Phonetic Biases in Voice Key Response Time Measurements€¦ · A voice key is a device that determines voice onset in real time. The voice key is connected to a microphone, which

Journal of Memory and Language 47,145–171 (2002)doi:10.1006/jmla.2001.2835

Phonetic Biases in Voice Key Response Time Measurements

Brett Kessler and Rebecca Treiman

Wayne State University

and

John Mullennix

University of Pittsburgh at Johnstown

Voice response time (RT) measurements from 4 large-scale studies of oral reading of English monosyllableswere analyzed for evidence that voice key measurements are biased by the leading phonemes of the response.

tribu-as whenosterior,

d an effectere largechniques

Mof v

eve and

tion denandHowdenChriKath

AKes71 W763

Words with different initial phonemes did have significantly different RTs. This effect persisted after contions of nine covariables, such as frequency, length, and spelling consistency, were factored out, as well variance associated with error rate was factored out. A breakdown by phoneme showed that voiceless, pand obstruent consonants were detected later than others. The second phonemes of the words also haon RT: Words with high or front vowels were detected later. Phoneme-based biases due to voice keys w(range about 100 ms) and pervasive enough to cause concern in interpreting voice RT measurements. Teare discussed for minimizing the impact of these biases.© 2002 Elsevier Science (USA)

Key Words:voice key; response time measurement; naming task; oral reading; measurement bias.

uch research in psychology uses measuresaddress issues about how speakers retri
ocal response latencies to make inferen
e

hg

o

rs

). of, &kenvari-

oroneey.

ice to

ureheeney,e.n

about underlying processes. In some expments, the time that it takes to initiate pronunation of a printed word is used to study tprocesses involved in reading and word recotion (e.g., Forster & Chambers, 1973). Othstudies use the time to start repeating a spoword to shed light on auditory word recogniti(e.g., Connine, Mullennix, Shernoff, & Yelen1990). In still other experiments, spoken sponses are used to examine higher levellanguage comprehension (e.g., StanovichWest, 1983). By having participants name ptures instead of printed words, researchers

This research was supported by National Science Fou

14

nedhec-er-

g-es a

re-ft-

Grant BCS-9807736. We thank the anonymous respts to our Internet poll; Nancy Ciaparro, Kira Rodrigue Joe Inman for their assistance with the reseaard Nusbaum, David Balota, Daniel Spieler, Mark Se

berg, and Gloria Waters for sharing their data; astopher Kello, Stephen Lindsay, Ronald Peereman,leen Rastle for helpful comments on an earlier draft.

ddress correspondence and reprint requests to Bsler, Department of Psychology, Wayne State Univers

. Warren Avenue, Detroit, MI 48202. Fax: (313) 576). E-mail: [email protected].

0749-596X/02 $35.005

cesri-

ci-eni-erkenn,e- of&

ic-can

produce words (e.g., Griffin & Bock, 1998Vocal response times are used in studiesmemory as well (e.g., Scarborough, CorteseScarborough, 1977). The naturalness of sporesponses makes such tasks well suited to a ety of populations and a variety of issues.

By far the predominant technology used fdetermining voice response times, and the used in the studies cited above, is the voice kA voice key is a device that determines voonset in real time. The voice key is connecteda microphone, which converts sound press(the physical correlate of the amplitude of tsound) into voltage. When a stimulus has bepresented, a computer arms the voice kwhich begins monitoring the microphonWhen the sound pressure reaches a predefitarget, the voice key is triggered. It notifies tcomputer, which stores the number of milliseonds that elapsed between arming and trigging the voice key. The typical voice key is trigered as soon as the sound pressure reachcertain level, which the experimenter can pdefine (e.g., Cedrus, 2000; Psychology Soware Tools, 2000).

da-on-z,

rch;i-nd

and

rettity,7-

© 2002 Elsevier Science (USA)All rights reserved.

N

ys ri&hln

&&he

o

a

ta

dcbo

ib

ss

tya

m

eicmucina

nt’sanheut-m-

o-avesow-seinex-ast,a-y

l.,

, itre-s.t ater-eoc-yrednsee-ceulte-ave8).o-bleof-llim-n a

le

heapshOf

146 KESSLER, TREIMA

Traditional voice keys are not the only waof measuring voice response times. More phisticated voice keys can be configured totriggered by a lower level of sound pressuprovided that it is maintained for a certalength of time (e.g., Hutzler, 1999; Rastle Davis, 2002). Digital signal processing tecniques can compute voice response time arithmically, or produce waveforms for visual ispection (e.g., Bachoud-Lévi, Dupoux, Cohe& Mehler, 1998; Fushimi, Ijuin, Patterson,Tatsumi, 1999; Kawamoto, Kello, Jones,Bame, 1998; Morrison & Ellis, 1995). Althougsome of our discussion will be relevant to thmore sophisticated methods, this paper conctrates on voice keys, which are the sourcetiming data for the large majority of publishestudies that use measurements of vocal resptimes.

The general goal of this paper is to increour understanding of the potential inaccuracyvoice keys. In particular, we focus on phonebias. Psychologists often wish to compare retion times across stimuli that provoke differeword or word-like responses. They are rarelyrectly interested in response time differenthat are due entirely to phonetic differences tween responses. Unfortunately, there are stra priori reasons for believing that systemaphonetic biases exist. A response beginnwith /s/, for example, might systematically detected later, or sooner, than a word beginnwith /m/, yielding different measured respontimes for no other reason than the phoneticthe words involved.

One obvious reason that phonemes mighdetected at different times is articulatorSome sounds take more time for the humvocal apparatus to initiate. Sakuma, Fushiand Tatsumi (1997) reported, for example, th/s/ can be initiated especially quickly. The seond major factor is acoustic. Even aftphonemes are initiated, it may take the vokey different amounts of time to detect theThe culprit in this case is the sound presscutoff that voice keys use. Setting the voikey too low (making it too sensitive) resultsan unacceptable number of suspiciously f
response times as the voice key is triggered
, AND MULLENNIX

so-bee,n

-go--n,

seen-ofdnse

se oficc-

nti-

ese-ng

ticngeinge of

be:ni,

atc-re.ree

st

nonspeech noises such as the participabreathing. However, any setting higher thzero means that part of the beginning of tvocal response will be missed, because allterances have a nonzero rise time to target aplitude. This acoustic factor can result in phnetic bias, because different phonemes hdifferent rise times. For example, fricativesuch as /s/ have less sound pressure than vels (Fry, 1979; Sacia & Beck, 1926). Becauvoice keys are triggered only after a certasound pressure level is detected, we mightpect that under some circumstances, at levoice keys may take longer to register a frictive than to register a vowel, or, indeed, manot register the fricative at all (Sakuma, et a1997; Rastle & Davis, 2002).

Although we focus here on phonetic biasesis useful to note that this acoustic factor can sult in other types of problems with voice keyIf for any reason one utterance is spoken afaster tempo than another, then the faster utance will more quickly reach the point in thspeech stream where the target amplitude curs. For example, if initial /s/ is routinelmissed, then the faster utterance will moquickly get to the following phoneme, antherefore have the quicker measured respotime, even if the veridical times to initiate the rsponses are identical. That could introdumuch variation into the data. It could even resin experimental bias if factors that influence rsponse time also influence tempo, as we hgood reason to expect (Kawamoto et al., 199Such effects would, of course, interact with phnetic biases. If a voice key has special troudetecting initial /s/, then artifactual effects slow speech will affect /s/-initial words disproportionately. In addition, differences in overaamplitude of the utterance can have effects silar to those caused by differences in tempo. Iword like grow, the /ɹ/ is louder than the /g/, andthe /o/ is louder still. If the utterance as a whois very loud, a voice key might trigger at the /g/.If the utterance is less loud, so that all of tphonemes are proportionately softer, perhonly the /ɹ/ or even the /o/ will have enougacoustic energy to trigger the voice key.

bycourse the tempo of the utterance will largely

I

cphs

e

e

nln

dn

atiifoohhdt

in-en-hey ationsn-ry,

atice-ne

heedthathone

theiceall.nd bytd,

thisen

Toreb-

ndnyof

PHONETIC VO

determine how long it takes to get to those laphonemes. Thus voice key measures can flate veridical response time, tempo, and amtude in a complex way that interacts with tdifferential acoustic natures of the phonemethe response.

Our specific objectives here are threefoFirst, we wish to determine what the magnituof phonetic bias actually is in voice key studiIs it a purely theoretical consideration of triflinimport, perhaps amounting to errors of a fmilliseconds, or is it a much larger problem thcould seriously impact typical studies? Secowe seek to discover how pervasive the probis. Could it be solved, for example, by treati/s/-initial words separately from other words,does it pervade the entire phonemic inventoIs the problem limited to the first phoneme,might it extend more deeply into the worDoes phonetic bias affect different experimeequally strongly? Third, we try to separate trphonetic bias from associations that might mediated by psycholinguistic factors. For exaple, if words beginning with /z/ consistently hslower measured response times than owords, we would want to rule out the possibilthat that is due, perhaps, to the fact that /z/-tial words tend to be of lower frequency and miliarity. One potential objective we do nadopt is that of separating out the articulatfrom the acoustic source of phonetic bias. Twould be of little practical benefit to those wwish to use voice keys or evaluate earlier stuthat used them. For those people, the sources of phonetic bias form a bundled en
whose magnitude and variability needs to
n

sednseofweceinesithithnsr,esce

dealt with or evaluated as a unit.

