Reviews

infection of ~nurinecells (15) and transgenic mice expressing human CD4 (16) and pro- vides a rationale for transgenic approaches to developing animal models of HIV disease.

REFERENCES AND NOTES

I . F. Cocchi et a/.,Science 270, 1811 (1995). 2. Y. Feng, C. C. Broder, P. E. Kennedy, E. A. Berger,

ibid. 272, 872 (1996). 3. M. Samson, 0. Labbe, C. Mollereau, G. Vassart, M.

Parmentier, Biochemistry 35, 3362 (1996); C. J. Raport, J. Gosling, V. L. Schweckart, P. W. Gray, I . F. Charo, J. 6/01, Chem. 271, 17161 (1996).

4. H. Choe et a/.,Cell 85, 1135 ( I996); B. J. Doranz et a/., ibid., p. 1149.

5. T. Dragic eta/., Nature 381, 667 (1996); H. Deng et a/.,ib@, p. 661; G. Akhatibetal.,Science 272, 1955 ( I996).

6. S. Gartner eta/.,Science 233, 215 (1986). 7. R. Atchison eta/., unpubl~shedobservatons. 8. L. Boring et a/.,J. Bioi. Chem. 271, 7551 (1996). 9. We cloned cDNAs encoding human or murine CCR5

n t o the expression vector pcDNA3 (Invitrogen)after engneering the FLAG epitope Into the NH,-termnus as descrbed (13).Expresson of each construct was determned by FACS w ~ t han antbody to FLAG (anti- FLAG) (Boehringer Mannhem), and relative expres- sion for each (see below) was calculated as the per- centage of cells expressng human CCR5 on the cell surface normalizedto the expresson of hCCR5 (de- fined as loo%), w ~ t hstandard errors of the mean. The mean fluorescence intensty of the postve cells from any sngle sample never variedfrom the average by more than 30% in a single experiment. Therefore, neither the relative number of postve cells nor the absolute expression levels w~thintransfected cells explains the differences in coreceptor activity. C ~ I - meric receptors were prepared by the overlap poly- merase chain reactlon (PCR) method (17). hCCR5 (HHHH), human CCR5 (100% relative expression); mCCR5 (MMMM), murlne CCR5 (126 2 49%); HMMM, NH,-terminus of human CCR5 [am~noacids (aa) 1 to 321 fused to murlne CCR5 (aa 35 to 354) (77 ? 22%): MHHH, NH,-termnus of murne CCR5 (aa 1 to 34) fused to human CCR5 (aa 33 to 352) (73 -t 17%); MHMM, extracellular loop 1 and a por- tion of transmembrane domain 3 of human CCR5 (aa 86 to 118) replacing the corresponding segment of the murlne receptor (aa 88 to 120) (37 -t 22%); MMHM, extracellular loop 2 and adjacent porions of human CCR5 (aa 134 to 210) replacing the corre- sponding regon of the murine receptor (aa 136 to 212) (81 2 30%); MMHH, NH,-term~nal half of mCCR5 (aa 1 to 162) fused to the COOH-termnal half of hCCR5 (aa 161 to 352) (80 -t 39%).

10. I . F. Charo et a/., Proc. Nati. Acad. Sci. U.S.A. 91, 2752 (1994).

11. C. Franci, L. M. Wong, J. Van Damme, P. Proost, I . F. Charo, J. /mmuno/. 154, 6511 (1995).

12. We cloned cDNAs encoding human CCR2B or chi- meras into the expresson vector pCMV4 (18) after engineeringthe FLAG epitope into the NH,-terminus as described (13).Express~onof each construct (see below) was determned as descrbed earlier. Chimer- ic receptors were prepared by the overlap PCR method (17). 5555, human CCR5 (100% relat~ve expression);2222, human CCR2B (87 ? 2%);5222, NH,-termnus of CCR5 (aa 1 to 32) fused to CCR2B (aa 45 to 360) (27 -t 5%); 2555, NH,-terminus of CCR2B (aa 1 to 44) fused to CCR5 (aa 33 to 352) (108 ? 17%); 2255, CCR2B (aa 1 to 136) fused to CCR5 (aa 124 to 352) (119 -t 33%).

13. F. S. Monteclaro and I. F. Charo, J. Biol. Chem. 271, 19084 (1996); F. S. Monteclaro et ai., unpubl~shed observations.

