9.5, 9.7 Received 12 November 1065

Distinctive Features and Errors in Short-Term Memory for English

•'AYNE A. •'ICKELGREN

Departmentof Psychology,Massach•setts insill,de of Technology,Cambridge, Massachusells 02139

Errorsin short-termrecall of 23 Englishconsonants were tabulated and relatedto threedistinctive-feature systems.The consonantswere always presented in initial positionin a -voweldiagram, and the vowelwas always/a/. Subjectswere instructed to copya listof consonantsasit wasbeing presented, followed by recallof the list. Perceptualerrors were excluded from the recall-errormatrix by scoringfor recallonly correctlycopied consonants. The data were also analyzed in sucha wayas to eliminatedifferences in response biasfor differentconsonants. Having controlled for responsebias, each feature system makes predictions aboutthe rankorder of differentintrusion errors in recall.Each of thethree feature systems was significantly moreaccurate than chancein thesepredictions, but the mostaccurate system was one developed in the presentstudy. This system is a slightlymodified version of theconventional phonetic analysis of cousonants in termsof voicing,nasalit>', openness of thevocal tract (mannerof articulation),and . The resultssuggest that a consonantis codedin short-termmemory, not asa unit, but as a setof distinctive features,each of whichmay be forgottenat leastsemiindependently.

INTRODUCTION by a conventionaldistinctive-feature analysis on two dimensions:place of articulation(front, back) and open- NTRUSIONlists are not errorsrandom.in immediateRecent studiesrecall byof Conrad verbal t ness of the vocal tract (narrow,medium, and wide). and Wickelgren2,'• on short-term recall of lists of letters Petersonand Barne? and Miller • hadfound previously and digits have demonstratedthat intrusionstend to that thesame dimensions are involved in theperception have a vowelor consonantphoneme in commonwith of ,suggesting that perceptionand STM use thecorrect item. 7his indicatesthat theinternal repre- the samesystem of internalrepresentatives. Although it sentative of a (verbal) item in short-term mentory is not possibleat presentto determinewhether this (STM) is not a singleelement, but a set of internal systemof internalrepresentatives is sensory or motor, representativesof the phonemescomposing the item. no support can be obtainedfrom any of thesestudies The phonemic-codinghypothesis petunits partial for- for the more "abstract" (nonsensoryand nonmotor) gettingof an item and accotintsfor the phonetalcsimi- level of the Chomsky-Hallefeature system.7 larity of intrusions to the correct item. In auditory perceptionof consonantsin noise,Miller If the STM representativeof a letter or digit is a set and Nicel? demonstrated that errors are nonrandom of representativesof phoneroes,it is natural to ask and tend to correlate with their distinctive-feature whetherthe representativeof a phonemeis a set of analysis,described in Table I. The two purposesof the representativesof its distinctive features. A previous presentstudy are to (a) determine if errors in STM for study4 indicated that this was the case for for consonants tend to have features in common with phoneroes,and almostperfect rank-order predictions the correctconsonant, and (b) determinewhat distinc- weremade of the frequencyof differentintrusion errors tive-feature system best predicts these errors. In

• R. Conrad, "Acoustic Confusionsin ImmediateMemory," a G. E. Petersonand H. L. Barney, "Control Methods Usedin a Brit. J. Psychol.55, 75-84 (1964). Studyof the Vowels,"J. Acoust.Soc. Am. 24, 175-184 (1952). 2 W. A. Wickelgren,"Acoustic Similarity and Intrusion Erorrs • G. A. Miller, "The Perceptionof Speech,' in For Roman in Short-TermMemory," J. Exptl. Psychol.70, 102-108 (1965). Jakobson,M. Halle, Ed. (Mouton& Co., The Hague,1956), pp. a W. A. Wickelgren,"Similarity and Intrusionsin Short-Term 353-359. Memory for Consonant-VowelDigrams," Quart. J. Exptl. • N. Chomskyand M. Halle, Sound•allern of œnglish(to be Psychoh 17, 241-246 (1965). published). 4 W. A. Wickelgren, "Distinctive Features and Errors in Short- s G. A. Miller and P. E. Nicely, "An Analysisof Perceptual Term Memory for English Vowels," J. Acoust. Soc. Am. 38, Confusionsamong Some English Consonants," J. Acoust.Soc. 583-588 (1965). Am. 27, 338-352 (1955).

