DOI: 10.1126/science.271.5245.77 , 77 (1996); 271Science

et al.Michael M. Merzenich,Impaired Children Ameliorated by Training Temporal Processing Deficits of Language-Learning

www.sciencemag.org (this information is current as of January 31, 2007 ):The following resources related to this article are available online at

http://www.sciencemag.orgversion of this article at:

including high-resolution figures, can be found in the onlineUpdated information and services,

http://www.sciencemag.org#otherarticles 33 articles hosted by HighWire Press; see: cited byThis article has been

http://www.sciencemag.org/cgi/collection/neuroscienceNeuroscience

: subject collectionsThis article appears in the following

http://www.sciencemag.org/help/about/permissions.dtl in whole or in part can be found at: this article

permission to reproduce of this article or about obtaining reprintsInformation about obtaining

registered trademark of AAAS. c 1996 by the American Association for the Advancement of Science; all rights reserved. The title SCIENCE is a

CopyrightAmerican Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005. Science (print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last week in December, by the

on

Janu

ary

31, 2

007

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om

refinements were performed with X-PLOR (27).20. W. I. Weis, K. Drickamer, W. A. Hendrickson, Nature

360,127 (1992); W. I. Weis and K. Drickamer, Struc-ture 2, 1227 (1994).

21. Multiple conformers [J. Kuriyan et al., Proteins 10,340 (1991)] (9) of the sub-MBP-A protein moleculeswere generated starting with the model that wasobtained after single-conformer refinement (19). Thewater molecules and Yb were restrained to theirinitial positions. Conformers of the protein were ren-dered invisible to each other in terms of chemicalinteractions. Each conformer in the ensemble con-tributed equally to the calculated structure factorwith the occupancy set to the reciprocal of the num-ber of conformers. Evaluation of the free R value as afunction of the number of conformers produced ashallow minimum for eight conformers, which wereused for subsequent analysis.

22. Vector difference electron density maps were com-puted by taking the vector difference between com-plex structure factors [I FobS exp(i40bs) - lFcaicIexp(i4()lJ)], the experimental structure factor withcentroid phases obtained from PA4() (17) minus that

computed from the refined single-conformer model. Aconventional (model-phased) difference electron den-sity map consisting of a Fourier synthesis of (IF,,bJ -IFcalcl) exp(i4ca1,) is an approximation of the vectordifference map [J. Drenth, Principles of Protein X-rayCrystallography (Springer-Verlag, New York, 1994),pp. 150-152]. As a consequence, vector differencemaps of sub-MBP-A were observed to have lessnoise than conventional difference maps.

23. M. M. Teeter, Proc. Natl. Acad. Sci. U.S.A. 81, 6014(1984).

24. G. M. Clore et al., Structure 2, 89 (1994).25. M. Levitt and R. Sharon, Proc. Natl. Acad. Sci.

U.S.A. 85,7557 (1988); Y. Komeijietal., Proteins 16,268 (1993); M. Karplus and G. A. Petsko, Nature347, 631 (1990); P. J. Steinbach and B. R. Brooks,Proc. Natl. Acad. Sci. U.S.A. 90, 9135 (1993); V.Lounnas et al., Biophys. J. 66, 601 (1994).

26. D. C. Rees and M. Lewis, Acta Crystallogr. A39, 94(1983); E. Arnold and M. G. Rossmann, ibid. A44,270 (1988); L. M. Rice and A. T. Brunger, ProteinsStruct. Funct. Genet. 19, 277 (1994).

27. A. T. Brunger, X-PLOR. Version 3.1. A System for

X-ray Crystallography and NMR (Yale Univ. Press,New Haven, CT, 1992); developmental version ofX-PLOR, available on request.

28. D. T. Cromer, J. Appl. Crystallogr. 16, 437 (1983);and D. L. Liberman, J. Chem. Phys. 53,1891

(1970).29. W. A. Hendrickson and W. I. Weis, unpublished data.30. L. M. Gregoret, S. D. Rader, R. J. Fletterick, F. E.

Cohen, Proteins Struct. Funct. Genet. 9, 99 (1991).31. We thank P. Gros for assistance in designing the

X-PLOR crystallographic language for phasing; V.Agrawal, S. Shah, and M. Willis for help with datacollection; C. Ogata for beamline support; and P.Adams, M. Gerstein, M. Levitt, and L. Rice for criticalreading of the manuscript. The coordinates and thediffraction data have been deposited with theBrookhaven Data Bank (accession number 1 YTT).Supported by the Howard Hughes Medical Institute(A.T.B.), NSF (DIR9021975, A.T.B.), and NIH(GM50565, W.I.W.). W.I.W. is a Pew Scholar in theBiomedical Sciences.

8 September 1995; accepted 1 November 1995

Temporal Processing Deficits ofLanguage-Learning Impaired Children

Ameliorated by TrainingMichael M. Merzenich,* William M. Jenkins, Paul Johnston,

Christoph Schreiner, Steven L. Miller, Paula Tallal

Children with language-based learning impairments (LLIs) have major deficits in theirrecognition of some rapidly successive phonetic elements and nonspeech sound stimuli.In the current study, LLI children were engaged in adaptive training exercises mountedas computer "games" designed to drive improvements in their "temporal processing"skills. With 8 to 16 hours of training during a 20-day period, LLI children improved markedlyin their abilities to recognize brief and fast sequences of nonspeech and speech stimuli.