14. J. Gosling eta/., unpubshed observat~ons. 15. P. J. Maddon eta/., Cell 47, 333 (1986). 16. P. Lores et a/.,AIDS Res. Hum. Retroviruses8,2063

(1992). 17. S. N. Ho, H. D. Hunt, R. M. Horton, J. K. Pulen, L. R.

Pease, Gene 77, 51 ( I989). 18. S. Andersson, D. L. Davis, H. Dahlback, H. Jornval,

D. W. Russell,J. Biol. Chem. 264, 8222 (1989). 19. M. A. Godsmth, M. T. Warmerdam, R. E. Atchison,

M. D. Miller, W. C. Greene,J. Virol. 69, 41 12 (1995). 20. COS-7 cells were transfected w ~ t h2 p,g of pasmid

DNA per well in a six-well plate as descrbed (19). DNA samples consisted of appropriate combina- t o n s of 0.5 p,g of a human CD4 expression plasmd [pCD4Neo (19)l or pla~nvector, and 1 5 p,g of a chemokine receptor-expressng plasmd or plain vector. About 30 hours after additon of DNA, the medium in each w e was replaced with 1.0 ml of medium contanng H V - I Ba-L (-1 00 to 170 ng of p24 per sample; source: N H ADS Reagent Re- pository, passaged on prlmary human macro- phages).About 10 hours later, an additona 1.0 ml of medium was added to each well. After 30 hours, the cells were recovered from the dish as described (19) and analyzed with a FacScan (Becton Dickin- son). Stainincl for intracvto~lasmicH V - 1 w24 was

(Coulter Immunology) and goat ant-mouse fuores- cein sothocyanate (FTC)-conjugated secondary ant~body(Becton Dick~nson).Cells were further sta~nedwith phycoerythr~n(PE)-conjugated antl- CD4 (Becton Dcknson).Appropriate controls n d - cated that the appearance of double-positve cells (FTC + PE) was dependent on cotransfection with both CD4 and human CCR5 expression plasmids and on the presence of HIV-I Ba-L.

21. H. Arai and I . F. Charo, J. 6/01, Chem. 271, 21814 ( I996).

22. We acknowledge the adv~ceof M. Warmerdam (transfecton-nfectionassay),E. Welder (FACS stud- ies), and L. Borng, H. Ara~,and R. Speck (scientf~c interpretation). We appreciate the ass~stanceof J. Carroll and M. Cenceros In the preparation of t h s manuscript. Supported In part by NIH grant HL52773 (I.F.C.)and by Pfzer (M.A.G.).

carr~edout ~ 7 t hthe Fix and Perm reagents (Caltag Laboratories), with a monoclonal antibody to p24 24 September 1996; accepted 24 October 1996

Statistical Learning by 8-Month-Old Infants Jenny R. Saffran, Richard N. Aslin, Elissa L. Newport

Learners rely on a combination of experience-independent and experience-dependent mechanisms to extract information from the environment. Language acquisition involves both types of mechanisms, but most theorists emphasize the relative importance of experience-independent mechanisms. The present study shows that a fundamental task of language acquisition, segmentation of words from fluent speech, can be accom- plished by 8-month-old infants based solely on the statistical relationships between neighboring speech sounds. Moreover, this word segmentation was based on statistical learning from only 2 minutes of exposure, suggesting that infants have access to a powerful mechanism for the computation of statistical properties of the language input.

D u r i n g early development, the speed and accuracy with which an organism extracts environmental information can be ex- tremely important for its survival. Some species have evolved highly constrained neural mechanisms to ensure that environ- mental information is properly interpreted, even in the absence of experience with the environment (1). Other species are depen- dent on a period of interaction with the environment that clarifies the information to which attention should be directed and the consequences of behaviors guided by that information ( 2 ) . Depending on the developmental status and the task facing a particular organism, both experience-inde- pendent and experience-dependent mecha- nisms may be involved in the extraction of information and the control of behavior.

In the domain of language acquisition, two facts have supported the interpretation that experience-independent mechanisms are both necessary and dominant. First, highly complex forms of language produc- tion develop extremely rapidly (3). Second, the language input available to the young child is both incomplete and sparsely rep-

Department of Brain and Cognitve Sciences, Unversity of Rochester, Rochester, NY 14627, USA.

resented compared to the child's eventual linguistic abilities (4). Thus, most theories of language acquisition have emphasized the critical role played by experience-inde- pendent internal structures over the role of experience-dependent factors (5).

