboth temporal orders (72 pairs), and each vowelbeing paired with itself (45 pairs). The ratio of"different" to "same" pairs thus was 8:5. Theinterstimulus intervals were 500 ms within pairs,2 s between pairs, and 5 s after each group of 13pairs.

114 Repp and Crowder

Table 1. Formant frequencies (in Hz) of the stimuliused in Experiment 1. (P = prototype, N = neighbor)

Stimulus Fl F2

/iJ P 270 2290Nl 245 2380N2 315 2340N3 295 2200N4 225 2240

N P 390 1990N1 370 2080N2 435 2030N3 410 1900N4 345 1950

lEI P 530 1840N1 538 1940N2 580 1825N3 522 1740N4 480 1855

Izl p 660 1720N1 700 1780N2 690 1640N3 620 1660N4 630 1800

10/ P 730 1090N1 767 1020N2 695 1010N3 693 1160N4 765 1170

I:JI P 570 840N1 580 740N2 520 820N3 560 940N4 620 860

lui P 440 1020N1 425 920N2 390 1045N3 455 1120N4 490 995

lui P 300 870N1 272 785N2 258 925N3 328 955N4 342 815

lal P 500 1500N1 500 1600N2 550 1500N3 500 1400N4 450 1500

2. Subjects and ProcedureTwenty-four undergraduate subjects partici­

pated in this study for course credit. Each subjectlistened to the tape once, and then to the firstthree blocks again, so that seven blocks werepresented in all. The first block was consideredpractice and was not scored. Presentation wasover loudspeakers in a quiet room. The task wasto respond "same" or "different" to each pair by

circling the appropriate response on an answersheet.

B. Results and DiscussionThe predictions of the neutralization hypothesis

were as follows: Pairings of prototype (P) vowelswith neighbors N1 and N3, which lie on the axispointing towards I';J/, should yield large ordereffects of opposite sign. Nl pairs should show apositive order effect (defined as more correct"different" responses when P comes first thanwhen it comes second), whereas N3 pairs shouldshow a negative order effect, perhaps of smallerabsolute size because of their more central loca­tion in vowel space. N2 and N4 pairs, on the otherhand, should not show any significant ordereffects. For pairs involving the I';J/ prototype therewere no clear predictions, except that order effectsshould be small. Any large order effects obtainedin this region would suggest that the true I';J/prototype is located elsewhere.

The results are shown in Table 2. It is evident,first, that discrimination accuracy was not veryhigh but obviously above chance: Hit rates("different" responses to nonidentical pairs) wereuniformly higher than false-alarm rates("different" responses to identical pairs). This per­formance level was optimal for observing largeorder effects. Pairs involving Iii received markedlyfewer "different" responses than the rest; other­wise, performance did not vary substantiallyacross the vowel space.

The results of interest, the order effects, areshown at the bottom of the table. These effectswere computed by subtracting the percentage of"different" responses for pairs in which P camesecond from that for pairs in which P came first. Itcan be seen that a number of stimulus pairsshowed large (> ± 10%) order effects, but thatpairs involving I';J/ showed only small effects, aspredicted. In the following statistical analyses,these latter pairs were excluded because they didnot follow the general stimulus design.

Two separate ANOVAs were conducted, one onN1 and N3 pairs, and the other on N2 and N4pairs. Each had the factors Vowel (8), Neighbor(2), and Order (2). For the first analysis, the neu­tralization hypothesis predicted a significantNeighbor by Order interaction, due to positiveorder effects in pairs involving N1 and negativeorder effects in pairs involving N3. This interac­tion was indeed highly significant, F(l,23) = 34.48,p < .0001, although there was also a significanttriple interaction involving Vowel, F(7,161) = 4.19,p = .0003. Two-way follow-up analyses were there­fore conducted on N1 and N3 pairs separately.4

Stimulus Order Effects in Vowel Discrimination 115

Table 2. Average percentages of "different" responses for all stimulus pairs in Experiment 1. and order effects(prototype first minus prototype second).

Prototype vowel (P)

Pair iii lei Izl 10/ I~I /vi lui kJI

Identical pairs

pop 6.9 17.4 16.0 153 11.8 13.9 22.2 26.4 4.2

NI-Nl 11.1 19.4 13.2 9.7 11.1 13.2 18.8 12.5 16.0N2-N2 12.5 13.9 15.3 153 10.4 18.8 21.5 22.9 6.3N3-N3 9.0 16.7 18.8 19.4 13.2 18.1 19.4 19.4 6.3N4-N4 83 16.7 16.0 18.1 9.7 16.0 16.7 18.1 10.4

Nonidentical pairs: prototype first

P-Nl 22.2 39.6 70.1 63.2 64.6 72.2 54.9 59.7 43.8P-N2 37.5 63.9 52.8 38.2 46.5 56.3 71.5 66.0 35.4P-N3 15.3 34.0 54.9 61.8 50.7 56.9 48.6 35.4 35.4P-N4 43.1 74.3 49.3 49.3 54.9 60.4 64.6 40.3 34.0

Nonidentical pairs: prototype second

NI-P 10.4 27.8 46.5 35.4 43.8 53.5 61.8 27.1 36.8N2-P 41.0 70.1 49.3 39.6 54.9 48.6 46.5 32.6 36.8N3-P 20.1 53.5 69.4 65.3 54.9 54.9 49.3 563 28.5N4-P 23.6 68.1 54.9 52.8 38.2 55.6 66.7 62.5 31.9

Nonidentical pairs: order effects (difference scores)