SURVEYS OF CURRENT PRACTICE

Awareness about possible bias in voice kevaries greatly. To gather information about rsearchers’ concerns and beliefs, we conducan informal Internet poll during the summer 2000, to which 60 people responded. Of the55 indicated that they understood that differeinitial phonemes could differentially affect thtriggering of voice keys. That understandiwas unanimous among the 39 who had p
lished research that used voice keys. In an op
CE KEY BIASES 147

teron-li-e in

ld.des.gwatd,

emg

orry?or?ts

uebem-dhertyni-a-tryatoieswotitybe

yse-tedofm,nteg

ub-

ended question about what factors may bevolved in such bias, several respondents mtioned that the intensity or, more precisely, tintensity rise time of the initial phoneme maaffect how quickly the phoneme is detected, ifall. Many respondents stated some observatin articulatory terms. Few individual respodents offered more than a small part of the stohowever. The most widely cited belief, thvoiced phonemes are detected faster than voless ones, was mentioned by only about othird of the respondents.

While 92% of the respondents believed tfirst phoneme has an effect, only 45% believthat subsequent phonemes have an effect; proportion crept up to 53% for researchers whad published voice key experiments. About othird of the respondents (19) suggested thatsecond phoneme may be important if the vokey is not triggered by the first phoneme at Almost as many (16) opined that the secophoneme could affect voice key responsesmodifying the pronunciation of the firsphoneme. However, few specifics were offereand many volunteered that they considered a theoretical possibility that they had never sedocumented.

What about researchers’ actual practices?gather information about how experiments aconducted, we surveyed the 220 articles pulished from 1997 through 2000 in theJournalof Memory and Language.(All of these articleswere accepted for publication before the secoauthor of the present article, who edited maof the surveyed articles, was aware of mostthe issues discussed here.) Our survey focuon experiments that compared vocal respotime across different sets of stimuli. Somethe experiments used the same stimuli;counted these groups of experiments only onin our summary statistics, giving 48 casestotal. The authors of a number of the articlwere aware of some potential problems wvoice key measurement and tried to deal wthese problems in some way. These solutiowill be discussed more fully later in this papebut we introduce them here briefly. In 5 cas(all in one paper), the acoustic biases of voi
en-keys were avoided by using a digital signal de-

N

ek

i

snle

cof

a

t

hal

tt

n

ot

thoa

n n

ipl

nt

er-ns,orktialngri-hattialticetents

re- are ab-ndad-thathers

also seem to expect that different voice key

of

eirfulbyd.byn-an-dybstadthe

eac-ings,or-is-of8

148 KESSLER, TREIMA

tection algorithm to find the onset of the rsponse. In 12 cases, a delayed naming tassimilar control task was used in addition tostandard naming task. Delayed naming is antempt to isolate acoustic influences on vokey measurements from psycholinguistic fators. Another tack, used in many of the 36 cathat did not use delayed naming, was to balathe stimuli by phonemic criteria. For exampif experimenters believed that the voice keybiased by the first phoneme of the word, thmight ensure that each group of stimuli hadequal number of words beginning with eaphoneme. We found a clear difference in hexperimenters treated initial phonemes andlowing phonemes. Experimenters attemptedbalance the sets of stimuli for the initiphoneme in 20 of the 36 cases. As a reasonthis practice, the authors sometimes stateddifferent word-initial phonemes trigger thvoice key at different times. Those authors wattempted to balance the stimuli for initiphonemes sometimes indicated that they wnot able to do so in all cases. In such cases,often attempted to match a phoneme by onewas similar, but they usually gave no detaabout what standard of similarity was used.another 15 of the 36 cases, the authors didreport any attempt to equate the stimuli for itial phonemes. One additional case revealedintermediate course of action: The authequated the stimuli for the manner class ofinitial phoneme. Two other authors also metioned manner class as being important in demining when the voice key triggers, one of ttwo offering the specific hypothesis that plsives trigger the voice key more quickly thfricatives.

The situation was different for the secophoneme. There was an attempt to matchstimuli for the second phoneme in 8 of thestimulus sets for which delayed naming wasused. In the remaining 28 cases, no such attewas made.

Few authors reported the specific typevoice key used or the conditions of its use. TJournal’s guidelines stated that details of equment need not be reported unless essentia
replication, and most authors seemed to fe
, AND MULLENNIX

-or

aat-cec-esce,iseyanhwol-tolforhateo

that the make of the voice key had no importaeffects.

Combining the results of our survey of expimenters’ beliefs and our survey of publicatioit appears that experimenters who publish wusing voice keys are aware that different iniphonemes can differentially affect the triggeriof voice keys. In practice, however, expementers do not always equate the stimuli tthey are comparing across conditions for iniphoneme. The gap between belief and pracmay reflect the fact that such matching is ofdifficult to achieve given the other constrainon stimulus selection. In addition, some searchers appear to believe that the effectsrelatively small and will wash out if there arereasonable number of stimuli. Far fewer pulished studies equate stimuli for their secophonemes than for their initial phonemes. In dition, fewer researchers endorse the idea the second phoneme is important. Researc

ereheyhatilsInnoti-anrshen-er-e-n

dthe36ot

mpt

ofhe-for

setups will yield similar results.

PRIOR RESEARCH

Although none of the authors in the surveyarticles from theJournal of Memory and Lan-guagecited published research to support thviews about voice keys, there is some helpresearch that contrasts the timings obtainedvoice keys with more accurate timings inferreby visual inspection of the voice waveformThe impact of this research has been limitedthe fact that several studies were written in laguages other than English and investigated lguages other than English. The first such stuwe know of was Pechmann, Reetz, and Zer(1989). They asked their participants to rethe same word five times and measuredvoice key error by comparing the response timas measured by the voice key to much morecurate measures determined by analyzwaveforms. They reported not only high errorbut also large differences between trials. Fexample, the wordFreudewas, on average, detected by the voice key 104 ms after it was vible in the waveform, and the average rangevoice key errors for a given participant was 9
elms. Most words had smaller errors, but the

IC

poi

thdmul

thsynar

aa

u

mi ice lepvevetdreves

ru

w

uafetdeo

thero-ob-ng-etheer-eticar-id

omtri-

of

teda-of or-

estedac-ofcedylla-weoficaltoaskr itske-on are89;

n,-d

yne97;e

1nd

TMedft-

ofame

PHONETIC VO

wide variation between words was perhamore alarming than high but uniform errrates would have been. While a reasonableterpretation is that phonetic bias underliesbetween-word variation, Pechmann et al.not measure more than one word with the sainitial phoneme, so we do not know conclsively to what extent the initial phoneme itsewas the cause of the errors. Sakuma et(1997) ran experiments contrasting most ofpossible word-initial phonemes in JapaneThey found that the voice key was alwaslower than the waveform measurement athat the difference was biased by manner ofticulation: Vowels and nasals had a small diffeence from the waveform value, the liquid hmore, plosives more yet, and voiceless frictives had a very great difference. The rangethe difference between the phoneme growas about 95 ms.

Possible contributions of the second phoneof the word to phonetic bias have been studless. Rastle and Davis (2002) were the first torectly investigate the possible effects of vokey artifacts caused by the second phonemthe word. Their experiments with speakersBritish English contrasted words with simp(/s/ plus vowel) and complex (/s/ plus /t/ or /onsets. When the data were studied by waform, words with complex onsets were namfaster than those with simple onsets. Howewhen a voice key of typical design was used,reverse pattern was obtained. These results ument how voice key artifacts may lead searchers to draw false conclusions when intigating questions such as the effect of oncomplexity on the reading of printed words.

Voice waveforms are much more cumbesome to use than voice key timings, so previoresearchers who directly compared the tmeasurement types have understandablystricted their experiments to a few stimultypes. Consequently, many questions remFor example, does the second phoneme afthe measured naming latencies in words thanot begin with the /s/ plus plosive clusters stuied by Rastle and Davis (2002), or are such (ratively infrequent) word-initial sequences
two voiceless phonemes an isolated spec
E KEY BIASES 149

srn-e

ide

-fal.ee.sdr--d-

ofps

eeddi-e ofof

/)e-dr,

heoc--s-et

-sore-sin.ct

do-l-f

case? We also noted that languages differ inway nominally equivalent phonemes are pnounced, and asked whether the resultstained for German, Japanese, and British Elish hold for North American English. Becausof these outstanding issues, we perceivedneed for a larger-scale study, using North Amican English. Because we investigated phonbias as a whole and did not seek to factor theticulatory from the acoustic components, we dnot need to use voice waveforms. This freedallowed us to economically run thousands ofals and also to compare our results with thoseprevious studies.

In the experiment reported here, we presena large number of words—virtually all the fmiliar one-syllable monomorphemic words English—to 20 native speakers in standardthography. We recorded their response timusing a typical voice key. Known or suspeccovariables such as word frequency were counted for statistically. Some (but not all) the lexical covariables could have been reduby asking participants to repeat nonsense sbles instead of reading words. However,placed priority on replicating the conditions speeded naming tasks, which are a more typapplication for voice keys; we did not want accidentally pass over any stage of the twhich might introduce phonetic bias, whethebe articulatory or acoustic in origin. The taalso allowed us to investigate variability btween experiments by affording compariswith three other naming megastudies. Thesethe studies of Seidenberg and Waters (19henceforth SW) at McGill University; TreimaMullennix, Bijeljac-Babic, and RichmondWelty (1995; henceforth TMBR), which studieconsonant–vowel–consonant words at WaState University; and Spieler and Balota (19henceforth SB) at Washington University. Wwill refer to our own study as KTM. See Tablefor the number of participants in each study, athe number of words that were tested. The Kand TMBR studies used the voice key includin the response box supplied with the MEL soware package (Schneider, 1988). In boththese studies, the voice key was left at the s
ialfixed setting for all trials. SB used a Gerbrands

4).

als.. Data un-

reiman et

Number of letters.i Number of words that differ by substituting one letter (Coltheart, Davelaar, Jonasson, & Besner, 1977).