It is undeniable that experience-depen- dent ~nechanis~nsare also required for the acquisition of language. Many aspects of a particular natural language must be ac- quired from listening experience. For exam- ple, acquiring the specific words and pho- nological structure of a language requires exposure to a significant corpus of language input. Moreover, long before infants begin to produce their native language, they ac- quire information about its sound properties (6). Nevertheless, given the daunting task of acquiring linguistic information from lis- tening experience during early develop- ment, few theorists have entertained the hypothesis that learning plays a primary role in the acquisition of more complicat- ed aspects of language, favoring instead experience-independent mechanisms ( 7 ) . Young humans are generally viewed as poor learners, suggesting that innate fac- tors are primarily responsible for the ac- quisition of language.

Here we investigate the nature of the

SCIENCE VOL. 274 13 DECEMBER 1996

experience-dependent factors involved in language acquisition. In particular, we ask whether infants are in fact better learners than has previously been assumed, thus po- tentially reducing the extent to which ex- ~erience-indevendentstructures must be

rnaterial that serves as a potential learning experience. They are subsequently present- ed with two types of test stimuli: ( i ) items that were contained within the familiariza-

word stimuli (18), with longer listening times for nonwords (Table 1). This noveltv preference, or dishabitllation effect, ind(- cates that 8-month-olds recoenized the dif-

tion rnaterial and (ii) items that are highly similar but (by some critical criterion) were not contained within the familiarization material. During a series of test trials that immediately follows familiarization, infants control the duration of each test trial by their sustained visual fixation on a blinking light (14). If infants have extracted the crucial information about the familiariza-

u

ference between the novel and the familiar orderings of the three-syllable strings. Thus, 8-month-old infants are capable of extract- ing serial-order information after only 2

posited. The results demonstrate that in- fants uossess ~owerfulmechanisms suited to learning the types of structures exemplified in linguistic systems. Experience may there- fore play a more important role in the ac-

lnin of listening experience. Of course, simple serial-order informa-

tion is an insufficient cue to word bound- quisition of language than existing theories suggest.

One task faced by all language learners is the seglnentation of fluent speech into words. This process is particularly difficult because word boundaries in fluent s ~ e e c h

aries. The learner must also be able to ex- tract the relative freauencies of co-occur-

tion items, they may show differential du- rations of fixation (listening) during the two types of test trials (15). We used this procedure to determine whether infants can acquire the statistical properties of sound sequences from brief exposures.

In our first experiment, 24 8-month-old infants from an American-English language environment were familiarized with 2 min of a continuous speech stream consisting of four three-syllable nonsense words (hereaf- ter, "words") repeated in random order (16 ) . The speech strealn was generated by a speech synthesizer in a monotone fernale voice at a rate of 270 syllables per minute (180 words in total). The synthesizer pro- vided no acoustic information about word

rence of sound pairs, where relatively low transitional probabilities signal word boundaries. Our next experiment examined whether 8-month-olds could oerform the

are marked inconsistently by discrete acous- tic events such as pauses (8). Although it has recently been demonstrated that 8-month-old infants can segment words

more difficult statistical computations re- quired to distinguish words (that is, recur- rent syllable sequences) from syllable strings spanning word boundaries (that is, syllable sequences occurring more rarely). T o take an English example, pretty#baby, we wanted to see if infants can distinguish a word-

- from fluent speech and subsequently recog- nize them when presented in isolation ( 9 ) , it is not clear what information is used by infants to discover word boundaries. This problem is complicated by the variable acoustic structure of speech across different languages, suggestiAg that infants must dis- cover which, if any, acoustic cues correlated with word boundaries are relevant to their

D

internal syllable pair like pretty from a word- external syllable pair like ty#ba.

Another 24 8-month-old infants from an American-English language environ- ment were familiarized with 2 min of a continuous s ~ e e c h strealn consisting of

native language (10); there is no invariant acoustic cue to word boundaries present in all languages.