•88 volume39 number2 1966 DISTINCTIVE FEATURES AN[) SHORT-'I'ERM MEMORV additionto the featuresystem proposed by Miller and TABLEI[. H distinctive-featuresystem. Nicely (MN), a featuresystem proposed by Halle9 (H) - - and Chomskyand Halle? for the parsimoniousdestrip- C,mso-Voic- Nas- Vo- Conso- Contin- St•i- tion cf English sound structure and a feature system nant ing alitv calic nantal uant dent Grave Diffuse proposedby the author (W) are investigated.These P (} 0 0 l 0 0 l 1 featuresystems are describedin Tables II and I[1. 1• l 0 0 1 0 0 1 l MX hasfive dimensions: voicing, nasality, affrication, m 1 l i) 1 0 0 1 l duration, and place of articulation. Place has three t 0 0 0 1 0 0 0 1 values;the rest have two values.MN has only been d 1 o l) 1 o o o 1 definedfor a setof 16 consonants.Within that set,each 1 1 o 1 0 0 0 1 dimensionis definedon every consonant,and all con- 0 0 0 l 0 1 0 0 sonantshave a uniquecharacterization in termsof their' valuesoneach ofthe five dimensions. However, if MN i I 0 0 I 0 1 0 0 wereto be extended toall English consonants, it would k (} 0 o 1 0 o l 0 requiresome additional dimensions or valuesto handle g 1 0 0 1 0 0 1 0 laterals,selnivowels, and the consonants /6', /)'/, and f 0 0 II 1 1 1 l 1 ,;h/. ,' 1 0 0 1 1 1 1 1 H has eight binary dimensionson which consonants 0 0 0 0 1 1 0 0 l are classified.The rather large number of dimensions •5 1 0 0 1 1 0 0 1 resultsfrom the decisionto use only two valuesper s o o o 1 1 l o dimension.Halle gives a moderatelycomplicated articu- z 1 0 0 1 1 1 0 1 latorydescription of the values of thedimensions in the • 0 0 0 1 1 1 0 0 H system,but the system is nnnaturaland inelegant • 1 0 0 l 1 1 0 0 as a descriptionof articulationand no attempt is made 1 0 0 0 l 0 I 1 by Halleto validatethe system on thesegrounds. 9The w featuresshould be consideredto be "abstract," not r 1 0 l l I 0 0 0 necessarilyhaving a simpleacoustic or articulatory I 1 0 i 1 1 0 0 1 description,though recognition and production of speech Y 1 0 o 0 1 0 0 demandthat therebe some,perhaps complex, relation- h 0 0 0 0 1 0 1 0 ship betweenthe abstractfeatures and their acoustic and articulatorycounterparts. The H systemis designed primarily to give a parsimoniousdescription of the ad- lnissiblesound sequences in English,but we are con- TABLE1. MN distinctive-featuresystem. cernedwith how accuratethe H systemis in predicting the rank order of different intrusion errors in short-term recall. Consonant Voicing Nasality Affrication Duration Place '[he W system,like the MN system,is moresimilar p 0 0 0 0 0 to conventionalphonetic analysis of consonantsthan b 1 0 0 0 0 the H system.Only four dimensionsare used,but open- m 1 1 0 0 0 nessof the vocal tract has three valnesfor consonants, t 0 0 0 0 1 and placeof articnlationhas five vaines.Voicing and d 1 0 0 0 1 nasalityare exactlythe sameas in the other two systems. n 1 1 0 0 1 The single dimension of opennessin the W system k 0 0 0 0 2 handles"," which is handled (in g 1 0 0 0 2 a somewhat different manner) by two dimensions f 0 0 l 0 0 (affrication and duration) in the MN system and by v 1 0 1 0 0 four dimensions(vocalic, consonantal, continuant, and strident)in the H system.Place of articulationis coded 0 0 0 1 0 l on a 5-pointscale in the W system,on a slightlyrougher • 1 0 1 0 1 3-point scalein the MX system,and on two binary s 0 0 1 1 1 dimensions(grave and diffuse)in the H system. z I 0 1 I 1 Notice that the opennessof the vocal tract and place • 0 0 l 1 2 of articnlatlon arc the s•tmc two dimensions that were so 2 1 0 I I 2 accuratein predictingthe errorsin STM for English vowels.Of course,the valuesof the opennessdimension for vowels would begin with a value greater than that • M. Halle, "On the Basesof Phonology," in The .5'tructureol' Language,J. A. Fodor and J. J. Katz, Eds. (Prentk'e-Hall, Inc., for the semivowels.Thus, opennessis conceivedto code EnglewoodCliffs, N.J., 19641,pp. 324-333. on a single6-point scale the dil'ferencebetween (1) stop

the journalof the AcousticalSociety of America 389 W. A. WICKELGREN

digits are not appropriate populations in which to Conso- Voic- Nas- Open- investigateintrusions as a function of feature similarity. nant ing ality hess Place On the other hand, a population of consonants(fol- p 0 0 0 0 lowed by the vowel /a/) has a constant degree of b 1 0 0 0 phonemicsimilarity betweenevery pair of items. rn 1 1 0 0 Sincefeature similarity correlateswith the errors in t 0 0 0 1 auditory recognition,it is important to ensurethat all d 1 0 0 1 intrusionsare errors in short-term recall, not errors in n 1 1 0 1 auditory perception.This can be accomplishedby re- • 0 0 0 3 quiringsubjects to copythe consonantswhile they are • 1 0 0 3 being presented,covering what they have copied,and k 0 0 0 4 then recallingthe consonants.Only the consonantsthat are copiedcorrectly are scoredfor recall.This procedure g 1 0 0 4 T,•BLE III. W distinc- was adoptedin the presentexperiments. f 0 0 1 0 tive-feature system.