Experiments conducted in human (1) andmonkey (2) neurological models of percep-

tual learning have demonstrated that thecapacity for segmentation of successiveevents in sensory inpLut streams can besharpened, apparently throughout life, bypractice. Electrophysiological studies oflearning-induced plasticity conducted inthe neocortices of monkeys have provided a

growing body of evidence about the neuralprocesses that underlie practice-based im-provements in both temporal segmentationand spectral (spatial) discrimination perfor-mances (3-5). These studies have shownthat the ability of an adult animal to makefine distinctions about the temporal or

spectral features of complex inputs can besharply improved, or degraded, by a periodof intensive behavioral training (2-4, 6).

In parallel with animal and human mod-

M. M. Merzenich, W. M. Jenkins, P. Johnston, C. Schrei-ner, W. M. Keck Center for Integrative Neurosciencesand Coleman Laboratory, University of California, SanFrancisco, CA 94143-0732, USA.S. L. Miller and P. Tallal, Center for Molecular and Behav-ioral Neuroscience, Rutgers University, 197 UniversityAvenue, Newark, NJ 07102, USA.

*To whom correspondence should be addressed.

els of temporal sequence learning, earlierstudies of the receptive deficits of LLI chil-dren (7) had shown that they have a "tem-poral processing deficit" expressed by limit-ed abilities at identifying some brief pho-netic elements presented in specific speechcontexts and by poor performances at iden-tifying or sequencing short-duration acous-

tic stimuli presented in rapid succession (8).Consistent with a temporal processing def-icit hypothesis, LLI children can distinguishthese brief speech features and can correctlyreconstruct stimulus input sequences ifstimuli are presented in slower forms or at

slower event rates (8).Taken together, these experimental

findings led us to hypothesize that the def-icits underlying the phonetic reception lim-itations of a LLI child might arise in earlylife as a consequence of abnormal percep-

tual learning that then contributes to ab-normal language learning (3, 4). For exam-

ple, a child might simply make more-limit-ed-than-normal use of temporal informa-tion as he or she learns to make distinctionsabout speech inputs as a learning alterna-tive, or a child might generate a represen-tation of phonetic information under early

SCIENCE * VOL. 271 * 5 JANUARY 1996

childhood conditions (for example, in thepresence of middle ear disease) under whichacoustic inputs are consistently muffled.Cortical plasticity studies indicate that tem-poral processing deficits like those thatemerge in LLI children would be expectedfrom these learning scenarios, if they wereto be undertaken in a training regime ap-plied in a monkey model (3, 4).

Visual psychophysical studies have al-ready shown that very great improvementsin the recognition of brief, successively pre-sented stimuli can be achieved with prac-tice in adult humans (1). In this study, weasked: Can we substantially alter the defi-cient temporal processing capacities ofyoung, school-age LLI children by similaracoustic practice?

For training tools, we produced two au-diovisual (AV) "games" designed around acircus theme to engage 5- to 10-year-oldchildren at high levels of attention andenthusiasm within highly repetitive learn-ing tasks. The first AV game (Circus Se-quence) was a perceptual identification taskin which a correct response in the exercisewas a faithful reproduction of the order oftwo-stimuli sound sequences by touch-screen button-press sequences (9). Thenonverbal stimuli applied in training were16 octave-per-second upward- or down-ward-gliding (U and D, respectively) fre-quency-modulated (FM) tonal pairs (U-U,U-D, D-U, or D-D). Stimuli in each FMpair swept across the same frequency range.These stimuli were in the range of sweepfrequencies and speeds that apply for theformant transitions of English consonantsthat LLI children have difficulty recogniz-ing (7, 8). The interstimulus intervals(ISIs) and FM frequencies were adaptiveparameters.

The second game was a phonetic ele-ment recognition exercise implemented as atwo-alternative forced-choice task in whichthe child was presented two consonant-

77

on

Janu

ary

31, 2

007

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om

vowel (CV) stimuli with contrasting con-sonants (for example, [be] versus [dcl pre-sented in rapid sequence) (10). The child'stask was to identify the sequence position ofa target CV (in this example, of either aprecued [be_] or [dei). The main variables inthis exercise were (i) the durations of syn-thetically produced consonants (and recip-rocally, of the following vowels; the totalCV duration was constant), (ii) the magni-tudes of a 0- to +20-dB amplification ofconsonant elements versus vowel intensity,and (iii) the ISIs between presented CVpairs.

These two games were begun with stim-uli that LLI children could easily distin-guish and recognize: They had long nonver-bal stimulus (60 ms) or consonant transi-tion (65 to 70 ms) durations, were present-ed with long ISIs (500 ins), and for CVstimuli, were presented with a maximumdifferential amplification of the consonanttransition (+20 dB). These variables werethen altered adaptively in training, trial bytrial, to drive each child in the direction ofnormal performance levels.

