Journal of Memory and Language 45, 412–432 (2001) doi:10.1006/jmla.2000.2783, available online at http://www.academicpress.com on

Segmental and Suprasegmental Mismatch in Lexical Access

Salvador Soto-Faraco and Núria Sebastián-Gallés

Departament de Psicologia Bàsica, Universitat de Barcelona, Spain

and

Anne Cutler

Max-Planck-Institute for Psycholinguistics, Nijmegen, The Netherlands

Four cross-modal priming experiments in Spanish addressed the role of suprasegmental and segmental infor- mation in the activation of spoken words. Listeners heard neutral sentences ending with word fragments (e.g., princi-) and made lexical decisions on letter strings presented at fragment offset. Responses were compared for fragment primes that fully matched the spoken form of the initial portion of target words, versus primes that mismatched in a single element (stress pattern; one vowel; one consonant), versus control primes. Fully match- ing primes always facilitated lexical decision responses, in comparison to the control condition, while mis- matching primes always produced inhibition. The respective strength of the contribution of stress, vowel, and consonant (one feature mismatch or more) information did not differ statistically. The results support a model of spoken-word recognition involving automatic activation of word forms and competition between activated words, in which the activation process is sensitive to all acoustic information relevant to the language’s phonol- ogy. © 2001 Academic Press Key Words: stress; vowels; consonants; features; lexical access.

Subjective experience tells us that recognizing even hundreds of thousands, some of them dif- words in a spoken utterance is an effortless fering only slightly from the word actually process. Words are simply heard, and then effi- uttered. There is ample evidence that speech ciently accessed in our lexical memory. How- input activates a number of different words with ever, the subjective impression may be highly which it is temporarily or partially consistent misleading, for the process is not at all trivial. (Connine, Blasko, & Wang, 1994; Connine, Understanding a spoken word involves comput- Titone, Deelman, & Blasko, 1997; Marslen-Wil- ing from a continuous and highly variable signal son, 1990; Zwitserlood, 1989). This can occur the information cueing one word among tens or even when the possible candidate words are em- bedded within longer words (Gow & Gordon, This work was partially supported by DGES research 1995; Shillcock, 1990) or when they span a word grant PB97-0977 of the Spanish Ministerio de Educación y Ciencia, and Catalan Government research grant SGR99- boundary in the intended utterance (Tabossi, Bu- 00083. The first author was supported by a fellowship of the rani, & Scott, 1995). The lexical candidates that Spanish Ministerio de Educación y Ciencia, FP95 are fully or in part consistent with the input (i.e., 77605578. We thank Mariette Koster for help in planning that are activated) compete with one another the experiments, Jaume Ramos and Ana Capdevila for help (Goldinger, Luce, & Pisoni, 1989; McQueen, with the preparation of the materials, Carolina Cortes for recording the sentences, and Xavier Mayoral for technical Norris, & Cutler, 1994). For instance, McQueen support; we further thank two anonymous referees for con- et al. (1994) found that detection of a word in a structive suggestions. nonsense string was more difficult (slower and Supplementary data for this article are available on less accurate) when this string was itself the be- IDEAL (http://www.idealibrary.com). ginning of a real word than when it was not. The Address correspondence and reprint requests to Salvador Soto-Faraco, Department of Psychology, University of word MESS, for example, was less efficiently British Columbia, 2136 West Mall, Vancouver, BC, V6T detected when it was embedded in the nonsense 1Z4, Canada. E-mail: [email protected]. string domess than when it was embedded in