Model G1341T voice-operated relay, and S

150 KESSLER, TREIMAN, AND MULLENNIX

TABLE 1

Summary Statistics for Each Study

Measure KTM TMBR SB SW

Participants 20 27 31 30Words, tested 3690 1327 2870 2897Words, analyzed 2982 1234 2525 2536RT (ms) 629 611 467 568

Error ratea 4.6 1.3 — 6.0

Bigram frequencyb 1493 1816 1529 1531

Consistency of onsetc .976 .966 .975 .976

Consistency of rimec .913 .906 .908 .907

Familiarityd 6.6 6.6 6.7 6.7

Frequency of spellinge 3172 2526 2559 2584

Homophonesf 1.1 1.2 1.1 1.1

Length of pronunciationg 3.5 3.0 3.5 3.5

Length of spellingh 4.4 4.1 4.4 4.4Neighborhood sizei 6.8 9.4 7.1 7.1

Note. Averages across words. For RT and error rate, the per-word measures are themselves averages across tria Percentage of mispronounced responses, excluding those with outlying response times or failures to respond

available for SB.b Average text frequency of the two-letter sequences in the spelling (Solso & Juel, 1980).c Proportion (by type counts) of words in the list that have the same spelling that also share the pronunciation (T

al., 1995).d Nusbaun, Pisoni, and Davis (1984).e Corpus frequency in Zeno, Ivenz, Millard, and Duvvuri (1995).f Number of words in list with this pronunciation (minimum 5 1).g Number of phonemes.h

sa oeipna

m

e

he-

wnteslcyr-dslerllye-ol-bseter-

om

used a custom-made voice key.

METHOD

Participants

The participants were 20 undergraduate dents from Wayne State University. Three pticipants dropped out from an original group23, leaving 6 men and 14 women. The studreceived course credit in exchange for partiction. All were native speakers of English. Nohad a history of speech or hearing disorder,none had uncorrected visual problems.

Materials

A set of 3690 words was taken from a coputerized version of the Merriam-WebsterPocket Dictionaryand Webster’s Seventh Coll
giate Dictionary that was developed for suc
W

tu-r-f

ntsa-end

-

-

studies as Nusbaum, Pisoni, and Davis (198Homographic heterophones (e.g.,wind) andmany unfamiliar words were excluded, but tlist included many words that differ only in inflectional ending (e.g., both blow and blownwere included) as well as many words unknoto the great majority of college undergradua(e.g., gneiss). In order to avoid the statisticanoise that would be introduced by redundanand by the inclusion of unknown words, we peformed our analyses only on the 2982 worthat also occur in the word list used for Kessand Treiman (2001). The latter list was carefucontrolled to have only morphologically nonrdundant words that are generally familiar to clege undergraduates. The selection of this suwas made before we looked at any of the expimental data.

We also analyzed, separately, the data fr
hthe three previous megastudies. Table 1 shows

IC

da

una

wr k 11.

dsirse rtr te

mi

tr twth nn

e-ofa

tly

thees,n-

re-of

erheai-wemeatrdsinsi-e,

rdsis

utdsi-l.tin

entnse

herTheel

PHONETIC VO

the number of words remaining in all four stuies after intersecting their stimulus list with thof Kessler and Treiman (2001).

Procedure

Participants were instructed to read alowords that appeared on a computer screequickly and accurately as possible. On etrial, the prompt GET READY appeared for 2in the center of the screen. The screen then blank for 1 s. The stimulus word was then psented in uppercase letters. It remained onscreen until the voice key picked up the sporesponse. Then the screen went blank foruntil the next warning message. After every trials, the participants received a 1-min break

All of the 3690 monosyllabic English worwere presented to each subject, in eight sesthat each tested from 461 to 463 words. Wowere presented in a different, randomly lected, order for each subject. Participants wadvised not to make any extraneous soundscould trigger the voice key. They were askedkeep their lips about 4 inches from the micphone throughout the experiment. An Elecvoice RE16 cardioid microphone was usedgether with the aforementioned voice k(Schneider, 1988). Trials with response timquicker than 100 ms or slower than 2000 were rejected from the analysis. The remaintrials were hand-coded to indicate whetherpronunciation was in error. A subset of the sponses (one randomly selected 461-wordfrom each of five subjects) was checked by different judges to determine accuracy of hand-coding; 96.7% of the judgments agreedthe correctness or incorrectness of a respoFor the purpose of error analyses, mispronu
ations were counted as errors even when
i

m

y

ourimere islse

n-imeirdore

participants subsequently corrected themselv

RESULTS AND DISCUSSION

We first present analyses dealing with vokey effects caused by the word-initial phonemWe then discuss effects of the second phone

Initial Phoneme

Overall analyses. First we grouped words b
their initial phoneme and asked whether tho
E KEY BIASES 151

-t

d, aschsente-theen s

00

sonsdse-re

thattoo-o-o-y

ess

nghee-listo

isonse.ci-thees.

cee.e.

groups differed significantly in their average rsponse times. We chose a Monte Carlo testsignificance (Good, 1995), as illustrated withsmall subset of the data in Table 2. We direcdetermined the likelihood that ourF ratio couldbe due to chance by randomly rearrangingdata between phoneme groups 10,000 timsubject to the constraint that each group cotains a constant number of words acrossarrangements and seeing what proportionthose rearrangements had anF ratio greaterthan or equal to ours. This method makes fewassumptions than are required for using tstandardF distribution and also providesstraightforward way of factoring out the contrbution of the second phoneme: Each timerandomly rearranged words across phonegroups, we only swapped pairs of words thhave the same second phoneme (i.e., wowere not swapped across the horizontal linesthe table). In effect, the data were cross-clasfied in blocks based on the second phonemand in each rearrangement the number of woin each of those blocks did not change. Thtechnique of blocking on variables to factor othe significance of their effect was usethroughout our analyses. We also blockedmultaneously on the participant in each triaBlocking allowed us to isolate the differensources of variance, yet still run all the dataone large test of significance.

The first row of Table 3, “Raw RT,” shows thresults of this analysis. Words with differephonemes have different measured respotimes, at a significance of .000, i.e.,p , .001.We ran the same analysis for the three otmegastudies and obtained the same result. only difference in those tests is that trial-levdata were not available, and so we used asbasic level of analysis the average response tfor each word. These results suggest that thea voice key bias: Words with different initiaphonemes have significantly different respontime measures.

One might suspect, however, that the relatioship between initial phoneme and response tis not direct, but is mediated through some thcovariable such as spelling length. We theref

setagged each of our words to identify the levels

fet

isch

r-

e

n r

aat

r

h

theord

, theple,ente-ach aree

renotil--me,esen

telye-es, orambe-am

ing: (1)e number

hefraction of rearrangements for which the metric is greater than or equal to that of the original is p.

of potential covariables. The mean values these covariables are listed in Table 1. Tablshows what happened when we tested whewords with different initial phonemes differ significantly for the various covariables. The stattical techniques used for these tests were exalike those used for response time, except tthese tests were conducted for each word,stead of for each trial. In almost all cases, wowith different initial phonemes have significantly different levels of each of the covariableThese tests leave open the possibility that thare in fact no phonetic voice key biases: The dferent phonemes could have different respotime measurements because words that haveferent phonemes also happen to vary with spect to the true causal factors.

However, an additional round of statisticblocking can factor out the effect of individucovariables. For each covariable, we teswhether there was still a significant (p , .05)difference between the response times for wowith different initial phonemes after we blockeon different levels of the covariable. That is, t
setup was similar to that illustrated in Table 2
or 3her--tlyatin-ds

s.re

if-sedif-e-

lled

dsde

except that the second blocking variable wascovariable instead of the speaker, and each wwas measured only once. In almost all casesassociation remained significant. For examthe footnotes (a) in Table 3 show that when wblock on that covariable, words with differephonemes still have significantly different rsponse times. (It should be noted that in ecell of the table, the number and the footnotereferring to two different tests. An entry lik“Familiarity .090a” means that when words agrouped by initial phoneme, the groups do differ significantly among themselves in famiarity, at p 5 .090; and they do differ significantly among themselves in raw response tiwhen blocked by familiarity.) These analyswith blocking show that the association betwephoneme and response time is not ultimadue to spelling consistency, familiarity, frquency of spelling, number of homophonword length as determined by pronunciationspelling, or neighborhood size. Tests of bigrfrequency were less readily interpretable cause virtually every word has a unique bigr


TABLE 2

Illustration of Monte Carlo Rearrangements

Blocking Variables Original One rearrangement

Vowel Speaker /b/ /d/ /b/ /d/

/{/ 1 bad 660 dab 748 dab 748 bad 660badge 806 dad 738 badge 806 dad 738

dam 797 dam 797

/{/ 2 bad 840 dab 550 dam 490 bad 840badge 629 dad 521 dab 550 badge 629

dam 490 dad 521

// 1 bib 815 did 826 bib 815 bid 857bid 857 dig 518 did 826 big 612big 612 dill 473 dill 473 dig 518

// 2 bib 508 did 565 bid 522 bib 508bid 522 dig 569 big 522 dig 569big 522 dill 737 did 565 dill 737

Sum 6,771 7,532 6,317 7,986Metrica 9,312,229 9,305,132

Note. Each rearrangement randomly shifts words and their response times across columns, subject to the followwords are not shifted across blocks (demarcated by horizontal lines); (2) each column of each block keeps the samof words.

a Sum of the mean squares of the column sums; this is the portion of the F ratio that varies between rearrangements. T

,frequency, resulting in block sizes that were too

i

d aa

t ctisr w anoh

aninte lntte n

d sd

ctarr-

dif-beticre-

ntal re-es-ta7)nt

ana-ion ofatifi-ul-

htthesy-e-ionit toto

b Raw RT not significant when blocked by this covariable.

small to allow significance to be determinedany event.