Correct performances at both of theseexercises were rewarded in several ways.The children received feedback about their

0-

100 -

200 -

300 -

400

E 500'a 0-it0 100-c

200-

E5 300-E*,, 400-0W 500-C

u -

100 -

200

300

400 -

CM

Late

X Early

FM 0.5 kHz

LateEarly

FM 1.0 kHz

.,Late

Early

FM 2.0 kHz

1 5 10 15 20Trial block

Fig. 1. Game performance functions at the se-quence recognition game for three of the fourstimulus sets (9) for one LLI child, recorded nearthe beginning (open dots) and in the final week(filled dots) of 4 weeks of testing (study 1). Allchildren in both studies who worked under re-sponse control at these games showed substan-tial, progressive gains in their abilities to sequencethese fast, brief stimuli. The symbols representchanges in performance level; for positive chang-es in level (decreasing ISIs) a dot may representfrom one to three correct responses. After eightreversals in difficulty level, the child progressed totraining with another FM stimulus set.

78

trial performances by an audio and visualsignaling of a correct response and by apoint accumulator that advanced for "hits"but not "misses." Entertaining reward ani-mations were presented at frequentlyachieved performance benchmarks. Chil-dren were also rewarded for their progressby working in a token economy in whichearned points could be traded for prizes.

The first trial of these temporal process-ing games was conducted with seven 5.9- to9.1 -year-old LLI children (I I). Training atthe nonverbal Circus Sequence temporaltraining exercise was applied for 19 to 28training sessions of 20 min each conductedover a 4-week training period. Five of theseLLI children benefited from the reinforce-ment contingencies used in the game andprogressed regularly for extended periods oftraining across the daily sessions of adaptivetraining. For the four children who wereunder the best psychophysical control, per-formances over successive daily practice ses-sions asymptoted at progressively higherperformance levels (Fig. 1). Two of thesefour children reached or exceeded normalperformance levels at this task; that is, afteradaptive training they could consistentlyorder stimuli that were presented in imme-diate or nearly immediate succession. Notethat normal children can reliably identify75-ms-duration stimuli that are separatedby -10 ms ( 12). Three other children madeseveralfold improvements in task perfor-mance with practice, operating with ISIsat or below 100 ms for some or all FMstimulus categories. One child never dis-played a consistent response pattern dur-ing this exercise and never progressed toISIs shorter than -300 ms. Another childshowed a clear progression in performance,but that performance achievement was lim-ited to stimuli separated by -200 ms for allstimulus categories.A benchmark test of temporal process-

ing ability, the Tallal Repetition Test (12),was conducted before and after training. Itrevealed statistically significant improve-

ments in the temporal event recognitionand sequencing abilities for tonal stimulicentered at 1.2 kHz, for all children (study1 in Fig. 2). Specifically, there were signif-icant performance improvements in thesechildren's abilities to sequence both stimuliof shorter durations and stimuli separatedby shorter ISIs, although all training for thelatter was conducted with stimuli of a fixed(60-ms) duration.

For the Phoneme Identification game,all children in study 1 performed at near-chance levels when they attempted to iden-tify stop consonants presented with briefformant transitions in the initial sessions.All reliably identified the same brief conso-nant transitions with a high performancereliability in some later sessions. A bench-mark test of phonemic reception (13) con-ducted before and after training showedthat six of seven children made significantimprovements at phonetic element identi-fication. Their average gain translated to-1.5 years in language development age.Because stimulus delivery in this trainingexercise was relatively slow and because thestimulus categories were not changed in afixed pattern, performance improvementsat this game were somewhat erratic andwere not easily compared between children.A second study was conducted to deter-

mine if these results could be replicated in alarger, independent sample of LLI children(14). The games were modified in an at-tempt to increase performance consistencyand to more strongly maintain the atten-tion of LLI children on these tasks (15).Children progressed at the task of identify-ing rapidly successive stimuli (the CircusSequence game) more reliably and underbetter response control than in the initialgame version, with 10 of 11 children sys-tematically improving their performancesover the majority of time spent at this ex-ercise. With extended stimulus sets thatincluded stimulus duration as a parameterand with more reliable on-task behavior,LLI children were also driven to higher

Fig. 2. Performance of A,,A Bstudy 1 and study 2 children Ebefore and immediately after - 500 Xi 120the conclusion of training, as 0 400 M2100-evaluated with a benchmark . 300 80measure of temporal se- 60quence recognition ability, E200 40E 7the Tallal Repetition Test. 100 E 20Bars are standard errors. -- .0 I 0The Tallal Repetition Test E Study 1 Study 2 Study 1 Study 2

measures (A) threshold ISIs ES Pretraining LI Posttrainingat which the sequence or-ders of rapidly sequenced tonal stimuli pairs of 0.8 and 1.2 kHz are correctly identified with 0.8 probability,and it also measures (B) the minimum durations of stimuli that can be sequenced [see (12)]. The effectsof training were significant for all pre- and posttraining comparisons. For study 1, pretraining versusposttraining threshold ISIs, ANOVA (repeated measures, two-tailed) F(1,6) = 36.7, P < 0.001; for study2 ISIs, F(1 ,6) = 12.8, P < 0.05. For study 1 durations, F(1 ,6) = 7.3, P < 0.05; for study 2 durations, F(1 ,6)= 28.9, P < 0.01.