0749-596X/01 $35.00 412 Copyright © 2001 by Academic Press All rights of reproduction in any form reserved. MISMATCH IN LEXICAL ACCESS 413 nemess; the former forms the onset of a real Eng- or lexical pitch accent). This may reflect merely lish word (i.e., domestic, and its morphological the fact that all current models are based on data relatives) that, being consistent with the input from experiments in English, and there is evi- signal, inhibits the target word mess. However, dence that in English lexical stress does not play the other nonsense string, nemess, activates no a strong role in word form activation (Cutler, competing word, so that mess receives no inhibi- 1986). However, there are now many experi- tion and thus can be recognized sooner. The out- mental demonstrations, from other languages, come of this multiple activation and competition of the importance of suprasegmental informa- process is, in most cases, that the input can be tion in lexical access; English may be atypical in unambiguously mapped onto a sequence of indi- this respect (see Cutler, Dahan, & Van Donse- vidual lexical items. laar, 1997, for a review). Although the processes of lexical activation Some recent behavioral studies have also and retrieval have been intensively studied in examined the effects of segmental mismatch recent years, most current models of spoken- between input and lexical representation. In word recognition are more concerned with cor- these studies the input is usually a nonword that rect capture of the phenomena of multiple con- in some way mismatches a real word. Connine current activation and interword competition et al. (1997) found that phoneme-monitoring than with the detailed simulation of the input responses were faster in nonwords that closely level to lexical access (see, however, Elman & resembled real words than in nonwords that McClelland, 1986, for an attempt to implement were unlike any real word. Boelte (1997), Cut- a detailed phonemic description of the input). ler, Sebastian-Gallés, Soler-Vilageliu, and Van Thus in the majority of models computational Ooijen (2000), and Van Ooijen (1996) investi- simulations begin with an input coded as a gated how listeners reconstructed real words string of phonetic segments, e.g., in Shortlist when they were given nonwords that differed (Norris, 1994) and in the Neighborhood Activa- from real words by a single phoneme (i.e., the tion Model (Luce, Pisoni, & Goldinger 1990). string kebra). When two solutions were avail- In TRACE (McClelland & Elman, 1986), able, one requiring substitution of a vowel (i.e., phoneme nodes are activated by input from a cobra) and the other substitution of a consonant bank of feature detectors. The first version of the (i.e., zebra), the vowel-substitution solution cohort model (Marslen-Wilson & Welsh, 1978) was consistently easier (faster and more accu- assumed an input consisting of a string of dis- rate) to reconstruct (Cutler et al., 2000; Van crete phonemes, but later work within the Ooijen, 1996). In a lexical decision task, re- framework of this model (e.g., Marslen-Wilson sponses are faster if an immediately preceding & Warren, 1994) incorporated the knowledge item overlaps with the current target in all but a that information about phoneme identity can single phoneme than when the preceding item overlap in time. In none of the empirical pro- is unrelated to the target (Cutler, Van Ooijen, & grams associated with these models, however, Norris, 1999; Radeau, Morais, & Segui, 1995), has explicit attention been given to testing these although the facilitation is transient (Cutler et input assumptions. In contrast, it is often ac- al., 1999). cepted that the assumptions are merely place- These studies certainly support the claim that holders for more detailed and faithful imple- lexical activation is sensitive to all available in- mentations to be undertaken at a future time formation; a partial match can produce partial (see, e.g., McClelland & Elman, 1986: p. 14; activation. But they do not fully elucidate the Norris, 1994: p. 208). Furthermore, no current role of match and mismatch between input model augments the segmental information (the and lexical representations in normal spoken- phonemes) in the input with suprasegmental in- language processing. In order to arrive at the formation (that is, information in the pitch con- correct sequence of spoken words in the input, tour, amplitude contour, or timing which varies listeners must be able to reject all words that with lexical identity: lexical stress, lexical tone, are not part of the correct sequence, even 414 SOTO-FARACO, SEBASTIÁN-GALLÉS, AND CUTLER though there may be very many words that dif- ation of spoken-word recognition. In the studies fer only minimally from those in the correct se- reported here listeners received input matching quence, and even though, as so much empirical the initial part of one word and mismatching an- evidence attests, those minimally differing other, and we measured the resultant facilitation words may be temporarily activated and thus versus inhibition (compared with a control con- may engage in the competition process. dition) for the matched and the mismatched The listener’s interest is best served if a mis- word respectively. This experimental technique, match between input and lexical representations sometimes called fragment priming, is a variant has an immediate and substantial effect on the of the cross-modal priming paradigm (see Zwit- pattern of activations. The original cohort model serlood, 1996, for a review) by which word- (Marslen-Wilson & Welsh, 1978) indeed pro- initial fragments reliably activate the representa- posed such a drastic effect: a single-phoneme tions of words with which they are compatible mismatch would produce total exclusion of mis- (Zwitserlood, 1989). Participants made lexical matched words from the set of words under con- decisions on visually presented words, and we sideration. However, this claim cannot hold; the compared response times and accuracy to a many demonstrations of activation despite par- given visual target as a function of the type of tial mismatch have established that. In models auditory prime that preceded it—a control word involving interword competition, such as fragment, a matching word fragment, or a mini- TRACE (McClelland & Elman, 1986), or the mally mismatching word fragment. We assume Shortlist model (Norris, 1994), the effect of a that compatible primes will activate the repre- mismatch on the activation of a word depends sentation of the target (as in Zwitserlood, 1989), on the competition process. As a function of the leading to facilitation in comparison to a control number of other words currently activated, and prime. The contribution of a given type of infor- the degree to which they in turn are mation to the lexical activation process will then matched/mismatched by the input, a word that be indicated by the extent to which activation is has suffered a mismatch may still play an effec- reduced—i.e., responses are inhibited relative to tive role in the ongoing competition. In Short- the control—by minimal mismatches in such in- list, for instance, a mismatch in the input will formation. We chose this task for various rea- penalize the mismatching word (i.e., reduce its sons. First, this study was only possible using activation); but a word penalized in this manner fragments of words (for reasons explained will not necessarily be excluded from the short- below) and the described technique is known to list of activated words. However, the competi- be sensitive to priming effects produced by par- tion process will then result in inhibition, tial input. Second, lexical decision in the prim- spreading from the more highly activated word ing paradigm taps into automatic word activa- that received the greater support from the input tion processes underlying speech perception to the less favored mismatched word; that is, the (rather than, for instance, explicitly directing continued presence of the mismatched word in observers’ attention to a given phonological as- the competition set will be observable from the pect of speech such as individual phonemes). fact that its recognition will be more difficult Using this method, we compared several (i.e., it is inhibited). types of information. First, and most crucially, This inhibition effect should be visible not we addressed the effect of suprasegmental mis- only when such minimal word pairs are pre- match and that of segmental mismatch. Al- sented as wholes but also with fragments of though evidence from lexical tone languages words that minimally mismatch part of another (e.g., Fox & Unkefer, 1985), lexical pitch- word (i.e., “solu” should produce activation of accent languages (e.g., Cutler & Otake, 1999), solution and inhibition of solicit, solicitor, and and lexical stress languages (e.g., A. Cutler & so on). Indeed, mismatch as a means of distin- W. Van Donselaar, submitted manuscript) attests guishing between such fragments is presumably that suprasegmental information can constrain the most common occurrence in the normal situ- lexical access, no study has undertaken a direct MISMATCH IN LEXICAL ACCESS 415 comparison of suprasegmental versus segmental versus SAbana (“sheet”; uppercase denotes the mismatch. Second, we compared, within seg- primary-stressed syllable). It might as a conse- mental mismatch effects, vowel mismatch quence be argued that cues to stress have little to against consonant mismatch. Again, evidence offer the Spanish listener, since it will not often from tasks involving the reconstruction of words be necessary to refer to stress pattern to distin- from nonwords has shown asymmetric effects guish one word-form from a segmentally identi- of vowel and consonant mismatch (Cutler et al., cal alternative. However, Spanish presents many 2000; Van Ooijen, 1996), but no study has com- minimal pairs of related words. Thus for exam- pared vowel versus consonant mismatch with ple, CAso is, besides a noun meaning “case”, the present activation paradigm. And finally, also the first person singular present tense of the within consonantal mismatch we compared the verb “to marry”, while caSO is the third person effect of a mismatch on one versus many phono- singular past tense of the same verb. Changes in logical features. This too is a dimension that has the stress pattern of stems frequently indicate been shown to affect responses given nonword differences in grammatical function (for in- input (e.g., Connine et al., 1997) but has not stance, noun versus verb) or, as in the above ex- been examined systematically using primes that ample, different forms of verbal inflectional activate competing words. morphology. Thus Spanish listeners can use In order to make our comparison between suprasegmental cues to distinguish between dif- suprasegmental and segmental mismatch as ferent forms of the same stem. Furthermore, close as possible, we needed to conduct our ex- with continuous-speech input that may activate periments in a language in which minimal pairs multiple candidate words overlapping with one of words can be distinguished either segmentally another, lexical stress information could help to or suprasegmentally, and in which segmental and distinguish between otherwise identical frag- suprasegmental structure are not necessarily in- ments of speech, and thus to provide mismatch terdependent. In many languages, there is such information that can help rule out potential com- interdependence, in that suprasegmental effects petitors. That is, listeners may quite regularly co-vary with segmental effects—for instance, have recourse to stress information in order to vowels are reduced in unstressed syllables in distinguish precisely such fragments as those English. Such confounds do not occur in Spanish. used in the present study. Experiment 1 ad- All polysyllabic Spanish words have one syllable dresses the constraints exercised by supraseg- marked for primary stress; this primary stress can mental information in lexical activation. occur in any syllabic position, and stressed and unstressed syllables do not differ in their vocalic EXPERIMENT 1 makeup. There is no vowel reduction; all vowels In the present experiment, as in the following are full, whether they occur in stressed or in un- ones, word onset fragments (the two first sylla- stressed syllables (see Navarro-Tomás, 1968, for bles) were presented at the end of carrier sen- detailed descriptions). Note that Castilian Span- tences as auditory primes. The visual target ap- ish has only five vowels, and that these are widely peared immediately at prime offset, and lexical separated in phonological space (Skelton, 1969; decision times to the target were measured from Stockwell & Bowen, 1965.) Thus minimal pairs this point. Reaction times and accuracy were of words, or of word fragments, can differ in evaluated as a function of prime type. In the stress but can be identical in segmental struc- match condition prime fragments consisted of ture—the necessary prerequisite for the compari- the onset of the target word pronounced aloud son we wished to undertake. Accordingly we (PRINci- for the target PRINcipe). In the mis- chose to conduct this study in Spanish. match condition prime fragments came from the As in other languages with variable lexical onset of a word with identical segmental infor- stress, it is not easy to find pairs of unrelated mation in the two first syllables but different words in Spanish that are only distinguished stress pattern (prinCI- from the word prinCIpio, suprasegmentally, such as saBAna (“savannah”) for the target PRINcipe). Finally, a control con- 416 SOTO-FARACO, SEBASTIÁN-GALLÉS, AND CUTLER dition was included to obtain a baseline measure for this experiment. All were native Spanish of the target word activation (the fragment mos speakers, with normal or corrected vision and from the word mosQUIto, for the target no reported problems in hearing. They received PRINcipe). One-syllable fragment primes were course credits for their participation. To balance chosen as unrelated controls in order to avoid the number of participants in each list, we ex- obvious lexical stress information being con- cluded data from three participants selected at veyed by the prime in this condition. Since lexi- random (leaving a total of 10 in each version of cal stress might be assessed (in Spanish) by the experiment). Therefore the analyses were comparison of pitch, amplitude, and duration based on data from 40 participants. between neighboring syllables, a two-syllable Materials. Twenty-four experimental word fragment prime would always be associated to a pairs were selected according to the following stress pattern. However, the use of prime frag- criteria: they were three or four syllables long, ments of different syllabic length than the ex- they were segmentally identical up to the onset perimental primes (one vs two syllables) may of the third syllable, they were not semantically seem a problem for interpreting inhibition or and/or morphologically related, and they dif- facilitation effects obtained. As a check on the fered in stress pattern. An example pair is neutrality of the control primes here used, it is PRIN.ci.pe–prin.CI.pio (“prince”–“beginning”; expected that they always yield performances syllable boundaries are marked with dots). The not better than those for the matching primes, selected words were 46 nouns and 2 adjectives; and not worse than those for the mismatching no items were compounds, and in no item did ones. This condition was met in every one of the the initial two syllables form a word. We experiments here presented. matched the frequency of the word pairs as far If the two types of experimental primes as possible (we used the LEXESP database that (match and mismatch) prove equally effective in contains frequency counts on a 5,020,930-word facilitating lexical decisions to the target, we body of written material; Sebastián-Gallés, may conclude that lexical stress is not used in Martí, Cuetos, & Carreiras, 2000); the mean log the lexical activation process, and that only seg- frequency of the target words1 was 3.61, SD ϭ mental information—the same in both types of 1.5; mean absolute difference in log frequency prime—is relevant. However, if positive priming between members of the same pair was 1.77, effects appear for the match condition but not SD ϭ 1.15. In addition, for each pair, one con- for the mismatch condition, we may conclude trol noun, without phonemic overlap in the first that, in Spanish at least, suprasegmental infor- syllable with its associated pair, was chosen (in mation is used in lexical activation, producing a the above example the control word was situation in which only one of the possible can- mos.QUI.to). A complete list of items is given in didate words completely matches the input Appendix A. while the other mismatches. The mismatching From each of these 24-word triplets, six condition may in this case be equivalent to the prime–target pairings resulted: Every possible control condition; this would suggest that the prime word in the triplet (the two experimental mismatching prime effectively ruled the target words plus the control) was paired with each of out of the initial activation set completely. Alter- the two target words (the two experimental natively, the mismatching condition may show words) in the same triplet. Table 1 shows the six inhibition relative to the control condition; this prime–target pairings for the chosen example. would be an indication of initial activation fol- For each of these 144 pairings (24 ϫ 6), one lowed by suppression as a result of competition sentence containing the prime word at the end from more favored candidate words. was constructed. All sentences were semanti- Method 1 These, and the remaining frequency averages presented Participants. Forty-three undergraduate stu- in the paper, were assessed using only the items included in dents at the University of Barcelona volunteered the analyses. MISMATCH IN LEXICAL ACCESS 417

TABLE 1 Example of the Six Prime–Target Pairings Constructed from One of the Selected Triplets and the Associated Carrier Sentences Prime type Sentence Prime Target Match Todos habían oido hablar del PRIN.ci.(pe) PRINCIPE (Everybody had heard about the. . .)

Él vió un libro sobre el prin.CI.(pio) PRINCIPIO (He saw a book about. . .) Mismatch Todos habían oido hablar del prin.CI.(pio) PRINCIPE

Él vió un libro sobre el PRIN.ci.(pe) PRINCIPIO Nadie supo leer la palabra mos.(QUI.to) PRINCIPE Control (Nobody knew how to read the word. . .)