v 1 0 1 0 I. 16-CONSONANT EXPERIMENT .• 0 0 1 1 Sixteenconsonant-vowel (CV) items were used,con- ti 1 0 1 1 sistingof the 16 consonantsptkfOsibdgviSz•mn followed s 0 0 1 2 by the vowel /a,/ (as in father). A set of 100 lists of z 1 0 1 2 six CV items each and a set of 100 lists of seven CV • 0 0 1 3 items each were constructed. No consonant was ever • 1 0 1 3 used twice in the samelist. Subjectslistened to a list w 1 0 2 0 of CV items presentedat the rate of one item/sec. r 1 0 2 1 Subjectscopied the initial consonantof each CV item 1 1 0 2 2 while the list was being presented,covering what they y 1 0 2 3 had copied.After copyingall items, they attempted to recall the entire list of consonants in the correct order h 0 0 2 4 (by fillingin boxes).Time for recallof thelist wasabout 18 sec,so one trial lasted about 25 sec. Subjectswere 33 MassachusettsInstitute of Tech- consonants;(2) ; (3) semivowels,laterals, nology undergraduates taking psychology courses. and/h/, (4) high (narrow-opening)vowels, (5) medium They constituteda rather broad regionalsampling of vowels,and (6) low (wide-opening)vowels. Values of the United States of America. Seventeensubjects re- the place dimensionfor vowelslie within the range for ceived the lists with six items each, and 16 subjects consonants. received the lists with seven items each. The speaker was a male who had grown up in Connecticut. While the W syste•n gives a uuique articulatorv descriptionof every consonant,it doesnot attempt to Careful instruction and numerous examples were given prior to the experiment on the distinction be- give a compleledescription of the articulation of each tween/0/ and/t$,.l and between/z/ and/•/. Subjects consonant.Thus, lip-rounding,tongue-tip retroflection, lateralization, and pharynx width are not indicated were instructedto write "th" for /0/, "th" for explicitly, althoughsome might considerthese special "sh" for /'i/, and "zh" for / z/. featuresto be morecharacteristic of/W/,/r,l,/1,/, and II. 23-CONSONANT EXPERIMENT /h/ than the W classificationin terms of the general positionof the massof the tongue(place of articulation). When a subject recalls the wrong consonantin a Also, voicing is indicated as a two-valued dimension, particular position in a list, it is much more likely to be even though the degreeof effort necessaryto produce another consonantfrom the same list (intralist intru- is greater for stop consonantsthan for fricatives, sion) than a consonantnot presentedin the list (extra- nasals,and semivowels.If STM is primarily in a speech- list intrusion). Furthermore, the intralist intrusion motor syste•n, rather than an auditory system, the tendsto be from a nearbyposition. The previousexperi- degreeof effort necessaryto producevoice might be the ment was not systematicallycontrolled with respectto more accuratedimension. However, the 2-valuedvoicing the frequencyof occurrenceof all pairs of consonantsin dimensionis a reasonablefirst approximation in this the same lists at different degreesof adjacency.This case. could resttit in some random error in the frequency Since phonereit similarity has already been demon- of differentintrusions to any givenpresented consonant. strated to correlatewith intrusion frequency,it is im- The second experiment was designed to reduce this portant to controlphonetalc similarity in the population source of error and to obtain a short-term-recall error from2.whichlists are constructed.Thus, letters and matrix for all 23 consonants that can occur in initial

390 volume39 number2 1966 DISTINCTIVE FEATURES AND SHORT-TERM MEMORY

TAu[.EIV. STM recallerror matrix for 16 Englishconsonants.