SCIENCE * VOL. 271 * 5 JANUARY 1996

on

Janu

ary

31, 2

007

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om

performance levels in this revised game.Despite a highly variable number of totaltrials for different children (640 to 3739 in17 to 22, 20-min training sessions), 10 of 11children showed progressive gains and wereable to complete a substantial program oflearcateprofgait

FMingage(seea petask(9)trait

em

- c00

.2-rzInle

Fia.

ms. Three children also progressed to aperfect (ISI = 0) performance level whenthe duration of the FM stimuli was loweredto 40 ms, and one child mastered the taskfor 20-ms-duration FM stimuli in this cate-gory. Differences between early and late

ning within at least two FM stimulus training performance levels were statistical--gories; 9 of 11 completed a substantial ly significant [ANOVA (analysis of vari-gram of learning and showed progressive ance) repeated measures, two-tailed:is in all four categories (Fig. 3). F(1,18) = 54, P < 0.0001; see Fig. 3].For example, for distinguishing between Although some individual performance dif-stimuli of 60-ms duration with a start- ferences by category were recorded, overall,or ending frequency of 1 kHz, the aver- performance gains did not significantly varyperformance improved about fourfold by FM stimulus category (16).Fig. 3A), with four children achieving Again, the performance gains measuredrfect (ISI = 0) performance level at this in these training experiments were mirrored

<. The initial reversal-defined threshold by results obtained before and after trainingaveraged 268 ms for these children; with on the benchmark Tallal Repetition Testning, it dropped to an average ISI of 77 (12). All 11 children showed gains by this

performance measure (see study 2, Fig. 2).A 1+ kHz FM The pre- and posttraining group differences300A were again significant. On the average,

2+ kHz FM minimal ISIs dropped to -1/15 of their200 pretraining values. The mean ISI threshold

of about 20 ms achieved after training ap-proaches normal performance abilities (8).

100- Similarly, trained children could identifymuch briefer stimulus tones presented in

0 almost immediate succession (Fig. 2). AfterEarly Late training, children could distinguish the se-

B 5quence order of almost immediately succes-

sive tonal stimuli that were, on the average,20 E 18 ms in duration, which was about one-

)o - / ^4 C fifth the pretraining value. This perfor-a0iance ability again approaches that of nor-

Y 40 mal children (8). These training effects on- the ability of study 2 LLI children to iden-2 tify the sequence order of briefer sound

20- 5 P = <6 6 60 3 stimuli was significantly greater than thoseO 3~~~~~~~~~for stLidy I children, possibly because they

0 5 10 15 20 were trained with stimuli that varied inTraining session duration. In contrast, 11 matched LLI chil-

3. (A) Avprqnp nPrfnrm-nr: cdiffPrPmps, nn dren who were undergoing classical lan-iFu. %J. ~PjrNA VC;0I cp8uF uiikJ a1l'1 ut111 1 u1 1 U;: VI

the Circus Sequence game for the 11 children instudy 2, recorded early and late in a 20-day (20min/day) training period. "Early" scores werefrom the first trials in which there was a clear dem-onstration that the children understood the "rulesof the game" (that is, played the game on-task)and reached a stable response level revealed bystimulus reversals within a training session. Theseinitial measures overestimate the early perfor-mance abilities of these children in some casesbecause children often practiced these games fora session or two before any clear stimulus rever-sals indicating stable performance limits were re-corded. (B) Progression in performance at the 1-or 2-kHz FM stimulus tasks for 6 of the 1 1 childrenin the second experimental series performing thestimulus sequence recognition (Circus Sequencegame) exercise (9). The highest difficulty levelsreached on every trial day on which childrenworked at these particular stimulus sets areshown. Similar progressive learning functionswere recorded for nine other children in these twostudies (that is, in 15 of 18 trained LLI children).More difficult stimuli had progressively shorter ISIsand durations.

guage training and equivalent periods ofnonadaptive video game playing per dayover the 20-day training period (7) did notsignificantly improve in their sequenced-stimulus recognition abilities, as measuredby this benchmark test (17).

The phoneme recognition game was alsorevised for testing in the 11 study 2 LLIchildren. Six of the 11 children progressedto stimulus categories for several CV con-trasts at which they could identify the brief-est consonants (35 ms) at the fastest ISIs(10 ms) (see examples in Fig. 4). Two otherchildren also learned to reliably identify thefastest forms of natural (nonamplified) con-sonants. Two others showed significant im-provements but, with this limited training,did not achieve normal recognition abilitiesfor CV stimuli in which consonants werepresented in very fast and nonenhancedforms. The 11th child, again, never consis-tently performed above chance at thisgame. Notably, performance gains wereagain mirrored by improvements in a pho-netic element recognition benchmark test(13). Analyses of results from these testsindicated that with training at these exer-cises supplemented (i) by two other gamesdesigned to promote speech perception gen-eralization (18) and (ii) by special languagetraining with acoustically modified speech(7), temporal processing improvements ap-plied generally to these children's languagecomprehension abilities [see (7)].