Todos los diarios hablaron del Mos.(QUI.to) PRINCIPIO (All newspapers talked about the. . .) Note. Each word triplet (experimental pair plus control word) yielded six prime–target pairings (two in each condition). Carrier sentences were balanced for each condition between the experimental pairings. Parentheses in the prime column indicate the cutoff part of the prime word. cally neutral and not syntactically biasing to- way. For each of the 96 sentences containing the ward either word of the pair. As some target experimental primes, the cut was made at the pairs had different gender, their carrier sen- offset of the second syllable of the last word tences were constructed in such a way that they (i.e., the prime). The 48 sentences containing were not syntactically biased toward one of the the control words were cut immediately follow- genders (i.e., using constructions like “El niño ing the first syllable of the prime. no sabía escribir la palabra. . .”; “The kid did In addition, 48 filler prime/target pairs were not know how to write the word. . .”). Sen- constructed and recorded, in 48 new neutral sen- tences corresponding to the same target in the tences (i.e., constructed in the same way as the experimental pairs were identical except for the ones described for the experimental and control last word2 (the prime). Sentences containing primes), with primes also placed at the end of control primes were different from those con- the sentences. For 32 of the sentences, the cutoff taining the experimental words. point was at the end of the second syllable of the We recorded the 144 sentences from a female prime. Of these 32 sentences, 16 were associ- native Spanish speaker using a digital audiotape. ated with a word target (YES response) that had The speaker, who was unaware of the goal of the no segmental overlap with the prime word frag- experiment, was instructed simply to read aloud ment, and 16 were associated with a nonword and clearly from a list in which the full sen- target (NO response) that overlapped phonolog- tences had been randomly mixed. Tape record- ically with the prime fragment. The remaining ings were digitized at 16 kHz and each sentence 16 filler sentences had the prime cutoff at the was saved in an individual audio file. The cutoff end of the first syllable, and were associated to a points in the prime words were established target nonword that had no phonological rela- using a sound editor (Cool Edit v. 1.52, from tion to the prime fragment. Syntrillium Software Corp.) in the following Four different experimental lists were con- structed from the materials as described. For 2 Therefore, even if one of the two targets were to be each sentence sextet, each one of the four exper- slightly favored by the sentence meaning, the effects would cancel out when averaged, given that across the whole ex- imental prime–target pairings was assigned to a periment, each of the four possible prime/target combina- different list. In addition, from the same sextet, tions was associated with each of the two sentences. each control sentence was associated to the two 418 SOTO-FARACO, SEBASTIÁN-GALLÉS, AND CUTLER lists that did not already contain an experimental TABLE 2 sentence with the target word of that control Distribution of Trial Types for Each List in Experiment 1 sentence. In the example in Table 1, each one of Prime type Cutoff point S.O. Response Number the match and mismatch sentences would be as- signed to a different list. Thus the control sen- Match 2 YES YES 12 Mismatch 2 YES YES 12 tence associated to the target PRINCIPE would Control 1 NO YES 24 be assigned to the two lists containing the target 2 YES NO 16 PRINCIPIO (either in the match or the mis- Fillers 1 NO NO 16 match condition); the control associated with 2 NO YES 16 the target PRINCIPIO would be assigned to the Note. The cutoff point refers to the number of syllables two lists containing the sentences associated presented as a prime (counted from the onset of the word). with the target PRINCIPE. In this way each list The column labeled S.O. (segmental overlap) indicates contained 24 experimental and 24 control sen- whether the prime fragment and its target onset overlapped tences in which none of the targets or the carrier segmentally. Response refers to the target lexical status (re- sponse “yes” to a word target, and response “no” to a non- sentences were repeated. From the 24 experi- word target). The last column displays the number of trials mental sentences, there were 12 matching of each type in every list. prime–target pairings, and 12 mismatching prime–target pairings3 (see Table 2). All 48 fillers were further added to each list for a total row of asterisks while at the same moment the of 96 trials per list. Finally, the sentences in each sentence began to be presented over the head- of these four lists were pseudo-randomly or- phones. At sentence offset (i.e., at the cutoff dered (the only restriction being that there could point of the incomplete prime word), the target not be more than three YES or NO responses in string, printed in capital letters, replaced the as- a row). terisks on the screen. The target presentation Procedure. Each participant was seated in onset started the computer’s clock; timing was front of a computer screen in an individual stopped by the button press, or after a timeout booth, wearing Sennheiser HD440II head- of 2 s. There was an interval of 1 s phones. A two-button response box (with labels before the beginning of the next trial. YES and NO) connected to the computer’s The entire experiment was under control of a parallel port was used to collect responses. Par- HP-Vectra VL2 4/66 personal computer running ticipants were instructed to respond (as fast as the EXPE programming language (Pallier, possible) with their decision as to whether the Dupoux, & Jeannin, 1997). The auditory sen- letter string displayed at the end of the auditory tences were played to headphones via a Proau- sentence was a word or not, while also trying to dio Spectrum 16 soundcard at a comfortable avoid errors. The trial sequence started with a sound pressure level. The response box was row of X’s presented at the center of the screen placed near each participant’s preferred hand. for 1 s. After that, the X’s were replaced by a Participants listened to two blocks of 96 trials corresponding to two different lists among the 3 Although no targets were repeated within any of these four described in the materials section. While lists, both phonologically related targets in the study were the order of trials was randomized within each included in the same list (i.e., PRINcipe and prinCIpio). block, the order of blocks was counterbalanced This was done to gain statistial power (24 observations per cell in the experimental conditions rather than 12). The op- (targets were repeated once between blocks but tion of including more items in the experiment was not vi- never within a block, as pointed out in the de- able because it was not possible to find sufficient additional scription of the lists). Block order was counter- Spanish words conforming to the criteria for the experiment. balanced with the type of trial, so if a target had Nevertheless, it was always the case that one of the targets been preceded by an experimental prime (match was preceded by the control prime and the other target by an experimental prime (either a match or a mismatch), and that or mismatch) in one block, it would be preceded the two trials containing targets of the same experimental by the control prime in the other block. Conse- pair were presented in different halves of the experiment. quently, the carrier sentences associated with a MISMATCH IN LEXICAL ACCESS 419 target were always different across the two were very low (2.05%, SE ϭ 0.02, overall), the blocks. The total duration of the experiment was same one-way ANOVAs were conducted for the about 25 min. accuracy data. The Prime Type factor did not reach significance in either analysis (F1[2, 78] Results ϭ 2.1, p ϭ .118; F2[2, 86] ϭ 1.3, p ϭ .263). Two items (together with their experimental Because there was a repetition of every target pairs) were excluded from the analyses because across two equivalent halves of the experiment, of a high error rate (15% or more, overall), leav- we also analyzed the results for each half of the ing a total of 44 items. None of the participants experiment alone. The pattern of results was made more than 10% errors on average in the equivalent to that obtained in the experiment as experimental conditions. Erroneous responses a whole (see Appendix B for the detailed analy- were excluded from the RT analyses. In this and ses). the following experiments, responses that had RTs faster than 250 ms or were timed out (over Discussion 2 s) were also excluded. Experiment 1 shows that two-syllable audi- Separate ANOVAs were conducted on RTs tory primes matching the target word’s on- and on accuracy, across participants and across set both segmentally and suprasegmentally items (see averages in Table 3). In each the speeded up response times as compared with main within-participants/items factor was unrelated control primes. However, prime frag- Prime Type (Match, Mismatch, Control). The ments that were segmentally identical but mis- RT analyses showed a significant main effect of matching in suprasegmental structure slowed Prime Type (F1[2, 78] ϭ 21.1, p Ͻ .001; F2[2, down responses. This result clearly shows that 86] ϭ 14.4, p Ͻ .001). Planned contrasts re- Spanish listeners use lexical stress information vealed that RTs were faster for matching primes in lexical access. The two lexical items segmen- than for control primes (F1[1, 39] ϭ 20.9, p Ͻ tally compatible with the prime fragment en- .001; F2[1, 43] ϭ 12.6, p Ͻ .005), indicating a tered the initial candidate set; information from significant facilitation effect (ϩ34 ms). The the stress pattern gave an advantage to one of planned contrasts between the mismatching these candidates, thereby biasing the competi- prime condition and the control condition were tion process against the mismatching lexical also significant (although only marginal by par- item. The presence of inhibition shows that the ticipants; F1[1, 39] ϭ 3.7; p ϭ .06; F2[1, 43] ϭ target mismatching the prime was indeed ini- 4.9, p Ͻ .05), indicating that targets preceded tially activated, but then adversely affected by by a word onset mismatching on the position of competition from the matching target. the primary stress, but otherwise identical, We next ask whether evidence for initial acti- slowed lexical decisions to the target (Ϫ18 ms). vation and subsequent inhibition can also be ob- Although error percentages in this experiment

TABLE 3 Average Reaction Times and Error Percentage (ϩSE in Parentheses) for Each Condition in Experiments 1 through 3 Prime type Experiment 1 Experiment 2 Experiment 3 Experiment 3, 1-feature mismatch several-features mismatch Match RT (ms) 615 (17.6) 617 (16.0) 619 (12.8) 623 (11.6) Errors (%) 1.6 (0.04) 1.2 (0.03) 1.4 (0.04) 0.3 (0.01) Mismatch RT (ms) 667 (18.8) 716 (21.2) 750 (14.7) 744 (12.9) Errors (%) 2.8 (0.05) 4.1 (0.03) 5.0 (0.09) 4.6 (0.08) Control RT (ms) 649 (16.3) 695 (16.9) 710 (14.5) 705 (12.3) Errors (%) 1.7 (0.03) 1.7 (0.05) 2.4 (0.04) 1.6 (0.03) 420 SOTO-FARACO, SEBASTIÁN-GALLÉS, AND CUTLER served with segmental mismatches. Experiment one additional participant were lost because she 2 begins the segmental mismatch investigation failed to understand the task; seven participants with vowel information; vowel match and mis- were then excluded to balance the numbers in match are manipulated under the same experi- each version of the experiment (they were se- mental circumstances as in Experiment 1. lected for exclusion according to their order of arrival). The analyses were therefore again EXPERIMENT 2 based on data from 40 participants, 10 for each In the match condition of Experiment 2, we version of the materials. again presented dissyllabic word onset frag- Materials and procedure. Materials were se- ments that completely overlapped with the tar- lected on the same basis4 as in Experiment 1 get word onset. An example is aban- (from the except that members of the experimental pairs word a.ban.DO.no; “abandonment”) for the tar- mismatched in the vowel of the second syllable get ABANDONO. In the mismatch condition, instead of in stress pattern (that was the same the prime fragment overlapped with the target for both members; see a list of materials in Ap- word onset except for a vowel in the second syl- pendix A). Examples of the selected stimuli are lable (e.g., abun- from a.bun.DAN.cia; “abun- shown in Table 4. The overall log frequency of dance”) for the target ABANDONO. That is, the targets used in this experiment was 2.52 (SD where segmental structure was held constant ϭ 1.59), and the mean absolute difference be- and stress pattern manipulated in Experiment 1, tween the log frequency of the members of each we here held stress pattern constant and manip- pair was 1.73 (SD ϭ 1.45). ulated the nature of a single vowel. Finally, so that any facilitatory/inhibitory effects would be Results and Discussion comparable across experiments, primes in the On the basis of error percentages, 5 items control condition were again one-syllable frag- (along with their associated experimental pairs) ments taken from the onset of unrelated were removed from the analyses (more than words—for example, e- from e.LAS.ti.co (“elas- 15% overall error rate), leaving a total of 38 tic”) for the target ABANDONO. items. Average RTs and error rates for each As in Experiment 1, we predicted that match- prime type are shown in Table 3. Analyses of ing primes would facilitate responses to the tar- variance on RTs as a function of Prime Type get in comparison with control primes. As be- showed a significant effect (F1[2, 78] ϭ 68.1, fore, we would interpret equivalent facilitation p Ͻ .001; F2[2, 74] ϭ 27.9, p Ͻ .001). The from mismatching primes as an indication that control versus match planned contrasts reached vowel information is ignored in lexical activa- significance (F1[1, 38] ϭ 95.9, p Ͻ .001; F2[1, tion, no difference between mismatching and 37] ϭ 37.0, p Ͻ .001), as did the mismatch ver- control primes as an indication that vowel infor- sus control comparisons, although the differ- mation is crucial in ensuring membership of the initial set of activated words, and inhibition for 4 In Experiment 2, due to an error during the recording, the mismatching condition relative to the con- two of the target pairs that differed in gender were assigned trol as an indication of initial activation and sub- to carrier sentences that were syntactically gender biased. sequent disadvantage in competition. Data analyses excluding these two target pairs along with their controls showed the same effects as the whole data set Method regarding the facilitation effects in RTs. The inhibition effect (Ϫ16 ms difference), did not reach significance in the RT Participants. Fifty-one participants from the analyses as assessed by the planned contrasts between con- same population as in Experiment 1 took part in trol and mismatch conditions (F1[1, 39] ϭ 2.5, p ϭ .121; this experiment in exchange for course credits. F2[1, 33] ϭ 1.9, p ϭ .177), although it was highly signifi- None had participated in Experiment 1. Data cant in the error analysis. In particular, the mismatch condi- tion (M ϭ 3.6% errors, SD ϭ 4.0) was significantly less ac- from three participants were discarded due to curate than the control condition (M ϭ 1.4% errors, SD ϭ high error rates (more than 10% overall in the 1.9) both by participants and by items (F1[1, 39] ϭ 8.8, p Ͻ experimental conditions), and the results from .005; F2[1, 33] ϭ 7.8, p Ͻ .01). MISMATCH IN LEXICAL ACCESS 421