Nl 11.8 11.8 23.6 27.8 20.8 18.7 -7.1 32.6 7.0N2 -3.5 -6.2 3.5 -1.4 -8.4 7.7 25.0 33.4 -1.4N3 4.8 -19.5 -14.5 -3.5 4.2 2.0 -0.7 -20.9 6.9N4 19.5 6.2 -5.6 -3.5 16.7 4.8 -2.1 -22.2 2.1

For Nl pairs, there was a highly significantpositive Order effect, F(I,23) = 54.78, p < .0001,but also a significant Vowel by Order interaction,F(7,16l) = 4.75,p = .0001. As can be seen in Table2, seven of the eight vowels showed large positiveorder effects, though their magnitude variedconsiderably; one vowel (lui), however, showed asmall negative effect. For N3 pairs, there was thepredicted negative Order effect, F(1,23) = 8.62,p =.0074, as well as a weak Vowel by Orderinteraction, F(7,16l) =2.73, p =.0106. Actually,only three vowels showed large negative effects;all other effects were of negligible size. Althoughthe neutralization hypothesis predicted smallerabsolute order effects in N3 than in Nl pairs

(which held for seven of the eight vowels), thislarge variability was unexpected.

For the analysis of the N2 and N4 pairs, theneutralization hypothesis predicted an absence oforder effects. There was, however, a significant(positive) main effect of Order, F(1,23) =8.46, p =.0079, and although the Neighbor by Order inter­action was not significant, there was a highly sig­nificant triple interaction, F(7,16l) = 7.28, p <.0001. A separate follow-up analysis of N2 pairsagain revealed a main effect of Order, F(1,23) =8.60, p = .0075, and a strong Vowel by Order in­teraction, F(7,16l) = 5.38, p < .0001. As can beseen in Table 2, two vowels (lui and lui) unexpect­edly showed large positive order effects; all other

116 Repp and Crowder

vowels showed small effects, as predicted. A sepa­rate analysis ofN4 pairs did not yield a significantmain effect of Order but again a significant Vowelby Order interaction, F(7,161) =4.53, P =.0001.Table 2 shows that two vowels (Iii, 101) yieldedsizeable positive order effects, and one (lui) a neg­ative effect, with the rest being negligible.

On the whole, these results confirm the mainpredictions of the neutralization hypothesis:Positive order effects for N1 pairs, negative effectsfor N3 pairs, and mostly negligible effects for N2and N4 pairs. There are a number of localdeviations from the predictions, however, whichare too large to be ignored. Some of these deviantresults could be explained by the ad hocassumption that back vowels changed in memorynot towards I-:J/ but towards a quality close to theN3 of I'JI (see Figure 1). This would predictpositive order effects for pairings of lui with its N1and N2, of luI with its N2, of I:JI with its N1, and of10/ with its N1 and N4, all of which were in factobtained; also, negative order effects for pairingsof lui with its N3 and N4, of lui with its N4, of I:JIwith its N3, and of10/ with its N2 and N3, whichwere clearly realized only in the case of lui butwere not strongly contradicted elsewhere; and noorder effects for pairings of luI with its N1 and N3,and for I:JI with its N2 and N4, which wasconfirmed (cf. Table 2). Front vowels and I~, onthe other hand, must still be assumed to decaytowards a quality near I':J/; otherwise, order effectswould have to be predicted for I-:J/ paired with itsN2 and N4, and for Ire! paired with its N4, none ofwhich were obtained. Thus, if different neutralpoints are assumed for front and back vowels, onlyone large order effect remains unaccounted for (Iiipaired with its N4).

Unfortunately, we have no independentjustification for assuming different neutral pointsfor front and back vowels. It was also surprisingthat pairs including Iii and its N3, which aresimilar to stimulus pairs that had yielded largenegative order effects in the studies of Repp et a1.

.(1979) and Cowan and Morse (1986), showed onlya negligible order effect here. This suggested to usthe possibility that the pattern of order effects isnot fixed but depends on the stimulus ensembleused in an experiment. Experiment 2 wasconducted to address this question.

II. EXPERIMENT 2In this study we reused four of the vowel sets of

Experiment 1, those grouped around the III, lei,lrel, and I':J/ prototypes. These are precisely thestimulus sets that yielded results supporting the

neutralization hypothesis. The critical set was lei,which was adjacent to each of the other three invowel space (see Figure 1). The stimulus pairsfrom that set were presented in three separatetests, each time intermixed with the stimuluspairs from one of the other three sets. If ordereffects were sensitive to stimulus context, theyshould follow significantly different patterns forthe same lei stimuli in the three different tests.The pattern of order effects for the context stimulishould also be changed in comparison toExperiment 1. Specifically, we suspected that thehypothetical neutral point might be in differentlocations in different contexts, perhaps closer tothe centroid of the stimulus ensemble used in aparticular test.

As an additional manipulation, we included twodifferent ISIs in our design. Cowan and Morse(1986) found that order effects in the flUII regionincreased with lSI, due to a more rapid decline indiscrimination performance for pairs in which themore Iii-like stimulus came first. They pointed outthat this provides important support for the neu­tralization hypothesis, whose main assumption isthat the memory representations of vowels changeover time. In the present study we intended toreplicate their finding by using two ISIs (200 msand 1 s) that straddled the lSI of 500 ms used inExperiment 1.

A. Methods

1. StimuliThe stimuli were the III, lei, Ire!, and I':J/ sets of

Experiment 1. Three separate test tapes wererecorded, the first containing lei and III stimuli,the second lei and lrel stimuli, and the third lei andkJI stimuli. Each tape contained six randomizedsequences of 52 stimulus pairs. These consisted of10 pairs of identical stimuli (each of the twoprototypes and each of the eight neighbors pairedwith itself once) and 16 pairs of nonidenticalstimuli (each of the two prototypes paired witheach of its four neighbors, in both orders), eachpresented with two ISIs: 200 ms and 1 s.