It is likely that performance at theseadaptive training tasks could be further im-proved by additional practice in almost ev-ery tested LLI child (19). The five childrenwho had the lowest performance resultsplayed the game for the least number ofsessions and trials. Moreover, the number oftraining sessions and training trials com-pleted by different children were directly

Fig. 4. Performance levels Perfectachieved early and late in

P*4 performance

training in five LLI children - Decreasing SIworking at the Phoneme 0 20

DecreasingIdentification game. Thecosntbottom of each colored bar 5 emphasismarks a stable performance a 1 Decreasinglevel achieved during the 0 consonantfirst five to six training days; duratothe top of each colored bar 0 Decreasing SI

marks the performance level 1 2 3 4 5achieved during the final Subjectdays of training. Note that in [ba] vs. [da] - [aba] vs. [ada]this exercise, training began - [be] vs. [de] - [bai] vs. [dai]with differentially prolonged - [fa] vs. [va]and amplified consonantsdelivered initially in CV stimuli presented at slow repetition rates. Every child could initially identify thesesynthetically disambiguated CV stimuli. As the child's performance at the task improved with training, taskdifficulty was progressively increased (see right margin of the graph). The ISI was first decreased in1 00-ms steps to initially stabilize at 200 ms; then the consonant duration was decreased progressively tonatural fast transition rates (35- to 40-ms durations) in seven steps; then the differential consonant versusvowel amplification (consonant "emphasis") was faded in seven steps; then the ISI between CVs werefurther incrementally decreased in eight nonlinear steps to an ISI of 10 ms.

SCIENCE * VOL. 271 * 5 JANUARY 1996

413...

79

on

Janu

ary

31, 2

007

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om

correlated with a benclhmnark measure oflanguLage outcome (20) for both study 1 [n

7; r (correlation coefficient) = 0.85; P0.011 and study 2 LLI clildren (n = 1 1; r =0.73; P - 0.01) (21).

These experiments once again confirmthat LLI children have major temporal pro-cessing, fast-speech-element recognitiondeficits (8). Those deficits have presumablybeen in place for most of the prior years ofthe speech reception histories of these chil-dren (22). With a total of 5 to 10 hours ofintensive practice at each of these two tasksin brief daily sessions extended over only 20days, a substantial remediation of these def-icits was achieved in nearly all the LLIchildren studied. These studies strongly in-dicate that the fundamental temporal pro-cessing deficits of LLI children can be over-come by training.

These studies also strongly indicate to usthat there may be no fundamental defect inthe learning machinery in most of thesechildren, because they so rapidly learn thesame skills at which they hacve been definedto be deficient. That finding suggests inturn that the physical differences and dis-tributed fLnctional response differences re-vealed in evoked potential and imagingstudies of the brains of LLI individuals (23)may substantially be effects of the learninghistories of these special children. Further-more, it may also imply that inherited fac-tors contributing to LLI origin (24) mayrelate to the initiation of a scenario thatembeds, through learning, a defective rep-resentation of speech phonetics-and doesnot necessarily mean that these childrenhave irreversible defects in the molecularand cellular elements of the learning ma-chinery of their brains.

REFERENCES AND NOTES

1. A. Karni and D. Sagi, Proc. Natl. Acad. Sci. U.S.A.88, 4966 (1991); Nature 365, 250 (1993); M. Ahissarand S. Hochstein, Proc. Natl. Acad. Sci. U.S.A. 90,5718 (1993).

2. G. H. Recanzone, M. M. Merzenich, C. E. Schreiner,J. Neurophysiol. 67, 1071 (1992); X. Wang, M. M.Merzenich, K. Sameshima, W. M. Jenkins, Nature378, 71 (1995); G. Bertini, A. Karni, P. De Weerd, R.Desimone, L. G. Ungerleider, Neurol. Abstr. 21, 276(1995); R. Beitel et al., ibid., p. 1 180.

3. M. M. Merzenich, C. S. Schreiner, W. M. Jenkins, X.Wang, in Temporal Information Processing in theNervous System: Special Reference to Dyslexia andDysphasia, P. Tallal, A. M. Galaburda, R. R. Llinms, C.von Euler, Eds. (New York Academy of Sciences,New York, 1993), pp. 1-22.

4. M. M. Merzenich and W. M. Jenkins, in MaturationalWindows and Adult Cortical Plasticity, B. Julesz andI. Kovacs, Eds. (Addison-Wesley, New York, 1995),pp. 247-272.

5. For reviews of this complex subject, see M. M. Mer-zenich and W. M. Jenkins, in Memory Concepts, P.Anderson, 0. Hvalby, 0. Paulsen, B. Hokfelt, Eds.(Elsevier, Amsterdam, 1994), pp. 437-451; M. M.Merzenich and C. DeCharms, in The Mind BrainContinuum, R. Llinms and P. Churchland, Eds. (MITPress, Cambridge, in press).

6. M. M. Merzenich, G. M. Recanzone, W. M. Jenkins,K. A. Grajski, Cold Spring Harbor Symp. Quant. Biol.

80

55, 873 (1990); N. N. Byl, M. M. Merzenich, W. M.Jenkins, J. Neurol., in press.

7. P. Tallal et al., Science 271, 81 (1996).8. P. Tallal and M. Piercy, Nature 241, 468 (1973); Neu-

ropsychologia 13, 69 (1975); P. Tallal, Brain Lang. 9,182 (1980); and R. E. Stark, Ann. Dyslexia32,163 (1982); , E. D. Mellits, Brain Lang. 25,314 (1985); P. Tallal, S. Miller, R. H. Fitch, in (3), pp.27-47. It has been argued that the fast-element(usually consonant) phonetic recognition difficultyarises because of an inappropriate "integration"with or "masking" by following or preceding speechelements (usually vowels). Although parametric psy-chophysical studies are lacking, this may apply onlywithin frequency channels as these children appar-ently have less difficulty identifying consonant-vowelpairs with brief consonants in which consonant tran-sitions and vowel formants are within largely non-overlapping channels.