Although those blocking tests show that invidual covariables are not responsible for thesociation between initial phoneme and meured response time, one might suspect thaassociation is mediated by a combination of variables. To test that possibility, our final statical test on initial phoneme groups was a gression analysis, using response time asresponse variable and all the covariables, as as error rate, as predictor variables. For eachvariable, we used the transformation that counted for the most variance in the respotime as averaged across all subjects, as cputed by a single-factor linear regression. Tpredictor variables in the multiple regression counted for 35% of the variance in the respotimes in the KTM study, 27% in TMBR, 28% SB, and 23% in SW. For each word, we demined the average residual response timeafter subtracting the predicted values accoufor by the linear regression. Then we teswhether there were significant differencesresidual response time for words with differeinitial phonemes.

The results of this analysis are presentethe last row of Table 3. For all four studiewords with different initial phonemes differe
significantly (p , .001) in residual response
n

i-s-s-theo--

e-theellco-c-sem-e

c-se

r-eftedd

int

in,

time. That is, even after factoring out the effeof all the psycholinguistic covariables by lineregression, there was still a significant diffeence in response times between words with ferent initial phonemes. This difference must due to factors other than the psycholinguisfactors, namely, phonetic voice key measument biases.

The converging evidence over four differestudies leaves little doubt that the initiphoneme itself has a direct influence on thesponse time. This corroborates previous suggtions from analyses of the SB and TMBR dathat were reported in Spieler and Balota (199and Treiman et al. (1995). However, the curreanalysis is more direct and convincing. Treimet al., for example, included the phonetic fetures of the initial phoneme in a large regresstest to predict the average response timewords from many predictors and found thsome of those phonetic features were signcant. However, there is no guarantee that a mtiple regression will assign variation to the rigpredictor variable: The variance assigned to phonetic features could actually belong to pcholinguistic covariables that are highly corrlated with those features. Our current regressanalysis is more conclusive because we use factor out all effects that can be attributed

PHONETIC VOICE KEY BIASES 153

TABLE 3

Significance of Differences in Response Variables Caused by Initial Phonemes

Response KTM TMBR SB SW

Raw RT .000 .000 .000 .000

Bigram frequency .000b .000a .000b .000b

Consistency of onset .000a .000a .000a .000a

Consistency of rime .000a .007a .001a .000a

Familiarity .090a .043a .017a .020a

Frequency of spelling .000a .000a .000a .000a

Homophones .000a .000a .000a .000a

Length of pronunciation .000a 1.000a .000a .000a

Length of spelling .000a .000a .000a .000a

Neighborhood size .023a .000a .000a .000a

Error rate .000a .012a — .426a

Residual RT .000 .000 .000 .000

a Raw RT remains significant (p # .05) when blocked by this covariable.

nonphonetic factors and then do association

N

oaAehhre

itirli z-

tedac epia

vr

aeofo

aoth h rdfooda

hasm

ith-

e,

hatef- in inet,atavele,

redr by byr-lyts. ofduehathe

if-ithurchutx-

lu-no-.

inse

utofanyokemon-ofe-eenis-

e-

154 KESSLER, TREIMA

tests only on the residuals. Also, we factor any contributions of the second phoneme,we reinforce the results by blocking tests. though the latter are limited to single variablthey reliably remove 100% of the effect of tcovariable in consideration without making tstatistical assumptions required for linear gression.

Analyses comparing pairs of phonemes. Evenafter being convinced that there is some vokey bias by initial phoneme, one could simagine that the effect is quite limited and thefore, perhaps, easily controlled. One possibiwould be that only particular phonemes areconcern. For example, Bates, Devescovi, Pizmiglio, D’Amico, and Hernandez (1995), responding to reports that fricatives and affricacause problems in voice key studies, addeflag in their regression analyses to indicwhether the word in question began with a fritive or affricate. To examine whether thereany possibility of addressing the voice key qution in such a straightforward way, we ran serate tests contrasting each pair of initphonemes. This may at first alarm readers ware alert to the fallacy of attempting to prothat variables are associated by running sepatests on each of their levels. However, we halready demonstrated that the initial phonemassociated with the response time. Our purpnow is to explore what factors may account the overall association.

The procedure for testing particular pairs wessentially the same as for the residual resptime analysis across all phonemes, except the presence of exactly two groups at a timelowed for a simpler metric abstracted from tstandard t test. Tables 4 through 7 present thesults. These tables are arranged by descenorder of differential residual response times the phoneme. These differentials show hmuch more slowly words with the indicatephoneme (in the first column) were named thwords with different first phonemes but tsame second phonemes. That is, for ephoneme, we computed first the average reual response time for words with that phonein second position (R2). Similarly, for each two-
phoneme sequence, we computed the avera
, AND MULLENNIX

utndl-s,ee-

celle-tyofa-

s atea-iss-a-l

hoeateve isser

snseat

al-ee-ingr

w

nechid-e

residual response time for words that begin wthat sequence (R1,2). Then, to obtain the differential residual response times for a phonemwe averaged the difference R1,2 2 R2 for all thetwo-phoneme sequences that begin with tphoneme. This computation factors out any fect of the second phoneme. In each rowTable 4, we present first the initial phonemeconsideration, its residual response time offsand the number of words that begin with thphoneme. Then we list the phonemes that hsignificantly faster response times. For exampthe 512 words beginning with /s/ are measuas 38.5 ms slower than can be accounted fothe levels of the covariables in those words orthe effect of the following phoneme. Furthemore, /k/, /t/, /n/, and so forth have significantsmaller (faster) residual response time offseThe word counts are important, because lacksignificance between two phonemes can be to the fact that there are only a few words tbegin with those phonemes, diminishing tpower of statistical tests.

The tables leave no doubt that significant dferences in response time between words wdifferent initial phonemes are pervasive. At osignificance level of .05, we would expect earow to have only one or two significant cells, balmost all rows have many more. The few eceptions are vowels, about which firm concsions cannot be drawn because very few mosyllabic words in English begin with vowelsThey are included in the tables primarily order to show their differential residual respontimes.

Analyses examining classes of phonemes. Ta-bles 4–7 contain a wealth of information abodifferences between individual pairs phonemes. Our next step is to ask whether generalizations emerge from the data. We toseveral passes over the data, breaking thdown by each of the three main classes of csonant articulatory features: voicing, place articulation, and manner of articulation. Bcause significance testing has already bdone, we apply only informal summary stattics in this phase of the analysis.

The patterning most often mentioned by r
gespondents to our Internet poll (35%) was that

sndi

st-

er)-edes,reerre

antage

olumn,tion of

f

voicing would affect the measured respontime. All of those who mentioned the directioof the effect said that voiced phonemes aretected faster than voiceless phonemes. A quglance at Table 4 shows that this cannot beabsolute rule: There are many voiced phonemthat are detected more slowly than voiceleones, and the slowest of all phonemes arevoiced /z/ and /i/. However, if we look at ten

a Response time residuals (milliseconds) after effect o

dencies rather than absolutes, the voicele

e

e-ckaness

he

phonemes do seem to cluster in the top (slowhalf of the table. One way to quantify that impression is to compare the number of voicphonemes that are faster than voiceless onand vice versa. We counted 97 times whevoiceless phonemes are significantly slowthan voiced phonemes, but only 12 times whevoiced phonemes are significantly slower thvoiceless ones; the voiceless have the advan

covariables is removed by regression.


TABLE 4

Differences in Response Time by Initial Phoneme Pairs (KTM Study)

Phone RTa N

z 40.6 9 f p dZ d h j b m ɹ w l θi 39.9 6 bʃ 38.6 76 k t n ð f g p dZ d a h j v b m ɹ w l ɔ θ ɑs 38.5 512 k t n ð f g p dZ d 2 h j v b ε m ɹ aυ w l ɔ e Ã θ o ɑ ɔtʃ 37.1 55 t n ð f g p dZ d h j v b m ɹ w l θk 21.8 239 t n ð f g p dZ d h j v b m ɹ w l ɔ θ ɑ ɔt 12.3 163 f g p dZ d a h j v b m ɹ w l θ ɑn 9.8 78 h j v b m ɹ w l θð 7.8 13 w l θf 4.6 189 d h j b m ɹ w l θ ɑg 3.9 150 d h b m ɹ w l ɑp 2.9 201 d h v b m ɹ w l θ ɑæ 0.2 16 aυ odZ 23.2 47 h j b m ɹ w l θd 26.0 138 h j b m ɹ w l2 26.3 7 ba 27.1 6 bh 29.1 134 w l θ 29.9 11 b ɔ oj 213.2 27 w l θv 213.3 36 w lb 214.5 247 ε aυ w lε 215.6 11m 217.4 124 w lɹ 217.6 131 w laυ 221.7 6w 222.6 127l 226.7 136ɔ 227.8 11e 228.0 12Ã 229.6 3θ 233.6 40 ɑo 235.1 8ɑ 237.9 11ɔ 250.2 1u 279.0 1

Note. Entries tell which phonemes have voice key response times significantly faster than the phoneme in the first cp # .05. Effect of second phoneme is blocked. See International Phonetic Association (1996, 1999) for explanaphoneme symbols.

ssin 89% of all the untied pairs. Applying the

tiath

n

ifenr

eugee

s

a

hees

atce-t aereeinter.ue

iveeyiase-ob-eyn

anip

so-

Note. See notes for Table 4.

same counts to TMBR (Table 5), we get a raof 75:6 pairs where voiceless phonemesslower than voiced phonemes (93%). Thus,rule for voiced phonemes having faster respontimes holds for both studies carried out usithe same equipment in the same lab at WayState University. The picture is somewhat dferent for the other two megastudies, howevIn SB (Table 6) the ratio is only 63:40 (61%); iSW (Table 7) the ratio is 80:29 (73%). Howeveit could be the case that voicing is only corrlated with some other feature that is the trcause of the response time difference. Toaround such problems, we next comparphonemes that differ only by the voice featurbut are otherwise identical. Comparing respontimes of such pairs in KTM, we found four pairsuch as /tʃ/ and /dZ/, where the voicelessphoneme is significantly slower than its voicecounterpart, but only one, /θ/ and /ð/, where thevoiced phoneme is slower (4:1). These ratios
4:0 for TMBR, 2:4 for SB, and 1:1 for SW.
oree

segne-r.