TABLE 4 First, however, we describe an additional exper- Examples of Prime–Target Pairings for Each Condition iment in which we again manipulated segmen- in Experiment 2 tal match versus mismatch, but in which the Prime type Prime Target manipulations concerned consonants rather than vowels. Match a.bun.(DAN.cia) ABUNDANCIA Mismatch a.ban.(DO.no) ABUNDANCIA Control e.(LAS.ti.co) ABUNDANCIA EXPERIMENT 3 Note. Parentheses in the prime column indicate the cutoff In many word processing tasks, robust differ- part of the prime word. ences appear in the contributions of vowels and consonants. Thus in English, phoneme detec- tion response times are significantly slower for ence was only marginal in the items analysis vowels than for consonants (Cutler & Otake, (F1[1, 39] ϭ 6.5, p Ͻ .05; F2[1, 37] ϭ 4.1, p ϭ 1994; Hakes, 1971; Van Ooijen, 1994), and .05). In the error analyses, the effect of prime Spanish patterns itself like English in this re- type was significant (F1[2, 78] ϭ 9.4, p Ͻ .001; spect (B. Van Ooijen, A. Cutler, R. Sánchez- F2[2, 74] ϭ 4.6, p Ͻ .05); planned contrasts re- Casas, & D. G. Norris, submitted manuscript). vealed a significant difference in percentage of These results have been interpreted in terms of errors between mismatching and control condi- listeners’ sensitivity to vowel variability in natu- tions (F1[1, 39] ϭ 11.6, p Ͻ .005; F2[1, 37] ϭ ral speech (Cutler, Van Ooijen, Norris, & 9.5, p Ͻ .005). The planned contrasts between Sánchez-Casas, 1996). Similarly, in the word the control and matching condition did not reconstruction task, in which listeners turn non- reach significance (both F Ͻ 1). Analyses words into real words by changing a single across each half of the experiment again re- sound, vowel changes are easier to make than vealed a pattern similar to that of the experi- consonant changes (Cutler et al., 2000, for ment as a whole (see Appendix B). Spanish and Dutch; Van Ooijen, 1996, for Eng- Thus we observed, again, a facilitation effect lish). It is therefore possible that vowel and con- in RTs for fragment primes that exactly sonant information might also make different matched the onset of the target. In addition, contributions in the present task. The closely both RTs and errors indicated inhibition for tar- similar patterns of results that we have observed gets preceded by vowel-mismatching prime in Experiments 1 and 2 cannot conclusively fragments (that were the onset of the competi- demonstrate that suprasegmental and segmental tor word). This effect can be interpreted in effects on lexical activation are parallel. Stress terms of multiple activation and competition differences between syllables are, after all, prin- processes that characterize lexical access. cipally carried by vocalic rather than by conso- Again, the results suggest that both the match- nantal portions of the speech signal; the equiva- ing and mismatching candidates were initially lent results might therefore reflect some activated, but as soon as the mismatching vowel property of vowels that would fail to hold for gave extra evidence favoring the matching can- consonants. In Experiment 3 we therefore con- didate, the competition process lead to inhibi- tinue our investigation with a comparison of tion of the mismatching word. The present pat- match versus mismatch using a consonantal tern of effects closely resembles that obtained manipulation. in the stress pattern manipulation (i.e., facilita- Moreover, in this experiment we included an tion for the matching prime and inhibition for explicit evaluation of the actual phonological mismatching primes). Thus, it appears that lexi- distance involved in a segmental mismatch. cal stress and segmental information play simi- Such a comparison, it should be noted, is possi- lar roles in constraining lexical activation for ble with consonant mismatches but difficult or word recognition in Spanish. The relation be- even impossible with stress or vowel mis- tween stress and segmental information is con- matches. Syllables are either stressed or un- sidered in more detail in the general discussion. stressed; a range of intermediate possibilities 422 SOTO-FARACO, SEBASTIÁN-GALLÉS, AND CUTLER does not exist. Vowels in Spanish (with its five- in either of the previous experiments, nor did vowel inventory) are all more or less equally any report problems in hearing or vision. Partic- distinct.5 Phonological distance effects can be ipants were divided into two groups according examined, however, in consonant mismatch, in to the set of materials with which they were Spanish as in other languages. The 20 conso- tested (one-feature mismatch and several-fea- nants of Spanish include pairs differing in a sin- ture mismatch, 40 in each). gle phonological feature—e.g., /t/ versus /p/, Materials and procedure. Two complete sets differing only in place of articulation—and of materials were selected according to the pairs differing in more features—e.g., /f/ versus same criteria as in the previous experiments ex- /l/, differing in place and manner of articulation cept that members of each experimental pair as well as in voicing. In Experiment 3 we inves- were matched in stress pattern and all vowels, tigated the effect of single-feature mismatches but they had a mismatch in the consonantal (an example is pa.TI.lla–pa.PI.lla) as well as of sound at the onset of the second syllable. In mismatches in several phonological features one set of materials the consonantal mismatch (e.g., bo.fe.TON–bo.le.TIN). The logic here par- consisted of one feature whereas in the other allels that of the previous experiments. Namely, set of materials it was of more than one feature we expect to replicate the typical facilitatory ef- (2.58 features on average, SD ϭ 0.5). fect of matching primes with respect to the con- The experimental procedure was exactly the trols, and use the amount of inhibition in the same as in previous experiments. Examples of mismatch condition as a measure of the magni- the selected stimuli are shown in Table 5, and tude by which the manipulated property (here, the materials are listed in Appendix A. The over- one or more than one phonological feature) all log frequency of the targets used in the one- contributes to reduction the activation of com- feature mismatch set was 2.58 (SD ϭ 1.82), and peting lexical candidates. the mean absolute difference between the log frequencies of the members of each pair was Method 1.42 (SD ϭ 0.88). The overall log frequency of Participants. Ninety-six participants were re- the targets in the several-features mismatch set cruited from the same population as in Experi- was 3.09 (SD ϭ 1.34), and the mean absolute ments 1 and 2. Data from one participant were difference between the log frequencies of the lost because of experimenter error, data from members of each pair was 1.35 (SD ϭ 0.84). three participants were discarded due to an over- all error rate above 10% in the experimental tri- Results als, and data from eight more participants were Three items (along with their experimental excluded in order to balance the number of par- and control pairs) were removed from the data ticipants in each version of the lists and each ex- collected with the one-feature mismatch set be- perimental group. No participant had taken part cause of a high error rate (above 15% overall), leaving a total of 42 items for that data set. No 5 Even in other languages, a vowel-based comparison of items from the several features mismatch set degrees of phonological distance would be difficult to had to be removed from the analyses because achieve. Although in English, for instance, the vowels of bat accuracy was above 85% for all of them. We and bet are very close while the vowels of bought and beet conducted ANOVAs on RT and the accuracy are far apart, it proves difficult to find pairs of words with three or more syllables in which the first two syllables are data, including prime type as a within-partici- identical except for a difference between two such selected pants factor and mismatch group (one- vs. sev- vowels. This fact reflects interesting characteristics of the eral-feature mismatch) as a between-partici- patterning of vowels and consonants in vocabulary structure, pants factor. and these may in turn underlie the vowel–consonant differ- In the RT analyses, neither the between- ences in some processing tasks; however, for the present Ͻ purposes, the effect is to render impossible an investigation participants factor mismatch group (both Fs of vocalic distance in a fragment priming experiment of the 1) nor the interaction between mismatch group kind used in the present study. and prime type (both Fs Ͻ 1) approached sig- MISMATCH IN LEXICAL ACCESS 423