Intrusion(consonant recalled) Omis- Total p b m t d n k g f v 0 •5 s z • 2 sions intrusions Consonantpresented and copied correctly P 901 29 18 36 23 21 30 11 10 2l 15 3 13 12 10 6 113 258

b 36 881 21 16 11 I5 9 24 13 31 13 0 11 17 4 12 93 233

15 11 998 17 5 29 24 14 1l 16 12 2 12 7 6 12 72 193

27 11 18 1055 8 23 19 20 21 8 26 9 18 18 9 12 86 247

39 20 18 21 994 45 21 35 17 16 27 11 11 20 1l 11 83 323

14 9 18 23 10 856 17 13 13 19 14 $ 13 14 5 10 62 197

k 35 12 16 18 8 1! 954 10 20 12 13 3 14 14 11 14 83 211

g 20 25 23 18 25 14 2l 83l 22 7 18 2 11 19 5 5 101 235

f 23 12 25 16 1 24 31 16 744 16 19 7 19 9 12 3 84 233

V 23 23 23 20 20 37 16 10 26 922 19 9 19 22 8 7 105 282

0 17 4 10 36 7 21 19 12 19 6 674 32 18 13 10 9 83 233

4 6 5 9 7 15 6 5 7 16 65 571 4 8 6 7 38 170

$ 16 13 19 21 9 13 16 13 18 12 25 4 911 22 35 13 69 249

g 13 9 11 18 13 7 14 32 10 3I 17 11 37 816 9 46 98 278

12 7 18 19 16 13 10 11 11 15 33 13 42 17 993 17 73 263

13 12 12 10 5 10 13 13 14 22 11 10 17 88 26 815 66 276

position in English (p,b,m,t,d,n/:,j',k,g,f,v,0,'tS,s,z,.•,•,be rememberedonly slightly better than chance. Bv w,r,l,y,h), followedby/a/as in father. making the list longer (nine instead of six or seven Lists consistedof nine different (37 items, with a items), we reducethe probability of correct recall of gronpof threeitems being presented in 1.5 secfollowed beginningand end items,and by introducingcue items b•' a 1.5-seepanse. Subjects copied the initial consonant we increasethe probability of recalling middle items. of every item as it was being presentedand covered Thus, the probability of correct recall can be more what they had copied. During the presentationof the nearly equated over the different positionsof the list list, they kept the recall boxescovered. Following the and set at some level that is nearly optimal for onr 1.5-seepause after the last group of three items, the purpose. subjectsuncovered the boxesin which they were to Secorot,we scorethe item in each position as if serial attempt to recall the list. One consonantin every group positionwere the only cue usedin recall. Obviously,it is of three had already been typed in its correct position not. Snbjectsare alsousing the adjacentitems that they in the recallboxes, so the subjectshad to recall only six have just recalled.By insertingcue items into every third of the nineconsonants. The typed-inconsonants (called position,we guaranteethat subjectswill never get very cne consonants)were in positions1, 4, and 7 on some far off in the cne items that they are usingin recall. trials, in positions2, 5, and 8 on other trials, and in Third, by groupingthe itemsto be recalledin three's, positions3, 6, and 9 on still other trials. Sincesubjects only two of which are to be recalled,we have a natural (lid not know during presentationof the list which con- •vayof controllingfor the frequencyof differentpairs of sonantswonld be typed in, they had to pay attention consonantsin the samelists. We simply use every one to all nine consonants. of the 23X 22= 506 pairs (approximately)equally often This relatively complicated procedure had three as the pair to be recalledin a groupof threeconsonants. purposes.Firsl, systematic errors are obviously not To useevery oneof thesepairs oncerequires a reason- goingto be obtainedwhen the memory trace is so strong ably large number of lists. It was decidedto use every thatthe subject recalls correctly, anti they are also not pair twice (and a few three times). At the sametime, likeIx, to be obtained when the memory trace has decayed the frequency of occurrenceof each consonant was COlnpletcly.Thus, it wouldbe desirai)le to testrecall at equate(l at 04 (except two that occnrred93 times). an intermediatedegree of strengthof the memory trace. l';achconsonant occurred (approximately) equally often The serial-positioneffect works against this, because in cuepositions. No consonantappeared more than once items at the beginningaud end of a list are generally in any two successivelists. Three setsof 120 lists each rememberedperfectly, while items in the middle may were used. 9 sec were allowed for recall of each list, 9