,-eetd,

se

d

re

What can account for these patterns? Tgeneral tendency to detect voiced phonemmore quickly is doubtless due to the fact thvoiced phonemes tend to be louder than voiless ones (Fry, 1979). If a voice key is set afairly high threshold, then it may skip over thsofter, voiceless segments to a greater degthan it skips over voiced ones, resultingshorter measured response times for the latThe differences between studies could be dto differences in thresholds. The more sensitthe apparatus (e.g., the lower the voice kthreshold is set), the less there should be a bdue to the difference in acoustic intensity btween voiced and voiceless sounds. A lessvious source of bias could arise if the voice ktrigger is so high that it tends to miss evevoiced obstruents, which are less loud thvowels. In such an event, the voice key will trearlier for words beginning with voicedphonemes, not because it detects those con


TABLE 5

Differences in Response Time by Initial Phoneme Pairs (TMBR Study)

Phone RTa N

tʃ 29.1 37 t f p g dZ n h ɹ b v m d l w θ j ðʃ 27.7 46 g dZ n h ɹ b v m d l w θ j ðs 27.4 83 t k f p g dZ n h ɹ b v m d l w θ j ðt 18.1 68 dZ ɹ b m d l w θ j ðk 17.1 75 g dZ n ɹ b v d l w θ j ðf 15.2 62 dZ ɹ b d l w θ j ðz 14.1 6 jp 9.4 87 b d l w θ j ðg 8.2 51 d l w θ j ðdZ 1.0 30 w θ ð2 0.0 1n 21.7 59 d l w θ j ðh 23.2 80 d l w θ j ðɹ 23.7 97 d w θ j ðb 29.8 93 w θ j ðv 213.8 23 ðm 214.8 79 j ðd 220.2 71 θ j ðl 225.2 88 w j ðw 225.9 59 j ðθ 248.7 13j 256.7 17 ðð 256.7 9

nants, but because voiced phonemes are about

a

c

ere

mo

t

r

re-e-l

mer

r-e

hypothesis; none contradict it (10:0). The ratios


40 ms shorter than voiceless phonemes (Kl1976).

In our poll, five people (8%) mentioned plaof articulation as a dimension along which intial phonemes could differentially affect voickey response time. Of those five, only two spondents offered a specific hypothesis. Ostated that labials have shorter response tithan other consonants. The other stated mgenerally that anterior consonants are detecfaster than posterior ones. The data conform
slight variation of the anteriority hypothesis
tt,

ei-

-neesre

tedo a

velar . coronal (alveolar, postalveolar, opalatal) . labiodental . dental . labial (bila-bial or labiodental), where “.” means that thesymbol to the left has the greater measuredsponse time. (A strict interpretation of the antriority hypothesis would place labiodentasounds after dental sounds, because the forinvolve the lips.) In the KTM study, if we lookat consonant pairs that differ only by place of aticulation, 10 pairs like /k/ and /t/ agree with th


TABLE 6

Differences in Response Time by Initial Phoneme Pairs (SB Study)

Phone RTa N

z 33.5 8 ð s g k θ p j b d h t dZ l n ɹ w m tʃ ʃf 18.1 161 s g k v θ p j b d h t dZ l n ɹ w e m tʃ ʃð 12.3 13 θ p j b d h t dZ l n ɹ w m tʃ ʃs 10.9 415 g p j b d h t dZ ε l n ɹ w e ɔ m { tʃ ʃg 10.5 136 k j b d h t dZ l n ɹ w e m tʃ ʃk 10.4 203 p ɑ j b d h t dZ ε l n ɹ w e ɔ m tʃ ʃv 10.3 30 j b d h t dZ l n ɹ w m tʃ ʃθ 6.9 36 j b d h t l n ɹ w m tʃ ʃp 6.9 176 i j b d h t dZ l n ɹ w m tʃ ʃɑ 4.5 6i 4.3 6aυ 3.0 6 2.2 6 bu 0.8 1a 0.4 4 bj 0.2 24 l n ɹ w m tʃ ʃb 21.4 219 h ε l n ɹ w m { tʃ ʃd 22.3 112 ε l ɹ w m tʃ ʃh 22.4 115 t ɹ w m tʃ ʃt 22.5 131 ε l n ɹ w m tʃ ʃ2 23.0 6dZ 23.0 39 l ɹ w m tʃ ʃε 23.2 8l 25.7 119 w m tʃ ʃn 25.9 68 m tʃ ʃɹ 26.0 125 w m tʃ ʃw 27.8 112 m tʃ ʃo 28.2 4e 210.3 10ɔ 210.3 6m 211.6 101 ʃ{ 212.3 9tʃ 213.5 45 ʃɔ 214.2 1ʃ 222.8 64

:for the other studies are TMBR 5:0, SB 7:2, and

n

beinioat

es

hend

oldalends,ftert


SW 12:0. Thus there are only two contradictioto the hypothesis in all four megastudies. Thconsistency suggests a real and pervasive that is not very sensitive to differences in expimental setups. For the plosives, this patterncorrelates with the length of burst and aspiratnoise: Velar plosives have the greatest burst aspiration components; labial plosives have least (Klatt, 1975). If voice keys tend to skip ether or both of these components, that could plain why they tend to register velars mo
slowly and labials most quickly. For the frica
sisiasr-gnndhei-x-t

tives, the relative ordering corresponds to taverage duration of the phoneme; Kent aRead (1992) suggested that /ʃ/ and /s/ are longerthan /f/, which is longer than /θ/. One may hy-pothesize that some voice keys have a threshsetting so high that they often skip over initifricatives entirely. This is corroborated by thfindings of Sakuma et al. (1997) and Rastle aDavis (2002), who found that typical voice keyon average, were tripped more than 100 ms athe true beginning of an /s/-initial word; tha


TABLE 7

Differences in Response Time by Initial Phoneme Pairs (SW Study)

Phone RTa N

z 55.6 8 tʃ dZ k f t j d b h ð θ n p l w m ɹa 38.2 4 g k f d b ps 37.9 418 tʃ g dZ k f v t j 2 d ε aυ b h e ɔ ð θ n p l w ɔ m o ɹ {ʃ 34.7 64 g dZ k f v t j ɑ d b h ð θ n p l w ɔ m ɹtʃ 25.0 46 f v t j d b h ð θ n p l w m ɹg 24.1 137 v i j d b h ð θ n p l w m ɹdZ 18.4 39 j d b h ð θ n p l w m ɹk 17.2 205 i j d b h ð θ n p l w m ɹu 15.8 1f 14.3 163 i d b h ð θ n p l w m ɹv 11.8 30 ð θ n p l w m ɹi 10.3 6 b pt 8.6 133 d b h ð θ n p l w ɔ m ɹ 4.4 6 b pj 3.9 24 n p l w m ɹ2 0.6 6 pɑ 20.9 6 bd 21.9 111 ε aυ b h ð n p l w m ɹε 24.7 8 b paυ 24.7 6 b θ pb 27.2 219 n p l w m ɹh 27.6 115 n p l w m ɹe 28.3 11 p {ɔ 212.4 1ð 214.0 13θ 217.0 36 mn 217.7 68p 220.5 175 ɹl 220.9 119w 220.9 112ɔ 221.0 6m 224.0 101o 225.4 4ɹ 228.8 125{ 240.0 10

-length of time corresponds to the entire duration

IC

r s a oot h

s rredhn

d.

ate

fr

d

cu

o

anti

tr

e

ionikethe-but es-

to amece

thew-la-ica-ly.

fferls), less

at-eo-

ru-B

de-

e it Inntly

olarem

ithhe ofs the

indicated feature and there is a significant difference be-tween their response times.

PHONETIC VO

of the phoneme. A prominent exception in omegastudies is in SB, where /ʃ/ is detected fastethan all other phonemes. That is explainablethe fact that /ʃ/ is the loudest of all obstruentand that fricatives, as mentioned before, takeshortest time for speakers to initiate (Sakumal., 1997). The differences in the treatment ofʃ/between the studies could be due to the vkey being set to trip a little lower (or the micrphone being a little more sensitive, or the paripants sitting a little closer to the microphonetalking a little louder) in the SB study than in tother three megastudies.