TABLE 5 Examples of Prime–Target Pairings for Each Condition in Experiment 3 Mismatch group Prime type Prime Target 1-feature mismatch Match pa.PI.(lla) PAPILLA Mismatch pa.TI.(lla) PAPILLA Control ce.(NE.fa) PAPILLA Several features mismatch Match bo.fe.(TON) BOFETON Mismatch bo.le.(TIN) BOFETON Control ga.(vi.LAN) BOFETON Note. Parentheses indicate the cutoff part of the prime word. nificance. The effect of prime type was signifi- about the phonological makeup of words in ex- cant (F1[2, 156] ϭ 254.7, p Ͻ .001; Ͻ F2[2, actly the same way whether this information is 176] ϭ 161, p Ͻ .001; see Table 3). Planned suprasegmental or segmental, vocalic or conso- contrasts showed that there were significant dif- nantal. Most remarkably, the effect of a conso- ferences between the control and matching nantal mismatch appears to be quite comparable prime conditions (F1[1, 78] ϭ 235.8, p Ͻ .001; irrespective of whether the mismatch involves a Ͻ F2[1, 88] ϭ 173, p Ͻ .001) and between the single phonological feature (inhibition due to control and mismatching prime conditions mismatch 40 ms) or more (39 ms). Phonological (F1[1, 78] ϭ 46.7, p Ͻ .001; Ͻ F2[1, 88] ϭ distance, in other words, is not a relevant factor 34.4, p Ͻ .001). in the activation and competition process: cackle The main effect of mismatch group also failed is as effective as camel in mismatching cattle. to reach significance in the accuracy analyses The absence of significant interactions between (F1[1, 78] ϭ 2.0, p ϭ .156; F2[1, 88] ϭ 1.5, p ϭ mismatch magnitude and prime type in the pres- .212). The interaction between mismatch group ent experiment suggests equivalence in the con- and prime type was also insignificant (both Fs Ͻ tribution of each type of consonantal mismatch 1). The effect of Prime Type in the accuracy data to lexical competition. Under the same logic, we was significant both by participants and by items addressed whether any of the mismatches evalu- (F1[2, 156] ϭ 30.3, p Ͻ .001; F2[2, 176] ϭ ated in the other experiments presented in this 21.7, p Ͻ .001). The planned contrasts showed study produced a significantly different amount that the accuracy in the mismatching condition of inhibition. Taking into account that every was lower than in the control condition (F1[1, other aspect of the method is equivalent across 78] ϭ 21.4, p Ͻ .001; F2 [1, 88] ϭ 15.3, p Ͻ the present experiments, differences in the .001) and that accuracy in the match condition amount of inhibition would suggest differences was higher than in the control condition (F1[1, in the contribution of the type of mismatch to 78] ϭ 16.2, p Ͻ .001; F2[1, 88] ϭ 14.2, p Ͻ lexical access. We conducted a cross-experi- .001). Separate analyses for each half of the ex- ment statistical comparison including data from periment within each group of participants all the experiments in the present study and showed the same pattern of results as that seen found no statistical differences in the amount of in the experiment as a whole (see Appendix B). inhibition that any type of mismatch (stress, vowel, consonant one-feature, and consonant Discussion several features) produced with respect to the Experiment 3 produced a pattern of results re- control condition. Detailed analyses are re- markably consistent with the findings of Experi- ported in Appendix C. ments 1 and 2; fully matching primes facilitate decisions to a target word, whereas mismatch- GENERAL DISCUSSION ing primes inhibit responses. These results Our series of experiments has assessed the ef- clearly suggest that listeners use information fects of suprasegmental (lexical stress) and seg- 424 SOTO-FARACO, SEBASTIÁN-GALLÉS, AND CUTLER mental (vocalic and consonantal) information in English (Cutler, 1986); in this language, lexical access, evaluating their contribution in a suprasegmental information is largely redun- cross-modal priming lexical decision task. The dant for word discrimination, since interword results were remarkably clear: In all experi- differences in stress pattern nearly always in- ments, we found that matching primes facili- volve vocalic differences, and thus listeners tated responses to a visually presented target, profit little from taking account of supraseg- whereas mismatching primes inhibited lexical mentals in word-form activation.6 However, decision responses. We observed comparable ef- when such redundancy is reduced (Dutch) or fects of segmental and suprasegmental mis- entirely absent (Spanish) suprasegmentals can match, of vocalic and consonantal mismatch, usefully reduce the number of potential candi- and of single-feature versus multi-feature mis- date words. We predicted, therefore, that an ef- match. These results motivate a number of gen- fective role of suprasegmental information eral conclusions. would be observed in Spanish, and our results First, consider the equivalence of the contri- were fully in accord with this prediction. For butions of suprasegmental and segmental infor- the first time, moreover, the relative contribu- mation to the activation of word forms. The tions of suprasegmental and segmental infor- differences between mismatch and control con- mation have been directly compared statisti- ditions were in the same direction (i.e., inhibi- cally, and, as we have observed, the contri- tion in the mismatch condition) in all four ex- butions appear to be equivalent. It is not clear periments and for each of the data sets analyzed that a direct metric can be established across (participants and items RTs and accuracy). This different types of features such as consonant or remarkable consistency in the data makes it stress. However, when the average RTs in all safe to conclude that, at least qualitatively, the four experiments were equated (therefore, con- effects of every type of mismatch were equiva- trols were at the same level of performance for lent. That is, mismatches in stress pattern and each; see Appendix C), there were no differ- in segmental structure apparently affected acti- ences in the actual size of the inhibitory effects vation of candidate words in the same way: produced by each feature mismatch, suggesting One activated word was favored, the other that their actual effects on speech recognition disfavored, and the consequent competition be- must not be very disparate in magnitude. tween these words led to inhibition when tar- Some previous investigations of spoken- gets were mismatched by primes. This is clear word recognition in Spanish have shown effects evidence that Spanish-speaking listeners take of lexical stress. Syllabic match effects in sylla- account of suprasegmental information in com- ble detection (faster responses to targets that puting the phonetic code that accesses stored correspond to the syllable divisions of the tar- lexical entries. As we described in the introduc- get-bearing word than to targets that do not) are tion, no current model of spoken-word recogni- stronger in disyllables with stress on the second tion takes account of the contribution of syllable (e.g., caSO) than in disyllables with suprasegmental information to word-form acti- stress on the first syllable (e.g., CAso; vation; our findings strongly suggest that they Sebastián-Gallés, Dupoux, Segui, & Mehler, should. Suprasegmental information can con- 1992). Phoneme-monitoring responses to word- strain activation in the same way as segmental initial sounds, likewise, show a larger advan- information does. tage of word over nonword items in caSO-type Similar conclusions have been drawn for than in CAso-type disyllables (Sebastián- Dutch, based on results from semantic judg- ment tasks (Koster & Cutler, 1997), gating 6 Vowel quality is perceptible by itself from a few pitch (Jongenburger, 1996), and word spotting (A. periods while stress requires at least more time, and usually reference to another syllable for comparison. This is the rea- Cutler & W. Van Donselaar, submitted manu- son that it is stress and not vowel quality that is overridden at script). Evidence from cross-modal associative this stage of processing (see Cutler, 1986, for a full discus- priming suggests a different state of affairs for sion of this matter). MISMATCH IN LEXICAL ACCESS 425

Gallés, 1996). These results were interpreted as tailored to the phonological structure of their possible evidence that unstressed syllables in native language (Cutler, 1997). For instance, al- Spanish are less efficient than stressed syllables though across many languages it appears that in activating word forms. Our present results do listeners exploit metrical structure to locate not lend direct support to such an argument, word boundaries in speech, the nature of the and we here propose an explanation of the ear- metrical structure and the consequent manner in lier stress effects in terms of the competition which it is exploited differ in languages such as process. Specifically, there are more Spanish English (Cutler & Norris, 1988; Cutler & But- words with unstressed initial syllables than terfield, 1992), French (Cutler, Mehler, Norris, with stressed initial syllables (87% vs 13% as & Segui, 1986; Mehler, Dommergues, Frauen- assessed in the LEXESP database of Spanish felder, & Segui, 1981), Catalan (Sebastián- words; Sebastián-Gallés et al., 2000), and since Gallés et al., 1992), and Japanese (Cutler & we now know that lexical activation is indeed Otake, 1994; Otake, Hatano, Cutler, & Mehler, sensitive to stress in this language, an un- 1993). Information that is available in a particu- stressed initial syllable will presumably acti- lar language but has no counterpart in other lan- vate more potential word candidates than a guages is easily exploited by native listeners— stressed initial syllable will. This increased ac- thus vowel harmony in Finnish is exploited as a tivation of word forms would then translate cue for word segmentation in that language into increased level of inhibition for mismatch- (Suomi, McQueen, & Cutler, 1997). Stress, ing candidates, leading to the effects observed then, can be seen as another such language-spe- in Sebastián-Gallés (1996). cific cue in the recognition of spoken language. What is clear is that the role of lexical stress It is irrelevant to word recognition in languages placement cues in the activation of words has without variable stress, and even within the been firmly established by the present findings, group of variable-stress languages it does not al- and that this role appears to be in no way disad- ways provide a useful cue; however, it is ex- vantaged in comparison to the part played in the ploited by listeners whose experience with the same process by segmental information. As Cut- native language has taught them that stress can ler et al. (1997) concluded, there is no reason to usefully contribute in lexical selection. view the contribution of suprasegmental and A second conclusion from our study con- segmental information in spoken-word recogni- cerns the equivalent effects of consonant and tion as differing in any principled way; listeners vowel information in constraining word-form exploit all information that can be of use to activation, contrasting with the robust differ- them. Lexical stress information is not always ences observed when using other tasks. Thus useful, of course. In English it is less useful, studies using the word reconstruction paradigm simply because it nearly always varies redun- (Cutler et al., 2000; Van Ooijen, 1996) have dantly with vowel quality. It is likewise of little shown that vowel information seems to be less value in languages in which the prosodic pattern constraining for listeners than consonant infor- of words does not vary freely, like in French mation (i.e., suggesting that vowels carry less (Dupoux, Pallier, Sebastián-Gallés, and Mehler, weight than consonants in lexical access). 1997). In Spanish, even though there are (as in Given a free choice, listeners are significantly all stress languages) very few (morphologically biased (both in RTs or in preference) toward unrelated) word pairs that are distinguished turning a nonword into a real word by substitu- solely by stress, stress pattern information is tion of a vowel rather than a consonant—i.e., useful because it can cut down the population of they prefer to change kebra into cobra rather competing word forms during the processing of than into zebra. These results hold in English, ongoing speech. To this end, as our study has Dutch, and Spanish, languages that differ in the shown, listeners use it. balance of vowels and consonants in their There are other examples supporting that lis- phonemic inventory. Other studies have also re- teners’ use of information in the speech signal is vealed processing differences between conso- 426 SOTO-FARACO, SEBASTIÁN-GALLÉS, AND CUTLER nants and vowels; phoneme detection is faster the discrepancy in the results. Both phoneme for the former than for the latter in both English detection and word reconstruction require lis- and Spanish (B. Van Ooijen et al., submitted teners to attend to a phonemic level of represen- manuscript). Vowels are in general longer than tation. Phoneme detection demands monitoring consonants; they are more resistant to noise for a phonemic target, and word reconstruction masking (Nooteboom & Doodeman, 1984) and requires substitution of a single phoneme of the less often misreported in slips of the ear (Bond input string; hence it asks the listener to con- & Garnes, 1980). It may thus seem somewhat sider the input as a sequence of individual remarkable that they should be disadvantaged phonemes. At this level of explicit decisions in these processing tasks. about phonemes, clear differences arise in the Nevertheless, robust as these vowel-conso- processing of vowels and consonants. Lexical nant differences appear to be in the phoneme decision, however, requires attention only to the detection and word reconstruction tasks, similar lexical level: is this input string (of letters or of asymmetries do not appear to arise in lexical de- sounds) an actual word or not? Prelexical pro- cision. Support for this claim, in addition to the cessing for lexical activation is sensitive to results of the present study, is available from a match versus mismatch between input and study by Cutler et al. (1999) involving two dif- stored representations, but there does not appear ferent types of lexical decision experiments. to be a categorical difference between vowels First, Cutler et al. tested whether spoken words and consonants in the type of contribution they like kebra would prime lexical decision re- deliver at this level. sponses to written words like COBRA or Our third conclusion from the present study ZEBRA. No difference was found between concerns the failure of phonological distance to these primes and an unrelated control condition; modulate the consonant mismatch effect. A certainly there was no evidence that mismatches consonantal mismatch is equally effective in vowels versus consonants exercised differing whether it involves difference in one or more effects. Second, Cutler et al. examined repeti- phonological features. This does not, of course, tion priming in auditory lexical decision in imply that phonetic similarity has no role to Dutch. A continuous sequence of spoken items play in spoken-word recognition. Studies of was presented for lexical decision, and response phonetic confusability (e.g., Wang & Bilger, times to a given target item were compared as a 1973) show that sounds that differ in more fea- function of whether a preceding item in the list tures are less often confused with one another was or was not similar to this target. Cutler et al. than sounds that differ in fewer features; under found that responses to Dutch words were facil- difficult listening conditions, such as in a noisy itated (in comparison to a control condition) if room or on a faint telephone connection, the the item immediately preceding the target mis- input may not be sufficiently clear for similar matched with the target on only a single phonemes to be distinguished (though clear phoneme (vowel or consonant). Responses to enough for dissimilar phonemes not to be con- the word kaper (“pirate”) were faster after fused). In such a case, if the vocabulary con- kamer (“room”) or koper (“buyer”) than after tains two words that happen to differ only by gretig (“greedy”), and this result also held for containing one each of two highly confusable nonword primes (i.e., responses to lepel, sounds, then both of these two words may be “spoon”, were faster after the nonwords lopel or equally well supported, and neither of them lemel than after the nonword gukte). Thus in mismatched. This is much less likely to happen lexical decision, in contrast to phoneme detec- with two words that differ in dissimilar tion and word reconstruction, there appears to phonemes, and in this way phonetic similarity be no observable asymmetry in the contribution is obviously important in activation. However, of vowel versus consonant information. when the input is clear, so that even similar There is a potentially important difference phonemes can be distinguished, a phoneme among these various tasks that can help explain that gives support to one word but mismatches MISMATCH IN LEXICAL ACCESS 427 another will add activation to the matched nodes are connected by mutual inhibitory links, word, which will then triumph over the mis- and the more any given node is activated, the matched word in the competition process, and more it passes inhibition to the other nodes at this will happen whether the difference at that the same level to which it is connected. Thus point between the two words involves similar the inhibition observed in our lexical decision or dissimilar sounds. experiments arises when (for example) one of Note that this mismatch effect is specific to two simultaneously activated words becomes the case where the distinction involves two activated to a greater extent than the other; it competing word candidates; it relies on a cru- succeeds in passing inhibition to the other, thus cial choice having to be made in favor of one reducing that word’s activation level, and this word that diverges at that point from another. process continues, with the first word’s activa- Any mismatch is enough; once one of the two tion steadily increasing and the second word’s competitors has been advantaged, the competi- activation steadily reducing. When separately tion process automatically leads it to inhibit its arriving information, however, demands (re-) competitor. The situation is different in the activation of the inhibited word in order for a case of a mismatch when no competitor is in- correct response to be made, the inhibition volved. Suppose an English listener hears en- arising from the competition process must first cyclo-; there is only one English word that is be overcome. As we pointed out in the Intro- likely to be activated by this onset fragment duction, experimental evidence already exists and it will thus be the most active candidate, in favor of such competition-based models with no real competitor. If the next phoneme (e.g., Goldinger et al., 1989; McQueen et al., in the input is /b/, instead of the appropriate 1994); the present study clearly adds further /p/, a mismatch will result, but it will have lit- support. tle effect (indeed it may not even be noticed); In summary, our three experiments have pro- substitution of a dissimilar phoneme for the vided evidence in favor of the general hypothe- /p/, for instance /z/, will, however, have a sis that listeners of a language will use all stronger effect. This is known from studies of available cues for lexical access that usefully mispronunciation detection (e.g., Marslen-Wil- serve to distinguish between words of the lan- son & Welsh, 1978) and phoneme detection guage in question. If the language contains (Connine et al. 1997). However, as Experiment pairs of words that differ suprasegmentally but 3 shows, when input is consistent with two are segmentally identical, then listeners will competing words except for a single mismatch, use suprasegmental information in lexical acti- the number of features involved in the mis- vation. Moreover, the relative contributions of match is immaterial. suprasegmental and of segmental information The fourth and final conclusion to be drawn in the lexical activation process seem quite from our findings concerns the role of inhibi- equivalent. Similarly, there seems to be no cat- tion in the lexical recognition process. The egorical difference in the contribution made by robust and consistent inhibitory effect that we vowels versus consonants, or by single-feature have observed in our experiments offers, we versus multi-feature differences between would argue, strong support for models of spo- phonemes. Any incoming speech information ken-word recognition in which lexical selection that favors one lexical candidate but mis- occurs as the outcome of a competition process matches a simultaneously active competitor between alternative candidate words simultane- will be equally effective; the competition ously activated by the input. Such models in- process is ruthless, and once one word has an clude TRACE (McClelland & Elman, 1986) advantage it will be able to triumph, regardless and Shortlist (Norris, 1994). In these models, of the type of speech information providing the the competition process is instantiated via the advantage. Finally, the results offer further sup- mechanism of lateral inhibition. At a given port for models of spoken-word recognition in- level of processing, simultaneously active volving automatic activation of word forms and 428 SOTO-FARACO, SEBASTIÁN-GALLÉS, AND CUTLER competition between activated words.