the journal of the AcousticalSociety of America 391 W. A. WICKELGREN

392 volume39 number2 1966 DISTINCTIVE FEATURES AND SHORT-TERM MEMORY

the journalof the AcousticalSociety of America 393 A. W icKEI. GP4EN

T,XBLEVII. Rankingof 22 otherpresented consonants

Intrusion (consonant recalled) p t, m t d n & ] k • f Rank of presented consonant 1 k j n b g m • r• 0 k v 9.43 5.71 13.22 6.62 7.25 13.35 7.02 5.75 10.47 8.89 9.89 2 b p r n 0 y • • •: j r 9.13 5.19 9.79 6.48 6.98 10.83 6.89 4.99 9.19 6.85 8.73 3 m d y p n d k 1 j r s 8.74 5.15 8.24 6.32 5.74 9.82 5.12 4.77 8.90 6.08 8.62 4 t k h v p t j 0 g b 15 8.48 5.12 8.04 6.07 5.64 9.82 5.02 4.65 8.90 5.94 8.52 5 w g w d y r 0 d 1• d O 8.43 4.76 7.95 6.0t 5.18 8.73 4.65 4.29 7.67 5.79 8.14 6 v y • '5 'iS v p p 1 • h 7.46 4.71 7.85 5.97 5.11 8.31 4.29 4.06 7.59 5.37 8.04 7 • w b z r 1 15 g 15 t k 7.13 4.58 7.31 5.87 5.03 8.03 4.26 3.93 6.82 5.13 7.82 8 0 r f s t p s z n z m 6.98 4.50 7.03 5.83 4.91 7.22 4.20 3.91 6.73 4.89 7.77 9 f v v • k j f b h f .• 6.81 4.49 6.97 5.52 4.85 6.85 4.18 3.88 6.62 4.84 7.60 10 h t d h j k 1 k f w b 6.38 4.24 6.87 5.20 4.79 6.74 3.90 3.77 5.93 4.82 7.31 11 r n p k b g w '• t p • 6.35 4.24 6.55 5.12 4.79 6.62 3.61 3.41 5.58 4.51 6.67

12 s O I r 1 h t • v m z 6.29 3.49 6.29 5.03 4.77 6.62 2.90 3.31 5.39 4.37 6.60 13 n z g 1 h s b r s • d 5.74 3.18 6.21 4.99 4.73 6.29 2.74 3.17 5.36 4.14 6.44 14 • h t •n z w g n z • p 5.46 3.07 6.03 4.85 4.40 6.27 2.69 2.99 5.13 4.04 6.32 15 y •: j f • f m t w 1 t 5.41 2.99 5.71 4.84 4.28 6.15 2.67 2.90 4.82 3.90 5.80 16 1 • 'iS g w b n h b h 1 4.99 2.85 5.40 4.55 3.61 5.71 2.49 2.84 4.79 3.55 5.42 17 g 1 • • f z r 5' d n • 4.97 2.82 4.99 4.13 3.30 4.65 2.38 2.82 4.51 3.49 5.37

18 z f s w m • h v r v n 4.65 2.63 4.90 3.61 3.16 4.28 2.36 2.47 4.23 3.15 5.24 19 d s 0 • v 15 y w y s y 4.08 2.56 4.65 3.33 3.15 4.26 2.35 2.41 4.23 3.03 5.18 20 • •5 z y •z ?: d f m '15 w 3.72 2.27 4.65 3.06 2.53 3.91 2.15 2.20 3.88 2.56 4.82 21 j m k j s E v s • y j 3.42 1.94 4.31 2.51 2.33 3.31 1.80 2.10 3.72 2.35 4.34 22 ti • •: 0 • 0 z m • 0 g 2.56 1.65 4.14 2.33 1.24 1.16 1.47 0.97 3.09 1.16 3.73

secfor presentationand copyingof the list, and 2 sec salnplingof the United Statesof America. The speaker for the "ready" signal;so a trial requiredabout 20 sec. was a femalewho spent the first 11 years of her life in Thns, each set requiredabout 40 min. Approximately Coloradoand who went to high schoolon Long Island, equalgroups of snbjectswere run in eachset, and there New York. were 71 subjectsaltogether. The subjectswere Massa- Beforethe recallexperiment started, the subjectswere chusettsInstitute of Technologyundergraduates taking given careful instruction and examples regarding the psychologycourses. They constituteda broad regional pronunciationand way of writing of each consonant.

394 volume39 number2 1966 DISTINCTIVE FEATURES ANI) SHORT-TERM MEMORY by percentsubstitution of eachintrusion consonant.