Manner of articulation was suggested asource of voice key bias by about half of thespondents in our poll, and it was also consideby some researchers in our analysis of publisexperiments. Twelve of the poll responde(20%) suggested that plosives are detecfaster and more reliably than other sounHowever, 4 people offered the opposite ideasecond hypothesis, offered by 16 responde(27%), was that fricatives are detected later,less reliably, than other sounds. The stricway to investigate whether manner is a caufactor is to contrast pairs of phonemes that dionly in manner of articulation. In all foumegastudies, /s/ is detected more slowly thanand in three studies (KTM, SB, SW) we seesignificant effect for /z/ being slower than /That is, fricatives are detected more slowly thplosives, at least for the coronal place of artilation. These results are fairly meager becathere are relatively few minimally contrastinpairs of phonemes one can check in Englishbroader pattern can be discerned if we relaxcriterion a bit and simply ask whether phonemwith different manners of articulation, agroups, differ in response times. For exampfor affricates, we counted all pairs where an fricate phoneme was significantly slower thanonaffricate phoneme, and divided that by number of all pairs where an affricate is signcantly slower or faster than a nonaffricaphoneme. Table 8 summarizes that computaacross seven manners of articulation, and fofour studies. A fairly strong pattern here is thobstruents (plosives, affricates, and fricativ
are detected more slowly than sonorants (nas
edtsteds. Antsndst

salfer

/t/, a/.anu-se

g. Aur

essle,f- ahefi-teion allats)

approximants, and vowels). The only exceptis that in SB, affricates are detected quickly, lsonorants. These ratios corroborate the hyposis that fricatives are detected very slowly,contradict the idea that plosives are detectedpecially quickly.

Physical causes for these patterns have,large extent, already been mentioned. Sotypes of sounds begin with a period of silencorresponding to complete air blockage in vocal tract (plosives, affricates, and often voels), which naturally slows their detection retive to other sounds; at the other extreme, frtives can be produced exceptionally quickAnother consideration is that sound types diin their loudness. Sonorants (especially voweare the loudest of all sounds; obstruents areloud and so may be skipped over entirely.

The inconsistent status of affricates can betributed to the intermediate status of postalvlar consonants in general (/ʃ/ as well as /tʃ/ and/dZ/), which tend to be the loudest of the obstents (Fry, 1979). It would appear that in Salone the voice key was sensitive enough totect them readily, and in the case of /ʃ/ this com-bined with the articulatory advantage to makthe most quickly detected of all phonemes.the other studies, the apparatus was apparenot sensitive enough to detect even postalveobstruents, and so words beginning with th

E KEY BIASES 159

ur

by,the et/ice-ic-ore

ae-

TABLE 8

Effect of Manner of Articulation of Initial Phoneme onResidual Response Time

Manner KTM TMBR SB SW

Affricate .82 .79 .19 .84Fricative .73 .56 .82 .78Plosive .65 .64 .67 .56Nasal .40 .50 .09 .00Vowel .13 — .12 .32Glide .07 .03 .26 .20Liquid .03 .33 .22 .00

Note. Number of phoneme pairs where the phoneme wthe indicated feature is significantly slower than tphoneme lacking the feature, divided by the numberphoneme pairs where exactly one of the phonemes ha

als,were detected slowly.

N

u

l b dtma

tei

n

eflash th

th

d

foeieprddi

ee

thestithif-

tor-

orilstn-

all

ses-

hate

f-tedu-sedsece

vi-ond

nd-

ofeas- theageionialrdsel-la--oci-redy

ol-

a

160 KESSLER, TREIMA

Summary of results for initial phonemes. Theresults clearly show a response time measment bias. The statistical tests showing thatsponse time varies across phonemes even wcovariables are factored out (Table 3) are btressed by finer-grained evidence at the levephoneme pairs (Table 4–7). The differencestween phonemes align along the well-knownmensions of voicing, place, and manner of arulation; it is not the case that every phonebehaves in an idiosyncratic fashion. The pterns are explicable in terms of our prior knowedge of the articulation and acoustics of phonemes. However, the details can vary widacross different phonemes and different studThus, while the fricatives /s/ and /ʃ/ are slow todetect in three of the megastudies, /ʃ/ but not /s/is extremely quickly detected in one of the fostudies (SB). It also appears that there maysubstantial variation beyond that found in ofour megastudies, in that we found no evideof a response time advantage for plosives,that was mentioned by several poll respondeand reported by Connine et al. (1990). Theralso the issue that several different factors inence response time measurements simultously. One cannot assume that by treating,fricatives, differently, one can readily correct tbias problem. Rather, all three dimensions ofticulatory features have an effect, so that, in example, fricatives will differ among themselves depending on their voicing and placearticulation.

Second Phonemes

For the most part, it has been assumedeffects of phonemes on voice key response timare limited to the initial phoneme of the worAs we have mentioned, less than a quarterthe experiments we surveyed included any efto control for the second phoneme. Howevsome recent research shows that second-posphonemes may need to be taken into considtion. Rastle and Davis (2002) reported that tycal voice keys trigger about 10 ms later on wothat begin with /s/ plus obstruent than on worthat begin with /s/ plus vowel. The analyspresented here asks whether this is an isola
problem or whether it is the tip of an iceberg.
, AND MULLENNIX

re-re-henut- ofe-i-

ic-et-l-hely

es.

ur beurceyetnts isu-ne-ay,e

ar-is

-of

ates.ofrtr,

tionra-i-ss

sted

Overall analyses. We performed the samanalyses as with initial phonemes, but this timused the second phoneme of the word as grouping variable and blocked by the firphoneme. That is, we asked whether words wdifferent second phonemes had significantly dferent response time or covariables, after facing out the effect of the first phoneme.

The results in Table 9 are similar to those finitial phonemes (Table 3), though a few detavary. A test of whether words with differenword-second phonemes have different raw (uadjusted) response times is significant infour studies atp , .001. The covariables, byand large, are less plausible as indirect cauthan they were for initial variables. As the superscripts indicate, even those covariables tvaried significantly with the second phonemcannot individually be responsible for the diferences in response time. When we conduca multiple regression to factor out the contribtion of all these covariables from the respontime, we found that words with different seconphonemes differed in their residual respontimes. This difference just reaches significanin SW, but is highly significant (p , .001) in theother three studies. Overall, we have strong edence of a voice key bias caused by the secphoneme of the word.

Analyses comparing pairs of phonemes atheir features. Tables 10 through 13 show figures for the individual phonemes and pairsphonemes. Recall that the response time murements are the average amount by whichresidual response time exceeds the averresidual response time of other second-positphonemes when they follow the same initconsonant. We ran these tests only for wothat begin with a consonant. We omitted vowinitial words from these analyses because retively few monosyllabic words begin with vowels, leaving very little data with which tconduct significance tests. On phonetic prinples, we would expect voice keys to be triggenear the beginning of vowel-initial words in ancase, and therefore to be little affected by flowing phonemes.

Rastle and Davis (2002) found that, with
typical voice key, words that begin with /s/ fol-

a

s

u

n e

un

e

o

s

eave&r,hass-nls-teranr-

/s/isng

ane-de-ssd;ar

ord-

dethrts

b Raw RT not significant when blocked by this covariable.

lowed by /t/ or /p/ were detected slower thwords that begin with /s/ followed by a voweCan their finding be broadened to other connant clusters? A quick look at Tables 10 throu13 reveals that vowels are interleaved with cosonants. Consonants are not clustered at theof the tables, as they would be if consonant clters were consistently measured as beginnlater than single-consonant onsets. When count, for example, the number of times consnants are significantly slower than vowels,striking patterns emerge. Even the positionsparticular consonants are rarely consistacross the studies. The semivowel /j/ makesvery slow response time measurements instudies, but its spelling after an initial consonais so unusual in English (typical spellings of /jencode the consonant and vowel simultaously) that we must reserve the possibility thcognitive factors are at play. The other appromants appear to speed up response time murements in comparison to vowels; but the effdoes not hold for SB. Of the remaining consnants, which appear only after /s/, only /k/ apears to consistently make for slower measuments than for words with a vowel after the /s

What can account for these patterns? Thevious a priori expectation is that if words thabegin with /s/ typically do not trigger voice key
at all, then words that begin with /s/ plus a fo
nl.o-

ghn- tops-

ingweo-oofntforallnt/e-atxi-eas-ct

o-p-re-/.b-t

lowing voiceless plosive (/p/, /t/, /k/) should fareven worse, because those consonants heven less amplitude than /s/ (KawamotoKello, 1999; Rastle & Davis, 2002). Howevethere are compensating factors. Researchshown that both of the initial consonants in cluters are typically quite a bit shorter in AmericaEnglish than they are when followed by vowe(Klatt, 1974; Umeda, 1977). After both elements in a cluster shorten, the /s/-plosive cluscan be almost as short as a single /s/ beforeaverage vowel. Umeda also noted a large diffeence in her measurements of the length ofbefore /p/ and those of Schwartz (1970). Itconceivable that this variation has some bearion the fact that /sp/ behaves differently in SWthan in our other studies. Other factors that caffect the outcome of consonant cluster rsponse time measurements include a partialvoicing of sonorant consonants after voiceleconsonants, making the vowel much less louand a retraction of /t/ and /d/ to the postalveolregion before /ɹ/, making them louder. As forinitial consonants, it is easy to see how mindifferences in acoustic thresholds between stuies and minor differences in speech amplituand phoneme timings between subjects (boidiosyncratic and dialectal) could account fomajor differences in how particular consonan


TABLE 9

Significance of Differences in Response Variables Caused by Second Phoneme of Word

Response KTM TMBR SB SW

Raw RT .000 .000 .000 .000

Bigram frequency .029b .001b .115a .105b

Consistency of onset .000a .800a .000a .000a

Consistency of rime .000a .000a .000a .000a

Familiarity .413a .714a .652a .551a

Frequency of spelling .236b .184b .709a .701a

Homophones .000a .000a .000a .000a

Length of pronunciation .000a 1.000a .000a .000a

Length of spelling .000a .000a .000a .000a

Neighborhood size .000a .000a .000a .000a

Error rate .000a .003a — .193a

Residual RT .000 .000 .000 .050

a Raw RT remains significant (p # .05) when blocked by this covariable.

l-behave in different studies. In addition, evi-

r

e&

rth tee

no

r

re-ntrein andmsanbe-dif-elsasne

wes,erels,pairall

cond col-

a Response time residuals (milliseconds) after effect of covariables is removed by regression.

dence has recently been adduced that wowith initial clusters are uttered sooner thawords with no clusters, apparently for cognitivreasons (Kawamoto & Kello, 1999; RastleDavis, 2002).