2. Subjects and ProcedureEighteen subjects from the same general pool

participated. Each subject listened to eachstimulus tape, in a balanced order. The procedurewas identical to that of Experiment 1.

B. Results and DiscussionThe results are presented in Table 3, with the

order effects at the bottom. Consider first theresults for the lei stimuli, shown in the last three

Stimulus Order Effects in VOUJel Discrimination 117

columns. A 4-way ANOVA was conducted on thesedata, with the factors Context (III, lrel, I'J!), Order(P first, second), lSI (short, long), and Neighbor (1,2, 3, 4). Several significant effects did not involveOrder: The main effect of lSI, F(l,17) = 8.15, p =.0109, reflected better discrimination performanceat the shorter lSI, which is not surprising; themain effect of Neighbor, F(3,51) =11.18,p =.0001,was due to much better performance for N3 pairs

than for the other pairs, a surprising finding; andthe Context by Neighbor interaction, F(6,102) =3.07, p = .0084, reflected a tendency towardsreduced discrimination performance for pairsinvolving the neighbor stimulus most dissimilar tothe context (N2 in the III context, N4 in the lrelcontext, N1 in the 1';)/ context), which suggests acontrastive influence of the context on theperceptual structure of the IE! category.

Table 3. Average percentages of "different" responses for all stimulus pairs in Experiment 2. and order effects(prototype first minus prototype second).

Prototype vowel (P)

Pair lSI /rI /lei lal Ie~ lelz lela

Identical pairs

pop 200 13.9 5.6 15.7 7.4 93 15.71000 5.6 93 93 11.1 13.0 13.9

NI-Nl 200 13.0 3.7 7.4 7.4 5.6 9.31000 13.0 5.6 16.7 7.4 10.2 5.6

N2-N2 200 7.4 3.7 93 5.6 7.4 13.91000 11.1 93 11.1 6.5 7.4 8.3

N3-N3 200 4.6 7.4 6.5 4.6 6.5 7.41000 83 5.6 93 5.6 83 93

N4-N4 200 3.7 10.2 18.5 7.4 4.6 13.91000 93 5.6 13.0 12.0 6.5 13.0

Nonidentical pairs: prototype first

P-Nl 200 62.0 713 74.1 63.9 63.0 63.91000 55.6 67.6 73.1 53.7 51.9 52.8

P-N2 200 60.2 52.8 48.1 51.9 51.9 61.11000 57.4 45.4 31.5 40.7 38.0 62.0

P-N3 200 463 72.2 68.5 722 75.0 74.11000 32.4 68.5 55.6 80.6 78.7 713

P-N4 200 75.9 66.7 74.1 59.3 62.0 7131000 75.0 64.8 65.7 58.3 57.4 63.9

Nonidentkal pairs: prototype second

NI-P 200 35.2 54.6 33.3 53.7 53.7 56.51000 27.8 39.8 19.4 38.9 44.4 32.4

N2-P 200 72.2 49.1 60.2 46.3 59.3 55.61000 74.1 24.1 46.3 48.1 66.7 47.2

N3-P 200 49.1 71.3 75.0 71.3 75.9 76.91000 38.9 62.0 72.2 42.6 61.1 583

N4-P 200 63.0 63.0 46.3 59.3 53.7 60.21000 41.7 45.4 333 50.9 33.3 56.5

Nonidentical pairs: order effects (difference scores)

Nl 200 26.9 16.7 40.7 10.2 93 7.41000 27.8 27.8 53.7 14.8 7.4 20.4

N2 200 -12.0 3.7 -12.0 5.6 -7.4 5.61000 -16.7 21.3 -14.8 -7.4 -28.7 14.8

N3 200 -2.8 0.9 -6.5 0.9 -0.9 -2.81000 -6.5 6.5 -16.7 38.0 17.6 13.0

N4 200 13.0 3.7 27.8 0.0 83 11.11000 333 19.4 32.4 7.4 24.1 7.4

118 Repp and Crowder

Effects involving Order were of primaryinterest: There was a highly significant maineffect of Order, F(1,17) =25.54, p =.0001, whichindicated a positive stimulus order effect overall.The interaction of Order and Neighbor fell short ofsignificance. Several other interactions weresignificant, however. One was between Order, lSI,and Neighbor, F(3,51) = 8.08, p = .0002, indicatingthat an Order by Neighbor interaction emerged atthe longer lSI. Another significant interactioninvolved Context, Order, and Neighbor, F(6,102) =4.05, p = .0011, indicating that the pattern oforder effects did change with test context. Thiswas particularly true at the longer lSI; thequadruple interaction was just significant,F(6,102) =2.24, P =.0449.

Since lei stimuli showed no really large ordereffects at the shorter lSI, the results at the longerlSI are of primary interest. Table 3 reveals thatN1 pairs showed a positive order effect, aspredicted by the neutralization hypothesis, thoughthe effects were smaller here than in Experiment1, especially in the lrel test context, despite thelonger lSI. N3 pairs, on the other hand, showedresults highly discrepant from those ofExperiment 1. Instead of the negative effectobtained there and predicted by the neutralizationhypothesis, these pairs exhibited positive ordereffects, with the effect in the IIi context conditionbeing more than twice as large as the effects inthe other two context conditions. Furthermore, N2and N4 pairs, which had not shown any ordereffects in Experiment 1 (as predicted by theneutralization hypothesis), showed some largeeffects here that depended on context: N2 pairsshowed a large negative effect in the lrel context,but a positive effect in the I':JI context. N4 pairsshowed a large positive effect in the /rei context.