9. Cortical plasticity and learning studies conducted inprimate and human models had shown that suchtraining (i) had to be applied with a heavy schedule ofpractice trials, (ii) would be ideally conducted on aseries of successive training days, (iii) would requirerelatively intense practice schedules designed todrive continuous performance improvements, and(iv) would have to be conducted under conditions ofhigh motivational drive. We chose to accomplishthese training objectives in LLI children by using en-tertaining and highly rewarding CD-ROM-mountedexercises disguised as games. The Circus Sequencegame was developed with Authorware Professionaland Director (Macromedia) software. The FM stimuliwere generated with a 22.05-kHz sampling rate and16-bit processor in Audio Interchange File Format(AIFF) with Matlab (Mathworks) software. Stimuliwere ramped on and off to reduce spectral spatter.The compact disks were produced with Sony Hybridsoftware and a Sony 900 CDR. Games were playedon Macintosh computers with CD-ROM drives orwith these exercises copied and played off of Macin-tosh hard disks for convenience. Children receivedsound stimuli through Sony model MDR-V600 head-sets. Four stimulus sets were mounted in this gamein version 1: 60-ms-duration FM sweeps with start-ing or ending frequencies at 0.5, 1 .0, 2.0, or 4.0 kHz.Three successive correct responses resulted in ashortening of either the ISI or the stimulus duration.An error resulted in a one-step lengthening of the ISIor stimulus duration. This learning schedule assuredthat at least 79.3% of the child's responses werecorrect. When the child achieved a predeterminednumber of positive and negative task difficulty rever-sals in any single training session, he or she wasadvanced to a new FM stimulus set. Training wasinitiated at a new stimulus set at a performance levelestablished by a prior session's performanceachievements.

10. The Phoneme Identification game was created withAuthorware Professional (Macromedia) software.Synthetic speech stimuli were produced with 16-bitprocessing and with a 22.05-kHz sampling rate byusing SenSyn (Sensimetrics) software, based on aKlatt cascade-parallel formant synthesizer [D. Klatt,J. Acoust. Soc. Am. 67, 971 (1980)]. Children instudy 1 were trained with [bh], [dE], and [ge] targetand foil stimuli. Children in study 2 were trained withfive CV pairs: [ba] versus [da], [bh] versus [dr], [fa]versus [va], [aba] versus [ada], and [bai] versus [dai].In the initially tested version of this game, childrenperformed trials under given parametric conditions intraining blocks of 10. When performance criteriawere met, consonant durations were shortened ordifferential amplification of fast consonant elementswas reduced (or both) for another set of CVs throughthe next trial block. In study 2, an adaptive staircasetraining procedure identical to that used in the CircusSequence game (9) was used.

11. Seven children (four females, three males) ranging inage from 5.8 to 9.1 years (mean age = 7.3 years,standard deviation = 1.5 years) who had a meannonverbal intelligence score of 106 (standard devia-tion = 18.25) participated in the study. The groupcomprised four whites, two Hispanics, and one Afri-can American. The children were from lower andmiddle socioeconomic status (SES) families. All chil-

dren demonstrated a severe delay in receptive andexpressive language development (mean languageage = 4.8 years) as well as marked temporal process-ing deficits. These school-age children also had read-ing deficits. All were without other primary deficits.

12. The Tallal Repetition Test determines the thresholdISI at which sequences of two tonal stimuli (in thiscase, of 1000 and 1400 or 800 and 1200 Hz) that are150, 75, 40, or 17 ms in duration are perceived withtheir delivery sequence reproduced with a 75% ac-curacy. ISIs vary in the test from 500 to 0 ms. See P.Tallal, in Non-Speech Language and Communica-tion, R. Schiefelbusch, Ed. (University Park Press,Baltimore, MD, 1980), pp. 449-467.

13. The Goldman-Fistoe-Woodcock Diagnostic Audito-ry Discrimination Test (American Guidance Service,Circle Pines, MN) was used as a standard bench-mark. It was designed to define an individual's abilityto identify phonic elements within words.

14. In study 2, 22 children (8 females, 14 males) rangingin age from 5.2 to 10.0 years (mean age = 7.4 years,standard deviation = 1.4 years) who had a meannonverbal intelligence score of 96.4 (standard devi-ation = 9.7) participated. The group included 18whites, two Hispanics, one Asian, and one AfricanAmerican. All children were from middle SES fami-lies. All children demonstrated a severe delay in re-ceptive and expressive language development(mean language age = 4.9 years; standard deviation= 1.4 years) as well as marked temporal processingdeficits. These school-age children also had readingdeficits. All were without other primary deficits.

15. After initial testing with these training exercises inthese seven children, both games were revised, thenretested in study 2 in 11 LLI children (14). In thesecond version of the Circus Sequence game, thenumber of stimulus variations in each set was ex-tended to 135 by including FM stimuli with durationsof 60, 40, and 20 ms. An animated performancebarometer was added to the game to further signalperformance progress in yet another compelling vi-sual manner, and points were subtracted for re-sponse "misses." To further ensure that childrenwere kept on task in these games, five misses in arow resulted in a suspension of a change in difficultylevel; task difficulty was then maintained constantuntil the child again recorded four "hits" in a row. Forthe Phoneme Identification game, an animated per-formance barometer was also added to the AV dis-plays, and points were subtracted for incorrect re-sponses. Most importantly, the game was recon-structed as a progressively adaptive exercise inwhich task difficulty was (i) increased progressivelywhen any three-trial block was completed withouterror, (ii) decreased by one step in difficulty for anyerror, and with the drop in game difficulty, (iii) haltedwhenever a child made five errors in succession.Task difficulty was increased by first reducing theduration of the consonant stimulus elements, then adifferential amplification of fast consonant elementsprogressively faded, then interstimulus ISIs for suc-cessive CVs were progressively reduced.