Perhaps the most intriguing pattern to emefrom Tables 10–13 concerns the vowels. Litresearch using voice response latencies mentioned the possibility that response timewords can be biased by the identity of an inrior vowel. The published research on voice kbiases has not looked at interior vowels. Whilminority of our poll respondents were willingin principle, to entertain the possibility thavowels could have an effect, they offered concrete hypotheses about which vowel factcould be involved.

It is obvious that there are significant diffe
ences between the vowels. In SW, there are
dsn

geleas

ofe-y a,tors

-

different vowels that are associated with sponse times that are significantly differefrom some other vowel or vowels; in SB, theare 8; in KTM and TMBR, 9. The differences residual response time between the fastestslowest vowel range from 12 ms in SB to 67 in TMBR. Thus the systematic differences cbe large, depending on factors that vary tween experiments. One way to explore the ferences between vowels is to break the vowdown by their major articulatory dimensions,was done earlier for the initial consonants. Oimportant dimension is front versus back. If look only at vowels that differ only in frontnesin no case do we find back vowels with slowmeasured response times than front vowwhereas all the studies have at least one where the front vowel is slower. Counting


TABLE 10

Differences in Response Time by Second Phoneme Pairs (KTM Study)

Phone RTa N

j 41.4 21 i u aυ a e o ɑ ɔ { ɔ ε 2 Ã ɹ υ li 16.9 142 aυ a e o ɑ ɔ { ɔ ε 2 Ã ɹ υ lu 16.9 94 a e o ɑ { ɔ ε 2 Ã ɹ lp 14.1 70 aυ a o { ɔ ε t l mf 12.7 2 10.8 225 o ɑ { ɔ ε 2 Ã ɹ υ laυ 7.3 42 k n ɔ Ã ɹk 6.4 74 a o ɔ ε t l ma 4.0 127 n ɑ ɔ t Ã ɹ υ le 3.6 160 ɑ Ã ɹ ln 1.0 25o 0.1 132 ɹ lɑ 0.1 140 { ɹɔ 20.6 21 l{ 22.2 186 Ã lɔ 22.4 149 wε 23.2 182 ɹt 23.9 1072 24.5 77Ã 25.3 180 ɹɹ 26.8 336υ 27.2 32l 212.1 253m 214.8 18w 223.0 77

Note. Entries tell which phonemes have voice key response times significantly faster than the phoneme in the seumn,p # .05. Effect of first phoneme is blocked.

7pairs that differ in frontness and significantly

4ecnn

tee

ii

or

InI

l

eo

pcao

isme,

ithandhanwe

rehc-er-il-sg-ly

atheindlse-Thencyelsn-ey


differ in response time, we find in KTM 1vowel pairs that show the front vowel as slowthan the back vowel; only 4 times are bavowels slower (14:4, front vowel is slower i78%). In TMBR, the ratio is 24:0 (100%); iSB, 12:6 (67%); in SW, 16:0 (100%). There isclear effect whereby front vowels are detecmore slowly, although the effect is a bit weakin SB.

The other important dimension in vowels their height. Among contrasting vowel pairs,KTM we find 24 cases where a high vowel is asociated with a slower response time than a nhigh vowel and only one case where the reveis true (24:1, high vowel is slower in 96%). TMBR, the quantities are equal, 6:6 (50%). SB, the numbers are 27:0 (100%); in SW, 16(94%). If we look at the four pairs that clearcontrast minimally (/i/-/e/, //-/ε/, /u/-/o/, /υ/-/ɔ/), SB has all four pairs with the high vowsignificantly slower; KTM and SW have three the four in that direction; TMBR, however, hanone in that direction, but it does have the cotrary /ε/ . //. With the rather mysterious excetion of TMBR, therefore, there is a strong effewhereby words with a high vowel after an initiconsonant have measured response times sl
than analogous words with a nonhigh vowel.
rk

adr

sns-n-se

n:1y

lfsn--tlwer

Thus we find pervasive evidence that therea response time bias due to the second phonesuch that front vowels are associated wgreater response times than back vowels,high vowels have greater response times tlower vowels. To the above-stated evidence might mention that /i/ is slower than // (signifi-cantly so in three of the studies); while both afront and high, /i/ is even more front and higthan //. Recall also that the semivowel /j/ as seond element slows the response time considably. Although we cannot discount the possibity that its odd spelling is a factor, it inoteworthy that /j/ is the only consonant in Enlish that is front and high and is in fact virtualidentical in articulation to /i/.

Are there articulatory or acoustic facts thcould account for the differences among tvowels? Articulatory biases seem unlikely light of the preliminary report of Rastle anDavis (2000) that words with internal vowediffering in frontness did not have different rsponse times when measured by waveform. effect may be due in part to a general tendeto lengthen consonants before high vow(Klatt, 1975; Schwartz, 1970). If the onset cosonant is so low in amplitude that the voice k


TABLE 11

Differences in Response Time by Second Phoneme Pairs (TMBR Study)

Phone RTa N

{ 33.0 128 i a ε ɑ e u Ã 2 o ɔ ɔ υaυ 30.6 23 ɑ e u Ã 2 o ɔ ɔ υi 20.0 110 u Ã 2 o ɔ ɔ υa 10.7 91 Ã o ɔ ɔε 5.3 95 Ã 2 o ɔ ɔ υɑ 4.5 76 2 o ɔ ɔ υe 1.1 123 Ã 2 o ɔ ɔ υb 0.0 1 26.5 140 ɔ υu 27.5 66 υÃ 213.0 1172 215.7 61o 219.0 88ɔ 219.6 75ɔ 231.9 11υ 234.4 29

sometimes misses it altogether, or if the time it

pitththol

aattl

-rs

re-ualre-reden

d-ig-irs.orynsta

tic-tis-

aveex-udyninerbe


takes the consonant to reach sufficient amtude to trigger the voice key is proportional to length, then the longer the consonant,greater the measured response time. Anopossible explanation involves the amplitude the vowel itself. Front vowels have less amptude than back vowels, and high vowels haless amplitude than low vowels (Fry, 1974).the voice key misses the initial consonant etirely and has to contend with the vowel, it mtake longer to detect weak vowels like /i/ thlouder vowels like /ɑ/. It is also possible thathrough coarticulation, the initial consonanthemselves may in part anticipate the articutory gestures of the vowels that make for damened amplitude.

Summary and implications of results for seond phonemes. The phonemes that follow wordinitial consonants have an effect on measuresponse time. General statistical tests acros
phonemes (Table 9) indicated that words wi
li-seerf

i-veIfn-yn

sa-p-

c-

ed all

different second phonemes had different sponse times, even when blocked by individcovariables. They also had different residual sponse times after all covariables were factoout. When we looked at the contrasts betweindividual pairs of phonemes occurring in worsecond position (Tables 10–13), we found snificant differences between many of the paThose differences are aligned along articulatdimensions that provide coherent explanatiofor most of the effects. The fact that our damake sense in terms of prior knowledge of arulatory and acoustic phonetics makes the statics especially convincing.

Could the second-phoneme effects we huncovered here have an impact on current perimental paradigms? Consider a naming streported by Peereman, Content, and Bo(1998; Experiment 1b). They asked whethprinted words containing sounds that can


TABLE 12

Differences in Response Time by Second Phoneme Pairs (SB Study)

Phone RTa N

j 17.6 18 i υ u a ɔ e l ɹ aυ ε ɑ o { 2 Ãf 12.2 2 op 8.3 53 w k ɔ l m n ε t o 2 Ãi 7.5 123 u ɔ e ɹ aυ ε ɑ o { 2 Ãw 6.9 58 k ɔ l ɹ o Ãυ 6.0 29 ɔ e ɹ aυ ε ɑ o { 2 Ãu 4.2 76 ε ɑ o { 2 Ãa 4.0 114 ε ɑ o { 2 Ãk 2.5 56 ɔ t o 2.3 193 aυ ε ɑ o { 2 Ãɔ 1.2 130 ɑ { Ãe 1.2 137 ε ɑ { Ãl 1.1 199 ɹ aυ ɑ o { Ãɹ 0.9 285 ɑ { Ãm 0.8 13 on 20.3 16 oaυ 20.3 39ɔ 20.8 19ε 22.1 164ɑ 22.9 117t 22.9 96o 23.3 118 Ã{ 23.7 1672 23.8 68Ã 24.4 162

thspelled in many different ways (feedback incon-

rath

rni

eoah

lhvn

uswng-

ny theme,nsig-d-rdsechetes

ntalrk

ses


sistent words) cause more difficulty in an oreading task than words that contain sounds are usually or always spelled the same w(feedback consistent words). The study was cried out in French. By way of example, the wogland /glɑ/ ‘acorn’ was considered inconsistebecause /ɑ/ has at least 13 different spellings French, whereas piste /pist/ ‘track’ was consid-ered consistent because the constituent souare rarely spelled otherwise than as in piste it-self. If there are feedback effects, one would pect the inconsistent words to be read mslowly and with more errors. Peereman et found that errors were marginally greater for tinconsistent words, but that the 6-ms increaseresponse time was not significant. Their concsion from this and other experiments was tfeedback consistency cannot be shown to hareliable effect on performance. However, mamore of the consistent words than the incons
tent words contained high front vowels an
lat

ayar-dtn

nds

x-rel.e inu-ate ayis-

glides in word-second position (39% vers11%). If high and front vowels and glides slomeasured response times in French, as in Elish, then that imbalance would counter afeedback consistency advantage. Perhaps ifstimuli had been equated for second phonethe difference between the two conditiowould have been of greater magnitude and snificance, supporting the conclusion that feeback inconsistency does play a role in wonaming. We mention this experiment becauthe authors were extremely careful to mattheir stimuli across many variables, and ythere is an imbalance in word-second phonemthat could lead to a bias against the experimehypothesis. This is a new issue that future woin this paradigm will need to correct for.