APPENDIX A: LIST OF MATERIALS IN EXPERIMENTS 1 THROUGH 3

Experiment 1 (stress mismatch; stressed syllable is underlined)

ANGULO (angle) ANGUILA (eel) INDICE (index) ARTICO (Artic) ARTICULO (article) CABALLO (horse) COMEDIA (comedy) COMEDOR (dining room) DELITO (crime) DISCIPULO (pupil) DISCIPLINA (discipline) BOLIGRAFO (pen) EJERCITO (army) EJERCICIO (exercise) CLASICO (classic) ESPECTRO (ghost) ESPECTACULO (performance) ARABE (Arabian) ESPIRITU (spirit) ESPIRAL (spiral) ALIJO (unloading) ESTAMPA (engraving) ESTAMPIDA (stampede) FISICA (physics) ESTATUA (statue) ESTATUTO (statute) MARIDO (husband) GALERA (galley) GALERIA (gallery) CAMARA (room/camera) INCENDIO (fire) INCENTIVO (incentive) PETARDO (firework) LITERA (bunk bed) LITERAL (literal) DECADA (decade) MARISCO (shellfish) MARISCAL (marshal) LAVABO (toilet) MEJILLA (cheek) MEJILLON (mussel) CANTARO (pitcher) PELOTA (ball) PELOTON (squad) MACETA (flowerpot) PRESIDIO (prison) PRESIDENTE (president) MODULO (module) PRINCIPE (prince) PRINCIPIO (beginning) MOSQUITO (mosquito) PROCESO (process) PROCESION (procession) MEDICO (doctor) PROYECTO (project) PROYECTIL (missile) CASTILLO (castle) SECRETO (secret) SECRETARIO (secretary) VICARIO (curate) SERENO (calm) SERENATA (serenade) REPLICA (answer) SUCESO (event) SUCESION (succession) TEJADO (roof)

Experiment 2 (vocalic mismatch)

ABONADO (subscriber) ABANICO (fan) OSADO (daring) ABUNDANCIA (abundance) ABANDONO (abandonment) ELASTICO (elastic) ACOTAR (to limit) ACATAR (to obey) ENFILAR (to thread) ALUMBRADO (lighting) ALAMBRADA (wire fence) ASENTIR (agree) APETITO (appetite) APATIA (apathy) ELEGIDO (selected) ASTURIANO (Asturian) ASTERISCO (asterisk) ENCIMERA (worktop) CABELLERA (hair) CABALLERO (gentleman) SOLEDAD (loneliness) COMPETIDOR (competitor) COMPATIBLE (compatible) TONTERIA (silliness) EMBESTIDA (charge) EMBUSTERO (liar) ANGUSTIA (distress) ESCONDITE (hiding place) ESCANDALO (scandal) ALMENDRO (almond tree) FRANQUICIA (franchise) FRANQUEZA (frankness) CRISTALINO (limpid) HISTERIA (hysteria) HISTORIA (history) ALMEJAS (clam) INDUCIDO (induced) INDECISO (undecided) ARTILUGIO (gadget) MINORIA (minority) MINERIA (mining) CALAMIDAD (disaster) OCTUBRE (October) OCTAVO (eighth) ESTUFA (heater) RELLENO (filling) RELLANO (landing) VISITA (visit) REPORTERO (journalist) REPERTORIO (repertoire) FILANTROPO (philanthropist) RESULTADO (result) RESALTADO (highlighted) COBALTO (cobalt) SARDINA (sardine) SARDANA (Catalan dance) MANTILLA (mantilla)

Experiment 3 (1-feature consonantal mismatch)

AFILADOR (sharpener) AGILIDAD (agility) UNIVERSAL (universal) APARATO (device) AVARICIA (greed) ELEGANTE (elegant) APESTOSO (stinking) ATESTADO (testimonial) EFECTO (effect) APROBADO (approved) ACROBACIA (acrobatics) EDREDON (quilt) CALIDAD (quality) CARIDAD (charity) JUVENIL (young) CAMERINO (dressing room) CAÑERIA (piping) MONITOR (monitor) CAMINO (path) CABINA (cabin) PELIGRO (danger) MISMATCH IN LEXICAL ACCESS 429

CATETO (yokel) CADETE (cadet) LINAJE (lineage) COCINERO (coock) COJINETE (bearing) DEGOLLADO (beheaded) CONCESION (concession) CONFESION (confession) GESTACION (gestation) DELEGADO (representative) DENEGADO (denied) FINALISTA (finalist) DIMISION (resignement) DIVISION (division) RELIGION (religion) EPICO (epic) ETICO (ethic) AGUILA (eagle) ESPIRAL (spiral) ESTIRON (jerk) ANDALUZ (andalusian) MATERIA (matter) MADERA (wood) TOMILLO (thyme) PAPILLA (baby food) PATILLA (sideburn) CENEFA (trimming) PROCESION (parade) PROFESION (career) CLARIDAD (clarity) RECADERO (messenger) REGADERA (sprinkler) LITERARIO (literary) REDENCION (redemption) RETENCION (retention) DIMENSION (dimension) TEMIDO (feared) TEÑIDO (dyed) MOROSO (bad payer) TRAFICO (traffic) TRAGICO (tragic) CRITICO (critic) Experiment 3 (2 or more features consonantal mismatch)

AFONICO (voiceless) ANONIMO (anonymous) ILICITO (ilegal) AGOSTO (August) APOSTOL (apostle) HORRENDO (horrible) ATENTADO (assault) AVENTURA (adventure) ILUSTRADO (illustrated) BOFETON (smack) BOLETIN (bulletin) GAVILAN (sparrowhawk) CALIDAD (quality) CAVIDAD (cavity) TOCADOR (dressing table) CARTELERA (billboard) CARCELERO (jailer) MERMELADA (jam) CINICO (cynical) CIVICO (civic) FASICO (phasic) COLISION (colision) COMISION (comitee) TEJEDOR (weaver) DEDICADO (devoted) DELICADO (delicate) MACERADO (macerate) INCENTIVO (incentive) INVENTARIO (inventary) ARGENTINO (Argentinian) INFECCION (infection) INYECCION (injection) HERMANDAD (brotherhood/sisterhood) INFERIOR (inferior) INTERIOR (interior) ESPIRAL (spiral) MAJESTAD (majesty) MALESTAR (discomfort) CORRECTOR (proofreader) MIMICA (mimicry) MITICO (mythical) COLICO (colic) PAPILLA (baby food) PASILLO (corridor) BELLEZA (beauty) PROTECTOR (protective) PROYECTIL (projectile) TRILLADOR (threshing machine) RECATADO (polite) RELATIVO (relative) CAMILLERO (stretcher bearer) REFERENCIA (reference) REVERENCIA (reverence) SOÑOLIENTO (sleepy) REPENTINO (sudden) RESENTIDO (resentful) CALENDARIO (calendar) SAGITARIO (Sagittarius) SANITARIO (sanitary) TORREFACTO (roasted) SEGADOR (reaper) SENADOR (senator) CABEZON (bigheaded) SINFONIA (sinphony) SINTONIA (tuning) DESMEDIDO (disproportionate) TONALIDAD (tonality) TOTALIDAD (whole) FACILIDAD (easiness)