16. For the main effect of stimulus category, F(1 ,1 8)1.7, not significant (n.s.); for the interaction of timeand stimulus category, F(1 1 8) = 2.5, n.s.

17. Group B children, who received equivalent languagetraining but with natural speech materials and whoplayed video games rather than these adaptive au-ditory-speech training games (7), did not improvesignificantly with respect to either ISI or durationthresholds measured by the Tallal Repetition Test.For ISI, ANOVA (repeated measures, two-tail) F(1,7)= 2.8, n.s.; for duration, F(1 ,7) = 3.9, n.s. In contrastto the experimental treatment group A, in which ev-

ery child improved at this benchmark, the majority ofgroup B children had equal or poorer performanceson this test after their 1-month-long training period.Although the mean performance of group B LLI chil-dren was modestly better than was that for group Achildren, children in these groups did not differ intheir pretreatment Tallal Repetition Test measures ofthreshold ISIs or durations.

18. Children in study 2 were also trained at two addition-al games, both designed to facilitate the generaliza-tion of training gains from these first two describedgames to the wider range of temporal sequence

SCIENCE * VOL. 271 * 5 JANUARY 1996

on

Janu

ary

31, 2

007

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om

4 S1:s

events and phonetic element contexts and contraststhat occur in natural running speech. The third game(Old McDonald's Flying Farm), produced with Direc-tor (Macromedia)software, was a limited hold reac-tion time task in which the child maintained a touch-screen "button" press while repeated stimuli weredelivered in regular sequence. The child's task wasto release the button when there was a change inphonetic element identity. The durations ot a widerarray of synthetic consonant elements and the inter-stimulus times between repeated stimuli were themain exercise variables. The fourth game (PhonicMatch), also developed with Director (Macromedia)software, was a sound-matching exercise in whichbutton presses resulted in soundings that the childhad to locate a match for, on a 2-by-2 to 5-by-5touch-screen button array. The button array size andthe temporal structure of elements and of elementsequences in individual consonant-vowel-consonantstimuli were game variables. Stimuli applied in thisexercise were synthetically processed to prolong anddifferentially amplify brief phonetic elements [see (1)].Children also played both of these games for approx-imately 2D mmn/day throughout the 2D-day trainingperiod. In general, children's performances at thesetwo games paralleled their progressive achievementsat the time order judgment and phonetic elementrecognition tasks described in this report. All LLI chil-dren who were trained at these games also under-went training with acoustically modified speech stim-uli, as described by Tallaletal. (7).

19. Children were still improving at their game perfor-mances when these exercises were arbitrarily ter-minated at the end of the 4-week training period.Their ultimately achievable performance limits areunknown.

20. The Token Test for Children (Teaching ResourcesCorporation, Boston, MA, copyrighted 1978) is de-signed to test the ability to follow auditory commandsof increasing length and grammatical complexity.

21. The intensity of practice at three FM stimulus catego-ries were all significantly correlated with Token Test(language outcome) results. For the 1+ kHz catego-ry, trial numbers versus language outcome, r= 0.75,P -_ 0.01; for 2+ kHz FM stimulus, trial numbersversus language outcome, r= 0.73, P .0.01; for4+kHz FM stimulus, trial numbers versus language out-come, r= 0.84, P< 0.01. The 0.5+ kHz categorypractice trial numbers were not significantly correlat-ed with language outcomes(r-- 0.48).

22. A. A. Benescish and P. Tallal, in (3), pp. 312-314;Infant Behav.Dev., in press.

23. A. M. Galaburda,G. F. Sherman,G. D. Rosen, F.Aboitiz, N. Geschwind, Ann. Neurol. 18, 222 (1985);J. P. Larson, T. Hoien,I. Lundberg, H. Odegaard,Brain Lang. 39, 289 (1990); T. L. Jernigan, J. R.Hesselink, E. Sowell, P. A. Tallal, Arch. Neurol. 48,539(1991); J. M. Flynn, W. Deering, M. Goldstein, M.H. Rahbar, J. Learn. Disabil. 25, 133 (1992); R. Du-ara et al., Arch. Neurol. 48, 410 (1991); J. 0. Hag-man, F. Wood, M. S. Buchsbaum, L. Flowers, W.Katz, ibid. 49, 734(1992); A. Kusch etal., Neuropsy-chologia 31, 811 (1993); A. M. Galaburda, Neurol.Clin. 11, 161 (1993); and M. Livingstone,Ann. N. Y. Acad. Sci. 682, 70 (1993); R. Naas, Curr.Opin. Neurol. 7,179 (1994).