Intrusion (consonantrecalled) v 0 •5 s z • • w r 1 y h Rank of presented consonant z •/ 0 • • s z v w y w m 6.11 4.54 11.63 9.26 1 I. 16 7.46 8.56 5.62 11.33 12.24 5.78 6.55 n v t z s • 0 1 d p h w 5.24 2.70 4.24 8.80 7.69 6.61 5.81 5.21 10.52 11.51 5.67 6.27 y f v 0 f •z f: p ! n I b 4.71 2.64 3.37 6.98 4.84 3.22 5.75 4.29 10.20 10.97 5.64 6.16 p k f f j j • r m h m 1 4.51 2.43 3.30 4.84 4.34 2.51 5.70 4.23 8.98 10.17 5.10 6.0; r • s r d 0 j • n w õ 4.50 2.07 3.03 3.97 4.29 2.33 4.11 4.13 8.73 9.16 4.99 5.79 m z •. v g •5 t m g g • f 4.37 1.71 2.99 3.82 4.14 2.27 3.12 4.13 8.69 8.00 4.96 5.7' b • z h ! t • • '•5 • j j 4.34 1.65 2.20 3.78 4.12 2.23 2.84 3.98 8.52 8.8 t 4.34 5.7 t f s k • • f v t k m z v 4.18 1.63 2.16 3.72 4.1)4 2.20 2.70 3.79 8.36 8.50 4.16 5.39 d r • g • m w f f f d t 4.08 1.59 2.14 3.11 3.68 2.18 2.41 3.74 7.9 l 8.35 4.08 5.36 •5 g y b w w b s h r f s 3.98 1.45 2.12 2.97 3.61 2.17 2.28 3.73 7.09 8.20 3.74 5.36 g h g p •5 z k g y t n 3.93 1.42 2.07 2.93 3.41 1.96 2.16 3.73 7.06 7.59 3.74 5.2.t w t j •z k d m b • b r 3.86 1.34 2.05 2.76 3.23 1.93 1.94 3.65 7.02 7.08 3.44 5.06 • 1 I ]5 b k y d b k • y 3.72 1.30 1.95 2.56 3.20 1.87 1.88 3.65 6.85 7.01 3.22 4.91: j d r j v g s n I) •: s g 3.65 1.29 1.85 2.28 2.92 1.66 1.86 3.24 6.32 6.90 3.03 4.76 0 n w I t y g .• s g t k 3.49 1.25 1.69 2.17 2.90 1.41 1.86 3.09 6.06 6.65 2.90 4.5• t m • w m h r y 0 j v r 3.35 1.21 1.65 2.17 2.67 1.18 1.85 2.38 5.81 6.39 2.70 4.50 I y d t n b h z •z s k z 3.25 1.18 1.50 2.01 2.49 1.14 1.65 2.20 5.75 6.29 2.43 4.40 •: j h m r p 1 h j d 0 n 2.99 I. 14 1.42 1.94 2.38 1.13 1.52 2.13 5.48 6.22 2.33 4.24 h b p y h v d k t v p d 2.84 1.14 1.35 1.88 2.13 1.12 1.50 1.89 5.36 6.07 2.26 4.08 s p n d y ! p j • z g p 2.33 O.OO 1.25 1.72 2.12 1.08 1.35 1.71 4.75 5.87 2.07 3.84 • .• b k p r f • z • •5 1.90 0.48 1.14 1.62 1.81 0.53 0.66 ! .36 4.65 4.06 1.99 3.• 2

k w m n 0 n n 0 v 0 b 1.62 0.00 0.73 1.50 0.00 0.50 0.25 1.16 4.49 3.49 1.83 2.33

Fifteen trials of copyingpractice were given, usingthe consonantswere rather obviouschoices: "ch" for/•/, same type of lists and rate of presentationas in the "j" for/j/, "g" for/g/, "sh" for/•/, "zh" for, z/, etc. recall experiment, with the experimenter writing and pronouncingthe answersafter the subjectshad !inishcd III. RESULTS copyingthe list for that trial. To reducegraphic fusions,subjects were instructed to write "0" for/0/ In both experiments,only consoratorsthat had been and"th" for/•5/. Thegraphic symbols used for theother copiedcorrectly in the correctposition were scored for

the journal of the AcousticalSociety of America W. A. WICI•ELGREN

TABLEVIII. Accuracyof binary predictionsmade by three featuresystems.

System Dimension STM TotalSTM Auditoryperception 16 cons. 23 cons. 16 cons. %corr. • %corr. N %corr. • %corr. • all Voicing 70 98 60 168 64 266 100 98 all Nasality 46 28 61 42 56 70 68 28 MN Affrication 64 56 75 56 70 112 59 56

gin Duration 54 28 57 28 55 56 61 28 M_NI Place 66 152 67 152 67 304 59 152 H Vocalic & consonantal ...... 55 40 55 40 ...... H Continuant 71 28 63 84 65 241 57 28 H Strident 54 28 50 42 51 70 46 28 H Grave 67 70 61 125 63 195 63 70