One important source of information that can, in principle, define word bound-

boundaries, resulting in a continuous strealn of coarticulated consonant-vowel syllables, with no pauses, stress differences, or any other acoustic or prosodic cues to word boundaries. A sample of the speech strealn is the orthographic string bidakupadotigola- bubidaku. . . . The onlv cues to word bound-

u

three-syllable nonsense words similar in structure to the artificial language used in our first experiment (19). This time, how- ever, the test items for each infant consisted of two words and two "vart-words." The part-words were created by joining the final svllable of a word to the first two svllables of

aries in any natural language is the statisti- cal information contained in seauences of sounds. Over a corpus of speech there are measurable statistical regularities that dis-

aries were the transitional probabilities be- tween syllable pairs, which were higher within words (1.0 in all cases, for example, bida) than between words (0.33 in all cases,

another word. Thus, the part-words con- tained three-svllable seauences that the in-"

tinguish recurring sound sequences that comprise words from the more accidental

fant had heaid duringLfamiliarization but statisticallv, over the corvus, did not corre- spond to words (20). These part-words could onlv be iudped as novel if the infants

sound sequences that occur across word boundaries (1 1). Within a language, the transitional probability from one sound to the next will generally be highest when the two sounds follow one another within a word, whereas transitional orobabilities

for example, kupa). T o assess learning, each infant was pre-

sented with repetitions of one of four three- svllable strings on each test trial. Two of

, " - had learned the words with sufficient spec- ificitv and comoleteness that seauences

L.

these three-syllable strings were "words" from the artificial language presented dur- ing familiarization, and two were three-syl- lable "nonwords" that contained the same syllables heard during familiarization but

crosskg a word 'boundary were reiatively ilnfamiliar.

Despite the difficulty of this word versus part-word discrimination, infants showed a significant test-trial discrimination between the word and part-word stimuli (21), with longer listening times for part-words (Table 1 ) . T~ILIS,2 min of exposure to concatenat- ed speech organized into "words" was suffi-

spanning a word boundary will be relatively low (12). For example, glven the sound sequence pretty#baby, the t r a n s ~ t ~ o n a lprob- ability from pre to ty 1s greater than the transitional probability from ty to ba. Pre- viously, we showed that adults and children can use information about transitional

not in the order in which they appeared as words (17).

The infants showed a significant test- trial discrimination between word and non-

probabilities to discover word boundaries in an artificial language corpus of nonsense words presented as continuous speech, with no acoustic cues to word boundaries (13).

We asked whether 8-month-old infants can extract information about word bound-

Table 1. Mean tlme spent llstenlng to the famlllar nonwords) and experlment 2 (words versus pat--wc times.

and novel stimuli for experlment 1 (words versus 3 r d ~ )and signlflcance tests comparing the listening

Mean listening tlmes (s) Experiment Matched-pairs t test

Famlllar ltems Novel ltems aries solely on the basis of the sequential statistics of concatenated s ~ e e c h .We used the famillarization-preference procedure de- veloped by Jusczyk and Aslin ( 9 ) . In this 1 7.97 (SE = 0.41) 8.85 (SE = 0.45) t(23) = 2.3, P < 0.04 2 6.77 (SE = 0.44) 7.60 (SE = 0.42) t(23) = 2.4, P < 0.03 procedure, infants are exposed to auditory

S C I E N C E VOL.274 13 DECEMBER 1996 1927

cient for 8-month-old infants to extract information about the sequential statistics of syllables. Moreover, this novelty prefer- ence cannot be attributed to a total lack of experience with the three-syllable sequenc- es forming part-words, as was the case with the nonwords in the first experiment. Rath- er, infants succeeded in learning and re- membering particular groupings of three- syllable strings-those strings containing higher transitional probabilities surrounded by lower transitional probabilities.

The infants' performance in these stud- ies is particularly impressive given the im- poverished nature of the familiarization speech stream, which contained no pauses, intonational patterns, or any other cues that, in normal speech, probabilistically supplement the sequential statistics inher- ent in the structure of words. Equally im- pressive is the fact that 8-month-old in- fants in both experiments were able to extract information about 'sequential sta- tistics from only 2 min of listening expe- rience. Although experience with speech in the real world is unlikely to be as concentrated as it was in these studies, infants in more n a k ~ r a lsettings presum-

, ably benefit from other types of cues cor- related with statistical information.

Our results raise the intriguing possibil- ity that infants possess experience-depen- dent mechanisms that may be powerful enough to support not only word segmen- tation but also the acquisition of other as- pects of language. It remains unclear wheth- er the statistical learning we observed is indicative of a mechanism specific to lan- guage acqu~sitionor of a general learning mechanism applicable to a broad range of distributional analyses of environmental in- put (22). Regardless, the existence of com- putational abilities that extract structure so rapidly suggests that it is premature to assert a priori how much of the striking knowl- edge base of human infants is primarily a result of expertence-independent mecha- nisms. In particular, some aspects of early development may turn out to be best char- acterized as resulting from innately biased statistical learning mechanisms rather than innate knowledge. If this is the case, then the massive amount of experience gathered by infants during the first postnatal year may play a far greater role in development than has previously been recognized.