VISCERA (entrails) VISPERA (eve) Experiment 1. RTs were faster (F1[1, 39] ϭ 101.1, p Ͻ TOMBOLA (raffle) .001; F2[1, 43] ϭ 106.4, p Ͻ .001) and responses more ac- curate (F1[1, 39] ϭ 13.5, p Ͻ .005; F2[1, 43] ϭ 5.3, p Ͻ A list of the complete carrier sentences can be found at .05) in the second half of the experiment as compared to http://www.idealibrary.com the first half. The interaction between experimental half APPENDIX B: ANALYSES INCLUDING THE and Prime Type did not reach significance for RT (both EXPERIMENTAL HALF AS A FACTOR Fs Ͻ 1), but experimental half interacted with prime type for accuracy (F1[2, 78] ϭ 8.4, p Ͻ .005; F2[2, As each participant was presented with two sets of ma- 86] ϭ 8.0, p Ͻ .005). Both the first and second halves of terials that included a repetition of the targets (albeit no the experiment showed a main effect of prime type (F1[2, given prime–target combination), we were concerned to es- 78] ϭ 6.4, p Ͻ .005; F2[2, 86] ϭ 4.1, p Ͻ .05 and F1[2, tablish whether repetition effects might have contaminated 78] ϭ 3.1, p Ͻ .05; F2[2, 86] ϭ 4.2, p Ͻ .05, respec- the overall effects. Here we present analyses testing the ef- tively). Planned contrasts for each half of the experiment fect of the experimental half and its interaction with the separately showed that in the first half of the experiment prime type factor. The main effect of Prime Type is ignored responses were more accurate in the matching condition here, as it can be found in the Results of the corresponding than in the control condition (F1[1, 39] ϭ 6.3, p Ͻ .05; experiments. F2[1, 43] ϭ 5.3, p Ͻ .05), and marginally less accurate in 430 SOTO-FARACO, SEBASTIÁN-GALLÉS, AND CUTLER the mismatching condition than in the control condition sponses were faster (F1[1, 39] ϭ 72.4, p Ͻ .001; F2[1, 47] (F1[1, 39] ϭ 3.1, p ϭ .08; F2[1, 43] ϭ 2.5, p ϭ .121]; in ϭ 161.0, p Ͻ .001) and more accurate (F1[1, 39] ϭ 17.9, p the second experimental half matching primes again Ͻ .001; F2[1, 47] ϭ 10.5, p Ͻ .005) in the second half of yielded more accurate responses than control primes (F1[1, the experiment than in the first half. The interaction between 39] ϭ 3.2, p ϭ .08; F2 [1, 43] ϭ 4.3, p Ͻ .05), but here experimental half and Prime Type did not reach significance there were no differences between control and mismatch in the RT analyses (F1 Ͻ 1; F2[2, 94] ϭ 1.9, p ϭ .145) or in prime conditions (both Fs Ͻ 1). the accuracy data (F1[2, 78] ϭ 2.6, p ϭ .075; F2[2, 94] ϭ Experiment 2. RTs were faster (F1[1, 39] ϭ 74.5, p Ͻ 1.9, p ϭ .154) .001; F2[1, 37] ϭ 136.1, p Ͻ .001) and responses were more accurate (F1[1, 39] ϭ 17.7, p Ͻ .001; F2[1, 37] ϭ 10.8, p Ͻ APPENDIX C: CROSS-EXPERIMENT ANALYSES .005) in the second half of the experiment than in the first The pooled data of all experiments were submitted to analy- half. The interaction between experimental half and prime ses of variance (we included the two groups of Experiment 3 type was significant for RT (F1[2, 78] ϭ 5.6, p Ͻ .01; F2[2, as two different experiments in these analyses). Given that 74] ϭ 6.1, p Ͻ .005); the difference between mismatching the main effect of Prime Type was significant in all experi- and control conditions was significant in the first half of the ments, it is not surprising that here too it was significant, experiment (F1[1, 39] ϭ 5.1, p Ͻ .05; F2[1, 37] ϭ 4.5, p Ͻ both in the RT and the error analyses; on both measures, the .05) but not in the second half (F1[1, 39] ϭ 2.8, p Ͼ .1; differences between the control condition and each of the F2[1, 37] ϭ 1.2, p Ͼ .2). The matching vs control difference other two conditions were separately significant. We report was significant in both halves (all p Ͻ .001). in detail only effects involving the factor Experiment, spe- The accuracy data also showed an interaction between cific to these pooled analyses. experimental half and prime type (F1[2, 78] ϭ 5.6, p Ͻ .01; In the accuracy analyses, there was no main effect of Ex- ϭ Ͻ F2[2, 74] 6.1, p .05). Here the error rate was signifi- periment and no interaction of this factor with Prime Type. cantly higher in the mismatching condition than in the con- In the RT analyses, the main effect of Experiment was mar- ϭ trol condition in the first half of the experiment (F1[1, 39] ginally significant by participants and significant by items Ͻ ϭ Ͻ 11.7, p .005; F2[1, 37] 8.4, p .01) but not in the sec- (F1[3, 156] ϭ 2.3, p ϭ .08; F2[3, 168] ϭ 7.4, ϭ ϭ ϭ ϭ ond half (F1[1, 39] 1.5, p .2; F2[1, 37] 1.2, p .2). p Ͻ .001), indicating overall differences in mean RT across The matching and the control conditions did not differ in ei- the different experiments. Pairwise comparisons (Bonfer- ther half of Experiment 2 (2.2% vs 2.1% in the first half, and roni) on the mean RTs across items indicated that responses 0.6% vs 1.1% in the second half). were faster in Experiment 1 than in Experiment 2 (p Ͻ .05), Experiment 3: One-feature mismatch group. The main and than in both groups of Experiment 3 (p Ͻ .001 in both). effect of experimental half was significant both for RT These differences did not reach significance in the analysis (F1[1, 39] ϭ 128.3, p Ͻ .001; F2[1, 41] ϭ 131.8, p Ͻ .001) by participants. No other pairwise comparisons reached sig- and accuracy (F1[1, 39] ϭ 21.3, p Ͻ .001; F2[1, 41] ϭ 25.7, nificance. The interaction between Experiment and Prime p Ͻ .001) in the same direction as in the preceding experi- Type was also significant (F1[6, 312] ϭ 9.6, p Ͻ .001; F2[6, ments. In RTs, the interaction between experimental half ϭ Ͻ and prime type was significant (F1[2, 78] ϭ 4.7, p Ͻ .05; 336] 5.1, p .001). Planned contrasts revealed that the F2[2, 82] ϭ 4.2, p Ͻ .05). In fact, the prime type effect was source of the interaction was a significant difference in the ϭ Ͻ significant in both halves of Experiment 3 (1st: F1[2, 78] ϭ size of the control vs match effect (F1[3, 156] 10.6, p ϭ Ͻ 92.0, p Ͻ .001; F2[2, 82] ϭ 56.2, p Ͻ .001; 2nd: F1[2, 78] ϭ .001; F2[3, 168] 6.2, p .001) across experiments. Pair- 72.1, p Ͻ .001; F2[2, 82] ϭ 53.0, p Ͻ .001); the source of wise comparisons (Bonferroni) showed that there was less the interaction was a difference in the size of the mismatch facilitation for matching primes in Experiment 1 than in the vs control effect (Ϫ55 ms in the first half and Ϫ29 ms in the other experiments (pϽ. 001 for all comparisons involving second, both significant, p Ͻ .001 and p Ͻ .05, respec- the matching effect of Experiment 1 with that of the other tively). The matching vs control difference was significant experiments). The size of the control vs mismatch effect, on and equivalent for both experimental halves (92 and 83 ms, the other hand, did not significantly differ across experi- p Ͻ .001 and p Ͻ .005). ments7 (F1[3, 156] ϭ 1.7, p Ͼ .1; F2[3, 168] ϭ 1.1, p Ͼ .3). The accuracy data also showed an interaction between The faster RTs in Experiment 1 than in Experiments 2 experimental half and prime type (F1[2, 78] ϭ 4.8, p Ͻ .05; and 3 are presumably due to the higher frequency of items in F2[2, 82] ϭ 3.8, p Ͻ .05); the effect of prime type was sig- Experiment 1 (M ϭ 3.61, SD ϭ 1.5) as compared to items in nificant in the first half of the experiment (F1[2, 78] ϭ 10.5, Experiments 2 (M ϭ 2.52, SD ϭ 1.59) and both groups of p Ͻ .001; F2[2, 82] ϭ 6.5, p Ͻ .005) but not in the second Experiment 3 (M ϭ 2.58, SD ϭ 1.82; M ϭ 3.09, SD ϭ 1.34, (F1[2, 78] ϭ 1.6, p ϭ .208; F2[2, 82] ϭ 1.9, p ϭ .153). for one- and several-feature mismatch sets, respectively). Planned contrasts for the first experimental half revealed that The fact that the significant differences were restricted to the the prime type effect was due to a significant difference be- tween the mismatching and the control condition (F1[1, 39] ϭ 8.6, p Ͻ .01; F2[1, 41] ϭ 4.5, p Ͻ .05), and a marginal dif- 7 This result remains the same after we exclude the gen- ference between control and matching conditions (F1[1, 39] ϭ der-biased item in Experiment 2; F1[3, 156] ϭ 2.0, p ϭ .113 4.0, p ϭ .06; F2[1, 41] ϭ 2.8, p ϭ .1). and F2[3, 164] ϭ 1.2, p ϭ .286 in the RTs analyses; and Experiment 3: Several feature mismatch group.Re- both Fs Ͻ 1 for the error analyses. MISMATCH IN LEXICAL ACCESS 431 items analysis further supports this account. However, this 385–400. overall difference in RTs, renders the interpretation of the Cutler, A., & Norris, D. G. (1988). The role of strong syl- rest of effects in the analysis difficult (since the detected in- lables in segmentation for lexical access. Journal of teraction may be related to overall differences in RTs). In Experimental Psychology: Human Perception and order to circumvent possible confounds due to overall dif- Performance, 14, 113–121. ferences in RTs, further analyses were conducted. We fil- Cutler, A., & Otake, T. (1994). Mora or phoneme? Further tered the data in the following way. For the participants’ evidence for language specific listening. Journal of analyses, we excluded the 12% of the participants with Memory and Language, 33, 824–844. faster average RTs in Experiment 1, and the 12% of the par- Cutler, A., & Otake, T. (1999). Pitch accent in spoken-word ticipants with slower average RTs in Experiments 2, and in recognition in Japanese. Journal of the Acoustical Soci- each of the groups of Experiment 3. For the items analyses ety of America, 105, 1877–1888. we excluded 25% of the items with fastest RTs in Experi- Cutler, A., Van Ooijen, B., & Norris, D. G. (1999). Vowels, ment 1 and the 25% of the items with slowest RTs in Exper- consonants, and lexical activation. Proceedings of the iments 2, and 3 (each group separately). Fourteenth International Congress of Phonetic Sci- Data filtering ensured that the main effect of the factor ences (Vol. 3, pp. 2053–2056). San Francisco. Experiment did not approach significance (both Fs Ͻ 1) in Cutler, A., Van Ooijen, D., Norris, D. G., & Sánchez-Casas, these analyses. However, the interaction between Prime R. (1996). Speeded detection of vowels: A cross-lin- Type and Experiment reached significance again (F1[6, 272] ϭ guistic study. Perception & Psychophysics, 58, 7.9, p Ͻ .001; F2[6, 302] ϭ 7.4, p Ͻ .001). The planned 807–822. contrasts indicated that the difference between the control Cutler, A., Sebastián-Gallés, N., Soler-Vilageliu, O., & Van and match conditions differed across experiments (F1[3, Ooijen, B. (2000). Constraints of vowels and conso- 136] ϭ 8.9, p Ͻ .001; F2[3, 151] ϭ 8.6, p Ͻ .001). Pairwise nants on lexical selection: Cross-linguistic compar- comparisons (Bonferroni) revealed significant differences in isons. Memory & Cognition, 28, 746–755. the size of the facilitation effect in Experiment 1 as com- Dupoux, E., Pallier, C., Sebastián, N., & Mehler, J. (1997). Ͻ pared to the rest of the experiments (p .005 for the com- A destressing “deafness” in French. Journal of Memory Ͻ parison with Experiment 2, and p .001 each for the com- and Language, 36, 406–421. parison with both groups of Experiment 3). Planned Elman, J., & McClelland, J. (1986). Exploiting lawful vari- contrasts between the control and the mismatch condition ability in the speech wave. In J. Perkell and D. H. ϭ Ͼ remained not significant (F1[3, 136] 1.7, p .1; F2[3, Klatt (Eds.), Invariance and variability in speech ϭ ϭ 151] 1.2 p .2), as in the previous analysis. processes (pp. 360–385). Hillsdale, NJ: Erlbaum. Fox, R. A., & Unkefer, J. (1985). The effect of lexical status REFERENCES on the perception of tone. Journal of Chinese Linguis- Boelte, J. (1997). The role of mismatching information in tics, 13, 69–90. spoken word recognition. Hamburg: Verlag Dr. Kovac. Goldinger, S., Luce, P. A., & Pisoni, D. (1989). Priming lex- Bond, Z. S., & Garnes, S. (1980). Misperceptions of fluent ical neighbors of spoken words: Effects of competition speech. In R. Cole (Ed.), Perception and production of and inhibition. Journal of Memory and Language, 28, fluent speech (pp. 115–132). Hillsdale, NJ: Erlbaum. 501–518. Connine, C. M., Blasko, D. G., & Wang, J. (1994). Vertical Gow, D., & Gordon, P. (1995). Lexical and pre-lexical influ- similarity in spoken word recognition: Multiple lexical ences in word segmentation: Evidence from priming. activation, individual differences, and the role of sentence Journal of Experimental Psychology: Human Percep- context. Perception & Psychophysics, 56, 624–636. tion and Performance, 21, 344–459. Connine, C. M., Titone, D., Deelman, T., & Blasko, D. G. Hakes, D. T. (1971). Does verb structure affect sentence (1997). Similarity mapping in spoken word recogni- comprehension? Perception & Psychophysics, 10, tion. Journal of Memory and Language, 37, 463–480. 229–232. Cutler, A. (1986). Forbear is a homophone: Lexical prosody Jongenburger, W. (1996). The role of lexical stress during does not constrain lexical access. Language and spoken-word processing. Unpublished Ph.D. thesis, Speech, 29, 201–220. University of Leiden. Cutler, A. (1997). The comparative perspective on spoken- Koster, M., & Cutler, A. (1997). Segmental and supraseg- language processing. Speech Communication, 31, 3–15. mental contributions to spoken-word recognition in Cutler, A., & Butterfield, S. (1992). Rhythmic cues to speech Dutch. Proceedings of EUROSPEECH 97 (pp. segmentation: Evidence from juncture misperception. 2167–2170). Rhodes. Journal of Memory and Language, 31, 218–236. Luce, P. A., Pisoni, D. B., & Goldinger, S. D. (1990). Simi- Cutler, A., Dahan, D., & Van Donselaar, W. (1997). Prosody larity neighborhoods of spoken words. In G. T. M. Alt- in the comprehension of spoken language: A literature mann (Ed.), Cognitive models of speech processing, review. Language and Speech, 40, 141–201. psycholinguistic and computational perspectives (pp. Cutler, A., Mehler, J., Norris, D. G., & Segui, J. (1986). The 122–147). Cambridge, MA: MIT Press. syllable’s differing role in the segmentation of French Marslen-Wilson, W. D. (1990). Activation, competition and and English. Journal of Memory and Language, 25, frequency in lexical access. In G. T. M. Altmann (Ed.), 432 SOTO-FARACO, SEBASTIÁN-GALLÉS, AND CUTLER