We turn now to the results for the contextualstimuli, which are shown in the first threecolumns of Table 3. We dispense with statisticalanalyses here, which presumably would showmany complex interactions, and instead discussthe pattern of substantial order effects in relationto Experiment 1. Even more consistently thanwith the lei stimuli, order effects increased inabsolute magnitude at the longer lSI, but therewere also a number of large order effects at theshort lSI here. The results for the !II stimuli werenot unlike those in Experiment 1, though theeffects differed in size and, overall, were lesscompatible with the neutralization hypothesis: Alarge positive effect for N1 pairs, but only anegligible negative effect for N3 pairs; a moderatenegative effect for N2 pairs, and a large positive

effect for N4 pairs. The results for lrel stimuli atthe shorter lSI were quite similar to theExperiment 1 results, showing only a positiveeffect for N1 pairs. At the longer lSI, however,positive effects emerged for N2 and N4 pairs aswell. The most discrepant results were obtainedfor I'J/ stimuli, which in Experiment 1 had notexhibited any large order effects at all. In thisexperiment, all pairs showed order effects. Thosefor N1 and N4 pairs were extremely large andpositive, those for N2 and N3 pairs smaller andnegative.

These data provide strong indications thatchanges in the test environment affected thepattern of order effects. Overall, the results aremuch less favorable to the neutralizationhypothesis than the results of Experiment 1. Thereason for this may be that the "neutral point"that vowels in memory drift towards is specific toeach stimulus ensemble. If such a point exists, itshould be possible to infer its location from thepatterns of order effects for the two stimulus setsin a given context condition: Arms of a stimuluscross that are associated with negative ordereffects point outward, towards the "neutral point,"whereas arms associated with positive effectspoint inward. If the results for each of the twostimulus crosses are internally consistent in thatthey point in a particular direction, then theneutral point is located at the intersection ofthesetwo directions.

In the test containing IIi and lei stimuli, thepattern of order effects for the IIi stimuli (N2 andN3 negative, N1 and N4 positive) points towardslei; the pattern for the lei stimuli in that context isnot internally consistent (both N1 and N3positive) but is most compatible with a neutralpoint at the lei prototype. Thus these data suggesta neutral point in the vicinity of the lei prototype,which incidentally is consistent with the data ofRepp et a1. (1979) and of Cowan and Morse (1986).

In the test containing lrel and lei stimuli, the re­sults for the Ire! stimuli (all positive) point inwardstowards the lrel prototype, whereas the results forlei stimuli (N2 negative, N3 and N4 positive) point"north of' Ire!. Thus the neutral point here mayhave been located near the lrel prototype.

Finally, in the test containing I':JI and lei, theresults for the I':JI stimuli (N2 and N3 negative, N1and N4 positive) point quite clearly to the"southwest" (Le., away from lei), whereas the leiorder effects are all positive and therefore indicatea "neutral" point at the lei prototype. Thus, thedata from this test are contradictory and do notsuggest a unique neutral point.

Stimulus Order Effects in Vowel Discrimination 119

Although these results do not provide very clearsupport for stimulus range specific "neutralpoints," they are even less supportive of the range­independent neutralization hypothesis of Cowanand Morse (1986). They replicate only theirfinding that order effects, regardless of theirdirection, increase with interstimulus interval. Itis noteworthy, however, that the results forindividual vowel sets (i.e., for each cross in Figure1) are nearly always internally consistent andthus "point" in a particular direction. Oppositeneighbors never yielded large negative ordereffects. Thus we need to consider the possibilitythat each vowel set has its own individual neutralpoint, as it were. This seems unparsimonious, butthere is in fact a plausible rationale. Each vowelcategory has a best exemplar or prototype (see,e.g., Grieser & Kuhl, 1989) which mayor may notcoincide with the prototype suggested by the dataof Peterson and Barney (1952). Moreover, thelocation of a vowel prototype is likely to besensitive to stimulus context. The pattern of ordereffects may tell us something about the actuallocations of vowel prototypes and their shifts withchanges in context. Rather than becoming moreneutral in memory, vowels may become moreprototypical.

Although this seems a very reasonablehypothesis, there is a serious problem with it: Itmakes just the opposite predictions of the neutral­ization hypothesis. The data of Experiments 1 and2 are not at all compatible with the idea thatvowels in memory become assimilated toprototypes, because they would imply that theseprototypes are often located centrally in vowelspace, which does not make sense. It is stillpossible that prototypes playa role, but that thisrole is not assimilative but contrastive in nature.Before discussing this idea further, we report theresults of a third experiment, in which weemployed the methodology of our original study(Repp et aI., 1979)-viz., stimulus continuaspanning two vowel categories-to partiallyreplicate Experiment 2 and to conduct anotherspecific test of the predictions of the neutralizationhypothesis. Replication of the Experiment 2results seemed desirable because of their strikinginconsistencies with Experiment 1.5

III. EXPERIMENT 3In Experiment 3 we employed two vowel

continua, one ranging from lei to lrel, and the otherfrom lei to I-:J/. The first stimulus series continuedwhere the Ii/-III-Iel continuum used by Repp et al.(1979) ended. Since the two endpoint vowels, lei

and Ire, are about equally peripheral in the vowelspace (see Figure 1), the neutralization hypothesisof Cowan and Morse (1986) predicts nopronounced order effects along this continuum.Indeed, in Experiment 1 pairings of lei with its N2and of Ire! with its N4, which lie approximately onthis continuum (see Figure 1), yielded no ordereffects. In Experiment 2, however, large ordereffects (negative and positive, respectively)emerged for these very same pairs at the longerlSI. Since Experiment 3 used the same long lSIand an even more restricted stimulus context, itwas expected to replicate the results ofExperiment 2.