REFERENCES AND NOTES

1. Certain species-spec~ficsk~llsdevelop w~thoutany experiential input, nclud~ngbat echolocat~on[E. Gould, Dev. Psychobioi. 8, 33 (197511 and cricket song [R. Hoy, Am. Zool. 14, 1067 (1974)l.

2. Examples of behaviors mediated by early experience are imprint~ng[E. Hess, imprinting (Van Nostrand, New York, 1973); M . Leon, Physioi. Behav. 14, 311

(1975)l and suckl~ngresponses n newborn rats [M H Te~cherand E M Blass, Science 198, 635 (1977)l.

3 These mlestones have been wel-documented both in English [for example, R Brown, A First Language [Harvard Univ. Press, Cambridge, MA, 197311 and cross-ingu~st~cally[for example, E Lenneberg, Bio- logicai Foundations of Language [Wiley, New York, 1967); D. Slob~n,Ed., vols. 1 to 3 of The Crosslin- guistic Study of LanguageAcquisition (Erlbaum,H s - dale, NJ, 1985, 1987, 199211

4. This "argument from the poverty of the stmulus" remains w~delyaccepted [for example, N. Chomsky, Aspects of the Theory of Syntax (MIT Press, Cam- bridge, MA, 1965); S Crain, Behav. Brain Sci. 14, 597 (199111.

5. D. B~ckerton,Behav. Brain Sci 7, 173 (19841, N. Chomsky, Rules and Representations (Columbia Univ, Press, New York, 1981); J. Fodor, Modularity of Mind ( M T Press, Cambrdge, MA, 1983), L. Gleit- man and E. Newport, in Language: An Invitation to Cognitive Science, L Gleitman and M. Liberman, Eds (MIT Press, Cambridge, MA, 19951, pp. 1-24

6. Examples n c u d e vowel structure [P K. Kuhl, K. A W~lliams,F. Lacerda, K N. Stevens, B Lindblom, Science 255, 606 (1992)],phonotactics [P Jusczyk, A. Frieder~ci,J. Wesses, V. Svenkerud, A. Jusczyk, J. Mem. Lang. 32, 401 (1993)],and prosodc struc- ture [P. Jusczyk, A. Cutler, N. Redanz,ChildDev. 64, 675 (199311.

7 Exceptions Include research on prenatal exposure to maternal speech [A. DeCasper, J.-P. Lecanuet, M.-C. Busnel, C. Granier-Deferre, R. Maugeais, in- fant Behav. Dev. 17, 159 (1994)] and early postna- tal preferences [J, Mehler et a / , Cognition 29, 149 (1988)l

8 R. Cole and J. Jakimk, n Perception and Production of Fluent Speech, R Cole, Ed. [Erlbaum, H~lsdale, NJ, 19801, pp 133-163.

9, P Jusczyk and R Asl~n,Cognitive Psycho1 29, 1 ( I995)

10. A. Chr~stophe,E Dupoux, J. Bertoncini, J. Meher, J. Acoust. Soc. Am 95, 1570 (1994); A. Cutler and D Carter, Comput. Speech Lang. 2, 133 (1987)

1I.Z Harris, Language 31, 190 (1955); J Hayes and H. Clark, n Cognition and the Development of Lan- guage, J Hayes, Ed. (Wiley, New York, 1970). See M Brent and T. Cartwright [Cognition 61, 93 [ l 996)l for a dscussion of related statstcal cues to word boundares

12. The transitional probability of frequency of XY

'IX = frequency of x 13. J Saffran, E. Newport, R. Aslin, J Mem, Lang 35,

606 [19 9 6 ) ; ,R Tunck, S Barrueco, Psychol. Sci., n press.