H Diffuse 54 56 62 84 59 140 57 56 W Openness 64 56 71 185 69 241 59 56 W Place 79 104 75 247 76 351 65 104 MN (A-t-D-t-P) 64 236 68 236 66 472 59 236 H (V&C+C+S-[-Gq-D) 62 182 60 375 60 557 58 182 W (O+P) 74 160 73 432 73 592 63 160

recall. Thus, perceptualerrors were eliminatedfrom particular consonant,independent of its strength in the recall data. Correctly copied consonantswere STM. With 16 or 23 consonants,the maximum dif- scoredfor correctordered recall. That is, a consonant ferencesin responsebias couldbe quite large.Tables was scoredas correctlyrecalled, if and onlv if it was VI and VII presentthe conditionalprobabilities for the recalled in the correct box on the answer sheet. For all two experimentswith the entriesin eachcolumn being casesin which the consonantpresented in a position ranked from greatestto least probability that the in- was correctlycopied in that position,Tables IV and V trusion consonant would be recalled instead of each of showthe frequenciesof correctrecall, onfission in recall, the 15 or 22 possible presented consonants.For ex- and each of the 15 or 22 intrusions in recall of each of the ample,let us examinethe caseswhere/p/was an in- 16 or 23 consonants. Table IV is for the 16-consonant trusionin recallin the 16-consonantexperiment (column experiment,and Table V is for the 23-consonant labeled"p" in Table VI). /p/ was givenincorrectly experiment. most often in responseto the presentedconsonant/k/, Each intrusionfrequency in Tables IV and V was next mostoften in responseto/b/, next most often in dividedby the total nmnberof intrusionsfor that par- responseto /d/, etc. The conditionalprobability of ticularpresented consonant (i.e., it wasdMded by the recalling/p/ when/k/ was presentedand when the intrusiontotal for that row). This yieldsthe conditional subjectrecalled some incorrect consonant was 0.1659. probabilitythat a subiectrecalls a particularincorrect Each of the three feature systemsmakes binary consonantgiven a particular presentedconsonant, and (greaterthan) predictionsabout variouspairs of con- given that he makes an intrusionerror (rather than sonants in each of the columns in Tables VI and VII. recallingcorrectly or makingan omission).The purpose For example,according to MN, /k/ shouldrank above of this transformationis to adjust for differencesin the /g/in column"p" (i.e.,/p/should be recalledfor/k/ frequencywith which different consonantswere pre- moreoften than/p/ is recalledfor/g/), because/p/ sented,copied correctly, recalled correctly, or omitted and /k/ are both unvoiced,while /g/ is voiced,and in recall.We want the entriesin onerow to be directly /k/ and/g/ areidentical in ever)'other feature dimen- comparablewith thosein anotherrow. sion.This predictionis referredto as a predictionmade Having adjusted for theseextraneous differences in by the t,oicing dimension of the gin system,although "opportunitiesfor intrusions,"we cannow compare eth deafly the other dimensionsof MN play a r01einsofar obtainedconditional probabilities in one row with those as /k/ and /g/ must have the same values on these in the same column in other rows. This comparison other dimensions.There is no way known at presentto within a column (rather than within a row) equatesfor test whether one dimensionis a significantpredictor of responsebias--that is, lhe subject'sbias to emit any error frequency completely independentof the other

3c)6 volume39 number2 1966 DISTINCTIVE FEATURES AN1) SIIOI(T-TERM MEMORY dimensionsin a featuresystem. However, this way of sablesound sequences in differentlanguages and has a testinga singledimension is at leastpartly independent morecomplex and less"natural" articulatory interpre- of one'shypotheses about the otherdimensions. For ex- tation. The resultsof the presentstudy suggestthat, at ample,voicing and nasality are identicaldimensions in least for random sequencesof English consonants, all three featuresystems, and the binary predictions STM usesan articulatory (or acoustic)code, not the madeby voicingand nasalityare alsoidentical in all moreabstract H code.Distinguishing between an articu- threefeature systems, despite the fact that the other latory codeand an acousticcode is very difficultat the dimensions are rather different for the three feature presenttime and is not attemptedhere. However,it is systems.Nevertheless, it is the total accuracyof the clearthat onedoes not needto postulateabstract levels predictionsmade by an entirefeature system that is that arenot easilyunderstood in acousticor articulatory unambiguouslyinterpretable, not theaccuracy of each termsin orderto predictthe errorsin STM for random featurein thesystem. Table VIII givesthe numberof sequencesof Englishconsonants. It remainsto be seen predictionsmade by eachdimension of eachof the three whethererrors in long-termmemory or errorsin STM featuresystems and alsogives the percentageof these for syntacticallystructured sequences of Englishconso- predictionsthat were correctfor the reckilldata in nantswould be bestdescribed by the H system. TablesVI and VII. The auditoryperception data of Why are the predictionsof the W systemnot 100% Miller and Nicel? werealso analyzed in the sameway, accurate?There are many possiblereasons. First, per- and the results are included in Table VIII. haps the dimensionsof the systemshould be redefined ß he predictionsof the dimensionsthat distinguish slightly. One possibilitythat appearspromising is ,to betweenthe threefeature systems (i.e., excludingvoic- make voicinga multivalueddimension. This possibility ing andnasality) are summedat the bottomof Table wasnot investigatedbecause it wouldhave substantially VIII to yieldan over-allcomparison of the accuracy of reducedthe numberof predictionsmade by the W thethree systems in predictingthe errors in bothpercep- systemunder the presentanalysis, and I havenot been tion and STM for Englishconsonants. The accuracyof able to think of a better method of analysisthan the each of the three feature systemsis consistentlyand present one. Second,perhaps redundant articulatory significantlyabove chance,indicating that intrusion dimensionsshould be included, such as tongue-tip errors in STM tend to have distinctive features in retroflexion,lateralization, lip-rounding,etc. This was commonwith the presentedconsonant. not investigatedfor the samereason as before.Third, Furthermore,the W systemis consistentlymore ac- estimatinga largematrix of probabilitiesrequires a very curate than the MN system,which is, in turn, con- largenumber of trials beforethe randomerror is so low sistentlymore accurate than the H system.The differ- that onecan believe the results of everyindividual paired encesin accuracyare not statisticallysignificant in the comparison.The presentexperiment was not intended dataof Miller andNicely, but arehighly significant in to achievethat objective.Thus, it is quite conceivable the total STM data of the presentexperiment. For the that the 73% accuracyof the W systemis as high a total STM data, W is significantlymore accurate than repeatableaccuracy as the variancein the data permits MN at the 0.02 level (x•=6.17) and significantlymore for that numberof binary predictions. accuratethan H at the 0.001 level (x•= 21.3). Is the samefeature systemused in auditory percep- tion of consonantsas in STM for consonants?Compari-