Cognitive models of speech processing: Psycholinguis- Sebastián-Gallés, N. (1996). The role of accent in speech tic and computational perspectives (pp. 148–172). perception. In T. Otake & A. Cutler (Eds.), Phonolog- Cambridge, MA: MIT Press. ical structure and language processing: Cross- Marslen-Wilson, W. D., & Warren, P. (1994). Levels of per- linguistic studies (pp. 171–182). Berlin: Mouton de ceptual representation and process in lexical access: Gruyter. Words, phonemes, and features. Psychological Review, Sebastián-Gallés, N., Dupoux, E., Segui, J., & Mehler, J. 101, 653–675. (1992). Contrasting syllabic effects in Catalan and Marslen-Wilson, W. D., & Welsh, A. (1978). Processing in- Spanish. Journal of Memory and Language, 31, teractions and lexical access during word recognition in 18–32. continuous speech. Cognitive Psychology, 10, 29–63. Sebastián-Gallés, N., Martí, A., Cuetos, F., & Carreiras, M. McClelland, J. L., & Elman, J. L. (1986). The TRACE (2000). LEXESP: Base de datos informatizada del es- model of speech perception. Cognitive Psychology, 18, pañol. Barcelona: Edicions de la Universitat de 1–86. Barcelona. McQueen, J. M., Norris, D. G., & Cutler, A. (1994). Compe- Shillcock, R. C. (1990). Lexical hypotheses in continuous tition in spoken word recognition: Spotting words in speech. In G. T. M. Altmann (Ed.), Cognitive models other words. Journal of Experimental Psychology: of speech processing: Psycholinguistic and computa- Learning, Memory, and Cognition, 20, 621–638. tional perspectives (pp. 24–49). Cambridge, MA: MIT Mehler, J., Dommergues, J. Y., Frauenfelder, U. H., & Segui, Press. J. (1981). The syllable’s role in speech segmentation. Skelton, R. B. (1969). The pattern of Spanish vowel sounds. Journal of Verbal Learning and Verbal Behavior, 20, International Review of Applied Linguistics in Lan- 298–305. guage Teaching, 7, 231–237. Navarro-Tomás, T. (1968). Manual de Pronunciación Es- Stockwell, R. P., & Bowen, J. D. (1965). The sounds of Eng- pañola (14th ed.). Madrid: Consejo Superior de Inves- lish and Spanish. Chicago: Univ. of Chicago Press. tigaciones Científicas. Suomi, K., McQueen, J. M., & Cutler, A. (1997). Vowel har- Nooteboom, S. G., & Doodeman, G. J. N. (1984). Speech mony and speech segmentation in Finnish. Journal of quality and the gating paradigm. In M. P. R. van den Memory and Language, 36, 422–444. Broecke & A. Cohen (Eds.), Proceedings of the 11th Tabossi, P., Burani, C., & Scott, D. (1995). Word identifica- International Congress of Phonetic Sciences (pp. tion in fluent speech. Journal of Memory and Lan- 481–485). Dordrecht: Foris. guage, 34, 440–467. Norris, D. G. (1994). Shortlist: A connectionist model of Van Ooijen, B. (1994). The processing of vowels and continuous speech recognition. Cognition, 52, 189–234. consonants. Unpublished Ph.D. thesis, University of Otake, T., Hatano, G., Cutler, A., & Mehler, J. (1993). Mora Leiden. or syllable? Speech segmentation in Japanese. Journal Van Ooijen, B. (1996). Vowel mutability and lexical selec- of Memory and Language, 32, 258–278. tion in English: Evidence from a word reconstruction Pallier, C., Dupoux, E., & Jeannin, X. (1997). EXPE: An ex- task. Memory & Cognition, 24, 573–583. pandable programming language for on-line psycho- Wang, M. D., & Bilger, R. C. (1973). Consonant confusions logical experiments. Behavior Research Methods, In- in noise: A study of perceptual features. Journal of the struments, and Computers, 29, 322–327. Acoustical Society of America, 54, 1248–1266. Radeau, M., Morais, J., & Segui, J. (1995). Phonological Zwitserlood, P. (1989). The locus of the effects of the sen- priming between monosyllabic spoken words. Journal tential-semantic context in spoken word processing. of Experimental Psychology: Human Perception and Cognition, 32, 25–64. Performance, 21, 1297–1311. Zwitserlood, P. (1996). Form priming. Language and Cogni- tive Processes, 11, 589–596.

(Received July 12, 2000) (Revision received December 19, 2000) Published online August 22, 2001