24. P. Tallal, J. Townsend, S. Curtiss, B. Wulfeck, BrainLang. 41, 81 (1991); B. A. Lewis, J. Learn. Disabil.25, 486 (1992); N. J. Cox, P. J. Bayard,Behav. Genet. 23, 291 (1993); B. F. Pennington, J.Child Neurol. 10, S69 (1995).

25. We thank T. Jacobson, B. Wright, X. Wang,G. Bedi,andG. Byma for their technical assistance, and C.Checko, N. Reid, and A. Lipski for assistance inprogramming the animation reward sequences forthese AV exercises. T. Realpe, I. Shell, C. Kapelyn, A.Katz-Nelson, L. Brzustowicz, C. Brown, A. Khoury,J. Reitzel, K. Masters, B. Glazewski, A. Rubenstein,and S. Shapeck assisted in the training of thesechildren at Rutgers University. This research wasfunded by the Charles A. Dana Foundation with sup-portive assistance by Hearing Research, Incorporat-ed. For further information about this and relatedsubjects, contact: http://www.ld.ucsf.edu/

6 October 1995; accepted 30 November 1995

Language Comprehension in Language-LearningImpaired Children Improved with Acoustically

Modified SpeechPaula Tallal,* Steve L. Miller, Gail Bedi, Gary Byma,

Xiaoqin Wang, Srikantan S. Nagarajan, Christoph Schreiner,William M. Jenkins, Michael M. Merzenich

A speech processing algorithm was developed to create more salient versions of therapidly changing elements in the acoustic waveform of speech that have been shown tobe deficiently processed by language-learning impaired (LLI) children. LLI children re-ceived extensive daily training, over a 4-week period, with listening exercises in which allspeech was translated into this synthetic form. They also received daily training withcomputer "games" designed to adaptively drive improvements in temporal processingthresholds. Significant improvements in speech discrimination and language compre-hension abilities were demonstrated in two independent groups of LLI children.

Exposure to a specific language alters aninfant's phonetic perceptions within the firstmonths of life, leading to the setting of pro-totypic phonetic representations, the buildingblock on which a child's native language de-velops (1). Although this occurs normallywithout explicit instruction for the majority ofchildren, epidemiological studies estimatethat nearly 20% of children fail to developnormal speech and language when exposed tospeech in their native environment (2). Evenafter all other primary sensory and cognitivedeficits are accounted for, approximately 3 to6% of children still fail to develop normalspeech and language abilities (3). Longitudi-nal studies have demonstrated a striking con-vergence between preschool language delayand subsequent reading disabilities (such asdyslexia). A broad body of research now sug-gests that phonological processing deficitsmay be at the heart of these language-learningimpairments (LLIs) (4, 5).

Tallal's earlier research has shown thatrather than deriving from a primarily linguis-tic or cognitive impairment, the phonologi-cal and language difficulties of LLI childrenmay result from a more basic deficit in pro-cessing rapidly changing sensory inputs (6).Specifically, LLI children commonly cannotidentify fast elements embedded in ongoingspeech that have durations in the range of afew tens of milliseconds, a critical time frameover which many phonetic contrasts are sig-naled (7). For example, LLI children haveparticular difficulty in discriminating be-tween many speech syllables, such as [ba]and [dal, which are characterized by very

P. Tallal, S. L. Miller, G. Bedi, G. Byma, Center for Mo-lecular and Behavioral Neuroscience, Rutgers University,Newark, NJ 07102, USA.X. Wang, S. S. Nagarajan, C. Schreiner, W. M. Jenkins,M. M. Merzenich, W. M. Keck Center for Integrative Neu-rosciences and Coleman Laboratory, University of Cali-fornia, San Francisco, CA 94143-0732, USA.

*To whom correspondence should be addressed.

rapid frequency changes (formant transi-tions) that occur during the initial few tensof milliseconds. Interestingly, LLI childrenare able to identify these same syllables whenthe rates of change of the critical formanttransitions are simply synthetically extendedin time by about twofold (8). A strong pre-diction is suggested by these findings: If thecritical acoustic cues within the context offluent, ongoing speech could be altered to beemphasized and extended in time, then thephonological discrimination and the on-linelanguage comprehension abilities of LLIchildren should significantly improve.

To test this prediction, we have conduct-ed two studies with LLI children who havebeen trained with the application of tempo-rally modified speech. These same childrenalso received training at making distinctionsabout fast and rapidly sequenced acousticinputs in exercises mounted in the format ofcomputer "games" (9). Modification of flu-ent speech was achieved by application of atwo-stage processing algorithm (10). In thefirst stage, the duration of the speech signalwas prolonged by 50% while preserving itsspectral content and natural quality. In thesecond processing stage, fast (3 to 30 Hz)transitional elements of speech were differ-entially enhanced by as much as 20 dB. Thistwo-step acoustic modification process wasapplied to speech and language listeningexercises that were recorded on audiotapes,as well as to the speech tracks of children'sstories recorded on tapes and on educationalCD-ROMs. The differential emphasis of fastelements also resulted in a speech envelopethat was more sharply segmented. This pro-cessed speech had a staccato quality inwhich the fast (primarily consonant) ele-ments were exaggerated relative to moreslowly modulated elements (primarily vow-els) in the ongoing speech stream. We rea-soned that amplifying the fast elementsshould render them more salient, and thus

SCIENCE *, VOL. 271 * 5 JANUARY 1996 881

on

Janu

ary

31, 2

007

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om