The second continuum ranged from lei to I-:J/.According to the neutralization hypothesis, strongnegative order effects should be obtained at the leiend of this continuum, but none at the I-:J/ end.These predictions were upheld in Experiment 1for pairings of lei with its N3 and of I-:J/ with itsN4, which lie almost exactly on the leI-I'JIcontinuum (see Figure 1). Again, the results ofExperiment 2 were contradictory: The very samepairs yielded a small and a large positive ordereffect, respectively. We wondered whetherExperiment 3 would replicate this curious pattern.

A. Methods1. Stimuli

The formant frequencies for the lei, lrel, and I-:J/prototypes, which served here as continuumendpoints, were the same as in Experiments 1 and2 (see Table 1). Five additional vowels wereinterpolated linearly between lei and lrel, andbetween lei and I-:J/, to obtain two 7-member vowelcontinua. Other stimulus characteristics were thesame as previously.

The stimuli of each continuum were recorded inpairs on separate experimental tapes. Theinterstimulus interval was 1 s within pairs and2.4 s between pairs. The pairs varied in the degreeof stimulus separation (measured in steps on thecontinuum). For each continuum, there were 37pairs: 7 identical pairs, 12 one-step pairs (6stimulus combinations, 2 orders), 10 two-steppairs, and 8 three-step pairs. Ten blocks of these37 pairs were recorded for each continuum, withdifferent random orders in each block.

2. Subjects and Procedure 0

Twenty-four undergraduate students served assubjects. Each participated in two sessions, ineach of which the same stimulus tapes werepresented. In one session, they were asked to givesame-different responses; in the other, they

120 Repp and Crowder

identified the second vowel in each pair. Theidentification data will not be reported here indetail.6 The order of these two conditions, and ofthe two stimulus sets within sessions, wascounterbalanced across subjects. The tapes wereplayed back monaurally in a quiet room using atape recorder and earphones of good quality. Halfthe subjects listened to the stimuli in their rightear, and the other half in their left ear; nosignificant ear differences were observed.

B. Results and Discussion1.1e/·leI continuum

Figure 2, left panel, shows the results for non­identical pairs from the Iel-/a~'/ continuum as afunction of location on the continuum, step size,and stimulus order. Predictably, the percentage of"different" responses increased as the step size in­creased. Scores also tended to be highest in themiddle of the continuum, which is in agreementwith the previously demonstrated tendency forisolated vowels to be perceived in a semi-categori.cal fashion (Repp et aI., 1979). Effects of stimulusorder are represented by the difference betweenthe closed and open symbols. There were largeorder effects for one- and especially two-step pairs;for three-step pairs, a ceiling effect probably pre-

vented order effects from emerging. At the lei endof the continuum discrimination performance wasmuch better when the more Ie/-like stimulus oc­curred second than when it occurred first; this dif­ference is particularly large for stimulus pair 1-3.In the middle of the continuum there were no pro­nounced order effects, but at the lrel end a reversaloccurred: Correct responses were more frequentwhen the lrel endpoint stimulus occurred second.

Analyses of variance were conducted on 1-stepand 2-step pairs separately, with the factorsstimulus pair and order. The stimulus pair byorder interaction, which reflects the change inmagnitude and direction of the order effect acrossthe continuum, was highly significant for 1-steppairs, F(5,1l0) = 6.79, p < .0001, and for 2-steppairs, F(4,88) = 39.09, p < .0001. In addition, therewas a main effect of stimulus pair for 1-step pairs,F(5,1l0) = 13.17, P < .0001, and for 2-step pairs,F(4,88) = 27.20, p < .0001, which reflects theaforementioned performance peak in the middle ofthe continuum, as well as the fact thatdiscrimination was better at the Ire! end than atthe lei end. For 2-step pairs, there was also a maineffect of order, F(1,22) =22.03, p =.0001, due tothe exceptionally large order effect for the 1-3stimulus pair.

CIJwenzoc.CIJwc:~zwc:wu.u.

~I­ZWCJc:wc.

80

60

40

20

a

•

123 4le.I

567l<el

I 2

lei3 4 5 6 7

lal

STIMULUS NUMBER

Figure 2. Results of Experiment 3: Percentages of "different" responses to pairs of nonidentical stimuli from /e1-1e1 (leftpanel) and /e1-1a1 (right panel) continua. Parameters are stimulus order (ascending versus descending direction) andseparation (one, two, or three steps).

Stimulus Order Effects in Vowel Discrimination 121

In the terminology of the earlier experiments,these results show negative order effects at bothcontinuum endpoints, which is inconsistent withthe neutralization hypothesis and with the resultsof Experiment 1. The results are more similar tothose of Experiment 2, where a large negativeeffect was obtained on the lei side, but a smallpositive effect on the lrel side. There, a "neutralpoint" near the lre/ prototype was suggested. Here,a neutral point is defined by the point on thecontinuum where no order effect is obtained (i.e.,the point at which the functions for the twostimulus orders in Figure 2 cross each other).That is somewhere between stimuli 4 and 5, whichis closer to Ire! than to lei. It is worth noting thatthe labeling data obtained from the same subjectsshowed the Iel-/rel category boundary to be in thesame location. These data, then, are reasonablyconsistent with Experiment 2; the differences maybe attributed to the changes in stimulus rangeand frequency (the prototype stimuli occurredmore often than other stimuli in the earlierexperiments, but not in Experiment 3).7

2. /e/·/Q/ continuumThe results for this continuum are displayed in

Figure 2, right panel. Again, there were stimulusorder effects that reversed direction along the con­tinuum. At the lei end, order effects emerged onlyfor I-step pairs and favored pairs in which themore lei-like stimulus came second. Effects at thekJI end were larger and in the opposite direction.The crossover point was closer to lei than to I';J/.