IV. DISCUSSION son of the presentfindings with thoseof Miller and •-icely is consistentwith thisinterpretation. But there Regardlessof whichof the threedistinctive feattire is an ambiguityin the interpretationof error matrices systemsone chooses, it is clear that consonant phoneroes for auditory perceptionthat is not presentin the in- are not the most elementaryunits in which speechis terpretationof errormatrices for memoryexperiments. coded in STM. Consonants are not remembered in an The ambiguity concernswhere the informationthat all-or-none manner. Some of the features of a consonant resultedin an error was lost. When auditory perception can be recalledwhen otherscannot, producing a syste- is testedunder noisy conditions,as in the Miller and matic tendencyfor the errorsin short-termrecall to Nicely experiment,much of the informationloss is have distinctive features in common with the correct undoubtedlyoccurring oulside the subject.This could consonant.This suggeststhat recall of a consonant result in errors that follow a simple acoustic-feature means recall of a set of features that defines that conso- systemfor reasonsthat have nothingto do with the nant in memory,and eachfeature is recalledat least natureof theorganism. If onestrives to makethe condi- semiindependentlyof the otherfeatures. tionsas noisefree as possible, it wouldtake an unreason- 'ihe W feature systemdescribes the errorsin STM ably long time to obtain •ccurate estimatesof the ex- somewhatmore accuratelythan the MN systemand tremelylow errorprobabilities. muchmore accurately than the H system.The W and When a subjectcorrectly identifies a consonantand MN systemsare relativelyeasy to interpretin con- later recallsit incorrectly,one can be surethat the infor- ventional articulatorydimensions. The H systemwas mation loss has occurredinside the organismand the developedfor theparsimonious description of theadmis- errormatrix is providinginformation about the coding

the journalof the AcousticalSociety of America 392' W. A. WICKELGREN systemused by the organism.Of course,the information matrices in short-term recall. However, it should be may be lost at se,eraldifferent levels of the organism's clear that we cannot conclude from this that the W controlsystem, each of whichhas a very differentcode. systemis optimalfor predictingother types of behavior. If this is the case,then what we seein the STM error The multilevel branchingorganization of individual xnatrixis a compositeof thesecodes, and precisein- modalitiesin the humannervous system indicates that terpretationbecomes very complicated.However, even there are manr different codesfor analyzingthe same if there is multilevel information loss in the present information.It is an openquestion whethe• all formsof STM experiments,it is likely that the amountof infor- speechbehavior, syntactically structured and unstruc- mation lost at eachlevel is not equal. If one level pre- tured, are perfonuedin one systemthat usesone code. dominatesin the inforotationloss, its codewill showup ACKNOWLEDGMENTS as the bestpredictor of the STM error matrix. Thus, we are on reasonablysafe ground in concluding This work was supportedprimarily by a arant from that the W feature system (voicing,nasality, openness the Xational Institutes of Health, Public Health of the vocal tract, and place of articulation• is a good Service, U.S. Department of Health, Education, and first approximationto thecode used in STM for English Welfare. Further aid was received from a National consonants.Certainly, lhe other two feature systems Aeronauticsand SpaceAdministration grant to Hans- are significantly less accurate in predicting the error Lukas Teuber.

398 volume39 number2 1966