14. Each infant was tested ndvidualy whle seated on the parent's lap in asound-attenuated booth. Syn- thetic speech was generated off-line by the Macin- Talk system and stored on disk at a s a m p n g rate of 22 kHz for on-line playback through an Audio- media board in an Apple Quadra 650 computer. An observer outside the testng booth montored the infant's looking behav~orw ~ t hthe use of a color video system, using a buttonbox connected to the computer to intate trials and score head-turn re- sponses. Both the parent and the observer listened to masking music over headphones to eliminate bias. Dur~ngthe 2 - m n fam~liarization phase, the infant's gaze was f~rstdirected to a bllnkng Ight located on the front wall of the test~ngbooth, and then the sound sequence was presented from two loudspeakers located on the s ~ d ewalls The in- fant's gaze was drected to one of two blinkng Ights on these side walls during fam~liarization,but there was no relat~onbetween lights and sound. Immediately after fam~liarization,12 test trials were presented (six words and SIX nonwords). Each test trial began with the central blinking hght When the observer s~gnaledwith a button press that the n - fant had fixated on the central ight, one of the two side blink~nglights was turned on and the center

light was ext~nguished When the nfant faced the s d e light (a head turn of at least 30" n the directon of the Ight), the three-syllable test strlng was played and repeated until the infant looked away from the light for 2 s or until 15 s of looking had occurred. The observer smply recorded the direc- t ~ o nof the Infant's head turn, and the computer measured o o k n g t~mes,determ~nedwhen the 2-s ookaway crlterlon had been met, and controlled the randomization and presentaton of s t m u l . Cu- mulatve look~ngt m e across each of the two types of test trals provded the measure of preference.

15 The drect~onof the f~xatonpreference depends on the degree of famlarity w t h the stmuli. If the Infants have become h g h y famliar w ~ t hthe stmuli, they show d~shabituat~onbehavior, preferr~ngthe novel stimu.

16. Two counterbalanced stimulus cond~tionswere generated. For each condition, 45 tokens of each of four trisyllabc nonsense words (conditon A: tu- piro, golabu, bidaku, and padoti; condition B: dapiku, tiiado, burobi, and pagotu) were spoken n random order to create a 2-min speech stream, w ~ t hthe s t p u a t o n that the same word never oc- curred t w c e In a row.

17. Test stirnull, tupiro, goiabu, dapiku, and tilado. In condton A, the first two strings were words and the last two strings were nonwords (the trans~t~onal probabllt~esbetween the syllables in the nonwords were all zero relat~veto the exposure corpus, as these syllable palrs had never occurred durng faml- iarizat~on).In cond~t~onB, the first two strings were nonwords and the last two strings were words. Ths between-subjects counterbalanced design ensured that any observed preferences for words or non- words across both conditions would not be artifacts of any general preferencesfor certain syllable strings. Each of the four test strings were presented [repeat- ed with a 500-ms Interval between test strngs) on three different trials, resutng n a total of 12 test trals per Infant.

18 There were no signficant differences between the in- fants n conditon A and conditon B: t(22) = 0 31. The data from the two groups were thus combined for the other analyses.

19 Condit~onA words pabiku, tibudo, goiatu, and da- ropi, condt~onB words, tudaro, pigoia, bikuti, and budopa

20 Test st~muli:pabiku, tibudo, tudaro, and pigoia. In conditon A, the frst two strngs were words and the second two strings were part-words. For ex- ample, the part-word pigola spanned the word boundary between daropi#goiatu and thus was heard dur~ngexposure In cond~tionB, the frst two strlngs were part-words and the second two strlngs were words The part-words were thus three-syllable sequences that the infants had heard during the course of the exposure period. The dif- f~cultyof t h s test dscrminaton can be seen by comparing the trans~tionalprobabillt~esbetween the syllables in the words (1.0 between syllables 1 and 2 and between syllables 2 and 3) to the tran- sitonal probabilites between the syllables n the part-words (0.33between syllables 1 and 2 and 1.O between syllables 2 and 3)

21. There were no signifcant d~fferencesbetween the Infants In conditon A and condtion B: 1122) = 0 49. The data from the two groups were thus combined for the other analyses

22. For example,this same general mechanism could be used to find an object, such as a human face, in the environment.

23 We thank J. Galipeau, J Hooker, P Jusczyk, A. Jusczyk, T. Mntz, K Ruppert, and J. Sawusch for ther help with varlous aspects of t h s research, and P, Jusczyk, S. Pollak, M . Spivey-Knowton, and M. Tanenhaus for their helpful comments on a previous draft Supported by an NSF predoctoral felowsh~p (J.R.S.), NSF grant SBR9421064 (R.N.A.), and NIH grant DC00167 [E.L.N.). The parents of all particl- pants gave Informed consent.

10 May 1996; accepted 30 September I 9 9 6

1928 SCIENCE VOL. 274 13 DECEMBER 1996