The statistical analyses showed the pattern ofresults to be very reliable. The stimulus pair byorder interaction was highly significant for I-steppairs, F(5,1l0) = 9.71, p < .0001, 2-step pairs,F(4,88) = 12.71, P < .0001, and even for 3-steppairs, F(3,66) =12.91, P < .0001. In addition, therewas a significant main effect of stimulus pair forI-step pairs, F(5,1l0) = 14.36, P < .0001, 2-steppairs, F(4,88) = 32.91, P < .0001, and 3-step pairs,F(3,66) = 11.96, P < .0001, due to betterdiscrimination performance in the center of thecontinuum, plus higher scores at the lei end thanat the I';J/ end. A main effect of order was obtainedfor 2-step pairs, F(1,22) =29.96, p < .0001, and for3-step pairs, F(1,22) = 7.96, p = < .0099, due to thelarge order effects at the kJI end of the continuum.

The Iel-/';J/ continuum thus yielded negativeorder effects at both endpoints, with the largereffects at the I';J/ end. The negative order effect atthe lei-end is consistent with the neutralizationhypothesis and with the data of Experiment 1.

However, the large negative order effect at the 1';;)/­end is in strong contradiction to both.Unfortunately, it also contradicts the findings ofExperiment 2 which showed a large positive effectfor I';J/ as well as a small positive effect for Ie/.These data, it will be recalled, were inconsistentin that they did not suggest a single neutral point;they remain mysterious. The present data suggesta neutral point somewhere between stimuli 3 and4 on the continuum (Le., closer to lei than to I';J/).Again we note that the category boundaryobtained in the labeling task fell there also.

IV. GENERAL DISCUSSIONThe purpose of the present series of experiments

was to test the generality of the neutralizationhypothesis proposed by Cowan and Morse (1986).We found many deviations from the predictions ofthis hypothesis, such as the large stimulus ordereffects in the vicinity of I';J/ obtained inExperiments 2 and 3. Only Experiment 1 yieldeddata that, on the whole, seemed to support thehypothesis. Although that experiment may seemto have been the strongest test because it includedthe largest variety of stimuli, it may actually havebeen the weakest: If stimulus order effects dependon the distribution of the stimuli in vowel space,then the most representative distribution has theneutral vowel at its center and therefore mayyield data that seem to support the neutralizationhypothesis. Only by using more limited stimulusdistributions can the range-specific nature of theorder effects be revealed.

The discrepancies among the results ofExperiments 1-3 provide ample evidence of suchrange-specific changes, though it must beadmitted that the pattern of effects obtainedcannot always be rationalized. On the whole,however, our data suggest that vowels change inmemory not necessarily towards the neutral vowelkJI, but towards a quality that lies within thestimulus range of a given experiment. What couldbe the reason for this?

It is well known, and the data from ourExperiment 3 confirm, that the perception of iso­lated vowels is weakly categorical: Discriminationtends to be best around the category boundaries.These discrimination peaks suggest that covertcategorization plays a role in the "same-different"task. Almost certainly, the first vowel in a stimu­lus pair is remembered in a dual code, one cate­gorical and the other continuous (Fujisaki &Kawashima, 1970; Pisoni, 1973, 1975). While, at ashort lSI, subjects can utilize the auditory stimu

122 Repp and Crowder

Ius trace for comparisons, at longer ISIs they mustrely increasingly on the category label assigned tothe first vowel in a pair. Since the size of order ef­fects increases with lSI, it seems likely that theseeffects are a phenomenon related to the covertcategorization of the vowel stimuli.S

We already noted, however, that simplephonetic classification does not predict the ordereffects that were in fact obtained. Phoneticcategorization amounts to an assimilation to theprototype, so that positive order effects would bepredicted at the ends of stimulus continua. Thenegative order effects obtained suggest thatvowels held in memory were assimilated towardssome standard(s) located between prototypes. Theonly "special" point in that ambiguous region isthe category boundary-the point of maximumuncertainty. Indeed, we found in Experiment 3that the "neutral point" suggested by the ordereffects coincided with the category boundary. Itseems, therefore, that the category boundarysomehow "attracts" vowels in memory; at thesame time, however, such a process cannot bereconciled with the idea of covert phoneticcategorization. Also, it is far from clear why theperceptually most stable vowels (the prototypes,and others near them) should exhibit the largestchanges in memory.

There is a way, however, of accounting for thesedata on the basis of phonetic categorization. Theapparent changes in the remembered quality ofthe first stimulus of a vowel pair may not occurautonomously during the silent interstimulus in­terval, but rather may be caused by the arrival ofthe second stimulus, which interacts with thememory trace of the first. This suggestion is sup­ported by three solid findings from earlier re­search. First, it is well known that successivelypresented vowels engage in contrastive interac­tions (Thompson & Hollien, 1970; Repp et a1.,1979), provided the interstimulus interval is nottoo short (Shigeno & Fujisaki, 1980) and they canbe perceived as belonging to different phoneticcategories (Shigeno, 1986). Contrast, of course, fa­cilitates discrimination. Second, vowels that areunambiguous representatives of a phonologicalcategory exert larger contrast effects than do moreambiguous vowels; it is the latter that are pushedaround in the context of less ambiguous neighbors(Crowder, 1982). Third, it has also been shownthat, when pairs of vowels are to be judged,retroactive contrast is larger than proactive con­trast (Repp et a1., 1979; Shigeno & Fujisaki, 1980),presumably because a memory trace is less stable

than a newly arrived stimulus. These three obser­vations together predict stimulus order effects ofthe kind found in Experiment 3 and earlier: Ateither endpoint of a vowel continuum, discrimina­tion should be easier when the more ambiguousvowel comes first and the less ambiguous vowelcomes second in a pair, because the retroactivecontrast effect in such a pair will be larger thanany proactive contrast effect obtained in the oPPOcsite arrangement. An increase in the order effectwith temporal separation between the stimuli isalso consistent with this explanation: As thememory trace of an initially stable vowel becomesweaker over time, its proactive contrast effect on afollowing unstable vowel will decrease, as shownby Crowder (1982). In fact, the labeling data inExperiment 3 revealed no significant proactivecontrast effects at all (footnote 6). On the otherhand, when an unstable vowel is followed by astable vowel, the retroactive contrast effect ex­erted by the latter on the former will stay thesame or even increase with temporal separation.Thus, according to this interpretation, order ef­fects do not occur because prototypical vowels be­come less stable in memory, but because unstablevowels shift away from following stable vowels.

Attractive as this explanation seems, there is aproblem with it. Repp et a1. (1979) found that aninterfering vowel sound eliminated retroactivecontrast effects but left stimulus order effects in­tact. Similarly, reanalysis of data from an unpub­lished vowel discrimination experiment by one ofus (RGC), in which interfering sounds were usedtogether with a long lSI, revealed large stimulusorder effects. These findings suggest that contrastand order effects are unrelated. The present ex­periments provide no additional information onthat point. Since no interfering sound was present,it is possible that retroactive contrast wasoperating. However, since retroactive effects areonly slightly larger than proactive effects (Repp eta1., 1979), the total absence of proactive contrasteffects in Experiment 3 suggests that retroactivecontrast, if present at all, was not very strong.Thus it seems that the retroactive contrast expla­nation may not be correct, after all.

A possible solution to this dilemma is suggestedby the psychophysical theory of Durlach andBraida (1969; Braida, Durlach, Lim, Berliner,Rabinowitz, & Purks, 1984), which has beenapplied to vowel resolution by Macmillan et a1.(1988). These authors distinguish between asensory trace and a more stable "context code."The context code is not limited to the phonetic

Stimulus Order Effects in Vowel Discrimination 123

labels listeners can apply; rather, it reflects theirmaximum labeling capacity. The context code thusis a subphonetic, quasi-categorical representation.It is also unstable and, as its name indicates,subject to influences of stimulus context. Althoughlisteners may assign a phonetic category to thefirst stimulus in a pair when it arrives, they mayuse its richer context code to compare it to afollowing stimulus. This context code may besubject to retroactive contrast, even when thephonetic category assigned to the first stimulus ina pair remains unaffected (Le., is not revised bythe listener).9

This interpretation receives independent sup­port from other psychophysical studies involvingnonspeech, even non-auditory, stimuli. Effects ofpresentation order in the method of constantstimuli have been noted since the earliest days ofpsychophysics (see Needham, 1934; Hellstrom,1985). A recent demonstration was provided byMasin and Fanton (1989) who used vertical linesin a visual length discrimination task. They con­cluded that subjects used a quasi-categorical code(i.e., a context code) to compare successive stimuli,and that "the categorical comparison isaccompanied by an additional inferential decisionprocess that uses only the category relative to thesecond stimulus because more weight is given tothat category, or because the category relative tothe first stimulus is momentarily forgotten" (p.485). They do not assume a change in the memorycode of the preceding stimulus through retroactivecontrast, but the same net effect is achieved by ahypothetical weighting process favoring the morerecent stimulus, which is really just anothermetaphor for memory degradation of the earlierstimulus. The argument is easily transferred tostimuli such as vowels, as long as it is assumedthat their context code is always in terms ofabstract labels that reflect the relative location ofstimuli in the range of all stimuli employed.Within the context code, continuum endpoints donot function as anchors (Macmillan et aI., 1988)and therefore can plausibly degrade in memory.

It seems, therefore, that stimulus order effectsin vowel discrimination represent an instance ofmore general presentation order effects in psy­chophysical judgment, not a phenomenon specificto the memory coding of speech sounds. The factthat there remain a number of unexplained irreg­ularities in our results may be attributed to theacoustic complexity of vowels, compared to thesimple unidimensional stimuli used in most psy­chophysical studies of "time order errors." In most

general terms, such time order errors are due to acontraction of the effective range for rememberedstimuli, a consequence of gradually substitutinggeneric information for specific information that islost (Hellstrom, 1985). The generic infonnation re­flects the recent stimulus history (Helson's, 1964,"adaptation level"). The neutralization hypothesisof Cowan and Morse (1986) may be seen as a spe­cific application of these general principles to cer­tain sets ofvowels whose adaptation level happensto be in the neutral region.

REFERENCESBraida, L D., Durlach, N. I., Lim, J. S., Berliner, J. E., Rabinowitz,

W. M., &: Purks, S. R. (1984). Intensity perception. XIII.Perceptual anchor model of context coding. JounuU of theAcoustiazl Soc~ty ofAmmaz, 76,722-731.

Chiba, T., &: Kajiyama, M. (1941). The rowel, Its 711Zture andstructure. Tokyo: Kaiseikan.

Cowan, N., and Morse, P. A. (1986). The use of auditory andphonetic memory in vowel discrimination. Journal of theAcoustiazl Soc~ty of Ammaz, 79, 500-507.

Crowder, R. G. (1982). Decay of auditory information in voweldiscrimination. Jounud of ExperimentAl Psychology: LeArningMemory, & Cognition, 8, 153-162.

Durlach, N. I., &: Braida, L. D. (1969). "Intensity perception. I.Preliminary theory of intensity resolution. JouT7llZl of theAcoustiazl Soc~ty ofAmmaz, 46,372-383.

Eimas, P. D. (1963). The relation between identification anddiscrimination along speech and non-speech continua.ungWlge and Speech, 6, 206-217.

Fujisaki, H., &: Kawashima, T. (1970). Some experiments onspeech perception and a model for the perceptual mechanics..AnnWll Report of the Engineering Research 11IStitute (University ofTokyo), 29, 207-214.

Grieser, D., &: Kuhl, P. K. (1989). Categorization of speech byinfants: Support for speech-sound prototypes. DeuelopmentalPsychology, 25, 577-588.

Healy, A. F., &: Repp, B. H. (1982). Context independence andphonetic mediation in categorical perception. Journal ofExperimentAL Psychology: Human Perception and Performance, 8,68-80.

Heison, H. (1964). Adaptation-Ieuel theory: An experimental andsystematic approach to behatrior. (Harper &: Row, New York).

Hellstrom, A. (1985). The time-order error and its relatives:Mirrors of cognitive processes in comparing. PsychologicalBulletin, 97, 35-61.

Kewley-Port, D., &: Atal, B. S. (1989). Perceptual differencesbetween vowels located in a limited phonetic space. JouT7llZI ofthe Acoustiazl Soc~tyof America, 85, 1726-1740.

Macmillan, N. A., Goldberg, R. F., &: Braida, L. D. (1988).Resolution for speech sounds: Basic sensitivity and contextmemory on vowel and consonant continua. JounuU of theAcoustiazl Soc~ty of Ameriaz, 84,1262-1280.

Masin, S. c., &: Fanton, V. (1989). An explanation for thepresentation-order effect in the method of constant stimuli.Perception & Psychophysics, 46,483-486.

Nearey, T. M. (1989). Static, dynamic, and relational properties invowel perception. JounuU of the Acoustical Society of Ameriaz, 85,2088-2113.

Needham, J. G. (1934). The lime error in comparison judgments.Psychological Bulletin, 31, 229-243.

124 Repp and Crowder

Peterson, G. E., &: Barney, H. L. (1952). Control methods used in astudy of the vowels. JounuU of tM Acoustical Socit:tyof America,24,175-184.

Pisani, D. B. (1973). Auditory and phonetic memory codes in thediscrimination of consonants and vowels. Perception &Psychophysics, 13, 253-260.

Pisoni, D. B. (1975). Auditory short-term memory and vowelperception. Memory & Cognition, 3, 7-18.

Repp, B. H., Healy, B. H., &: Crowder, R. G. (1979). Categories andcontext in the perception of isolated steady-state vowels.JournAl of ExperimentAl Psychology: HumAn Perception AndPer{rmrwra,5,129-145.

Shigeno, S. (1986). The auditory Tau and Kappa effects for speechand nonspeech stimuli. Perception & Psychophysics, 40, 9-19.

Shigeno, 5., &: Fujisaki, H. (1980). Context effects in phonetic andnon-phonetic vowel judgments. AnnuAl Bulletin RILP (Tokyo)14,217-224.

Thompson, C. L., &: Hollien, H. (1970). Some contextual effects onthe perception of synthetic vowels. lAngUAge And Speech, 13, 1­13.

FOOTNOTES·JounuU of the Acoustical Society of Ameriaz, 88, 2080-2090 (1990).tAlso Department of Psychology, Yale UniVersity.lWe note, with some embarrassment, that the effect is incorrectly

described in Experiment 2 of Repp et al. (1979, p. 138). We areconfident that this is a mistake in the text, and that the dataconformed to the description given here and in Experiment 1 ofRepp et al. (1979, p. 134).

2They generously credit us (Repp et al., 1979) with this idea,though we did not state it explicitly.

3We follow common practice in referring to the Peterson-Barneyfor the synthesis of isolated vowels, even though these dataderive from vowels produced in a Ih_dl context

4We are aware of the dangers of conducting multiple analyses onthe same data without adjusting the p levels. However, thesefollow-up analyses serve the sole purpose of clarifying complex

interactions, and the significance levels are generally so high asto make adjustments superfluous.

SWe are taking the liberty of describing Experiment 3 in theseterms for expository reasons. Actually, Experiment 3 wasconducted before Experiments 1 and 2.

6The purpose of this condition was to assess proactive contrasteffects. Somewhat surprisingly (see, e.g., Crowder, 1982), nosignificant effects were found. Note that the occurrence ofretroactive contrast effects is not precluded by these findings(see General Discussion).

7Another demonstration that order effects change with stimulusrange is obtained from a comparison with the old data of Reppet al. (1979). They showed a small negative order effect at the1£1 end of the 1i/-/l/-I£1 continuum. In those pairs, however,the stimulus paired with the 1£1 prototype was more hi-like,while in the present pairs It was more lei-like; hence thepresent negative order effect at the I £1 end is contrary to theeffect obtained previously.

8Accordingly, stimulus order effects should be smaller in tasksthat force subjects to rely more on the stimulus trace. Kewley­Port and Atal (1989) conducted experiments with four sets ofvowel stimuli <Ii/-hi, I£I-/e/, lui-lui, and le/-/a/-hl>arranged in prototype-neighbor configurations, but the taskrequired numerical dissimilarity judgments for the two vowelsin a pair. We re-analyzed their raw data (kindly provided byDiane Kewley-Port) and found stimulus order effects to be smalland follo~ga consistent pattern in only one of the sets, lu/­lui. That pattern agreed with that obtained in our Experiment 1,suggesting that lui-like vowels drifted towards lui.

9Macmillan et al. also noted that points of perceptual stabilityserve as "anchors" for the context code. Although anchors areoften located at the ends of stimulus continua, Macmillan et al.deduced from their vowel discrimination data that boundArystimuli served as anchors on their 11/-/1/-1£1 continuum. Thissurprising (and somewhat tentative) conclusion is in agreementwith the order effects obtained in our experiments, though Itleads back to the neutralization metaphor and should perhaps beregarded with caution.