Possible Solutions and Recommendations

We have documented large phonetic bia


TABLE 13

Differences in Response Time by Second Phoneme Pairs (SW Study)

Phone RTa N

j 27.5 17 i aυ e ɑ u o ε { Ã ɔ l 2 ɔ ɹ wf 24.8 2 wi 14.8 123 e ɑ u o ε { Ã ɔ l 2 ɔ ɹ wk 14.2 56 a e ɑ n ε { ɔ l p w ma 9.6 114 t ɑ u n o ε { Ã ɔ l 2 ɹ p 9.5 194 o ε { Ã ɔ 2 ɔ ɹυ 8.8 29 Ã ɔ l ɹ wt 7.3 96 ɑ ε { ɔ p w maυ 5.8 39 ɔ ɹ we 5.4 138 ε Ã ɔ ɔ ɹɑ 4.6 116 ɹu 2.0 76 ɹn 20.4 16 wo 20.5 121 ɹε 20.6 163 ɹ{ 22.7 168 ɔ ɹÃ 23.8 162 ɹɔ 23.9 131l 24.3 199 ɹ2 24.8 69ɔ 26.5 19ɹ 27.4 287p 28.7 55 ww 215.5 58m 219.0 13

dthat come into play when voice keys are used to

te

W

rc

ec

apm

-eric nntfi

aara.g

ormi

tmsfos

af-o-ds

er-rdteneyofuld

Sot ace nous

eirin-dt-

ras-estwas-

odre-ns,eyar-

,iashetheat av-thech-

7),97). and

beddyhe

166 KESSLER, TREIMAN

measure vocal response times. These biaalign along several phonological dimensionextend at least two phonemes deep into word, and are very sensitive to differences in perimental procedures. How can these problebe dealt with? There are several alternatives.will discuss the solutions roughly in increasinorder of usefulness, but all have their pros acons.

First let us briefly consider the possibility ogetting around the issue by not using voice sponse times at all. For example, lexical desion tasks measure how fast it takes a participto decide whether a stimulus is a word and prthe appropriate key. However, the lexical desion task is often viewed as a less “clean” tathan naming, because in addition to a lexical cess component, it includes a decision comnent, which may add bias to response timeasurements (Balota & Chumbley, 1984; Morison & Ellis, 1995). Lexical decision performance also varies with such things as the naturthe nonword foils (e.g., Lewellen, GoldingePisoni, & Greene, 1993). Another way to avoor supplement voice response times is to foon error rates or error types (e.g., AndrewsScarratt, 1998; Laxon, Masterson, & Mora1994). However, when studying adults’ readiof familiar words, it may be difficult to elicienough errors to make for statistically signicant differences, unless special measures taken (Kello & Plaut, 2000). Because alterntives to voice response time measurement hlimitations of their own, many researchechoose to use a combination of these proaches and look for converging results (eConnine et al., 1990; Wurm & Ross, 2001).

Another lateral solution is to adopt a protocwhereby voice response times are compaonly against different utterances of the saword. In many types of research, the crucial dference is not in the stimuli themselves, butsome other experimental factor such as prime that precedes the stimulus. If one copares the response times to multiple instanceexactly the same word, there is less room phonetic bias to enter. Even here, however,may need to beware of possibilities such a
prime’s causing a response to be spoken more
, AND MULLENNIX

sess,hex-ms

egnd

fe-i-

antssi-skc-o-e

r-

of,dus&,g

-are-vesp-.,

lede

f-inhe

- oforne a

less loudly, or more or less fast. That could fect the amplitude and duration of initial consnants that may fall below detection thresholfor some voice keys.

If one cannot get around the need to detmine vocal response times for different wotypes, one optimistic idea would be to eliminauncertainty by strictly controlling the setup ivoice key experiments. Lowering the voice ktrigger until one cannot tolerate the number false starts caused by nonspeech noise woclearly reduce much phonetic variance. would encouraging participants to speak aconstant loud amplitude, at a fixed distanfrom the microphone, and to make absolutelynonverbal noise. It would also be advantageoto encourage participants to pronounce thwhole response as quickly as possible, thus mimizing the duration of utterance-initial souncomponents that fall below the voice key cuoffs. It is difficult to infer on the basis of foustudies just how far one can go with such meures. We might note, however, that the smallrange in phonetic bias between phonemes 41 ms, in SB (for initial phonemes, ignoring uncommon ones such as /z/), which is still a gochunk of time. Also, there is a limit to how faone can go with forcing all participants to bhave ideally. Even under the best of conditiowith very loud speakers and very low voice kthresholds, some phonemes will reach that tget amplitude faster than others.

Another idea, attractive in its simplicitywould be to compensate numerically for the bintroduced by voice keys. After measuring tresponse time to a word that begins with /s/,experimenter would look in Table 4, note thwords in /s/ are detected 39 ms later than theerage word, and so subtract 39 ms from measured response time. This is one of the teniques considered by Kondo and Wydell (199using the Japanese data of Sakuma et al. (19Of course one would offset that number byvalue corresponding to the effect of the secophoneme as well. Such an approach mightbetter than doing nothing, but it is not gooenough. The numbers that obtain in one studo not necessarily apply to a different study; t
ordifferences range from the trivial to the alarm-

IC

bnl ffmth

lign

ndane)

vrdr aseroci

imyiathim

lmecnaam

th a

thb

den

ate-ro-ti-t-e-ef-re

ener-l-edge-r-i-

rs,

-p-

hate-im-l-e

c-s-

toatairelyi-or-reldceu-d

mene,w

PHONETIC VO

ing. Numerical compensation could conceivawork if the computation for the correction icluded factors for voice key threshold, amptude of the beginning of the utterance, andforth. However, as soon as one goes to the eof measuring amplitude, one might as well copute the entire waveform, which obviates need for a voice key in the first place.

The delayed naming paradigm is a popuway of addressing phonetic bias. This paradwas introduced by Eriksen, Pollack, and Motague (1970). In addition to asking participato read a word as quickly as possible immeately after it was presented (immediate, or stdard, condition), in later sessions participapronounced the same word after a variable dof from 1 to 2 s (delayed naming conditionPresumably, 1 s should be enough time for ethe most difficult of words to be read. Therefoin the delayed condition, no response time ferences should be attributable to lexical trieval processes in the broad sense. Thoseferences should show up only in the immedicondition. In contrast, acoustic voice key biashould contribute equally to the measured sponse time under both conditions. Therefthe response times in the delayed condition be treated statistically as equivalent to the nocomponent of the response times from the mediate condition. Eriksen et al. found that slable count (their variable of interest) had a snificant effect on measured response time that there was a significant interaction such syllable count was more important under the mediate naming condition. Under the assumtions just outlined, that analysis convincingindicated that syllable count affects naming tiin a way that cannot be explained by phonbiases alone. Another statistical approawhich does not require factorial analysis awhich may be just as convincing, is to subtrthe delayed response time for each participfrom the immediate response time for the saparticipant and perform all computations on difference (e.g., Fushimi et al., 1999; LangeContent, 1999). Delayed naming is better ththe idea of subtracting standardized offsets was mentioned in the previous paragraph,
cause the compensating numbers are taken fr
E KEY BIASES 167

ly-i-soort-e

arm-

tsi-n-tslay.ene,if-e-dif-tes

e-reansy-

l-g-ndat-

p-ye

tich,dctnte

e&nate-

the experiment itself and so will not be polluteby possible systematic differences betwevoice keys or between participants.

There are, unfortunately, several things thcan go wrong with the delayed naming procdure. Researchers sometimes simplify the pcedure to the extent that it is no longer statiscally convincing. Often they simply show thathe variable of interest significantly affects response time in the immediate but not in the dlayed condition, but neglect to show that thdifference between the two conditions is signiicant. Also, many researchers may not be awaof the great variance in voice key readings evfor the same speaker producing the same uttance (Pechmann et al., 1989). To our knowedge, all researchers take only a single delaynaming trial for each word instead of averaginover several trials. In some instances the dlayed naming data are taken from different paticipants from those who produced the immedate naming data (e.g., Forster & Chambe1973).

A much more important problem with the delayed naming corrective is that its core assumtion may be wrong. The technique assumes tall word-specific biases that show up in the dlayed condition are the same as those in the mediate condition. If that provision is not fufilled, then, while the technique is removing onsource of bias with the right hand, it is introduing a new source of bias with the left hand. Posible sources of such a problem are easyimagine. For example, the delayed naming dmay be gathered under conditions that inspthe participants to speak more loudly or softthan for the immediate naming trials. Particpants who must hold a word for a second more before uttering it may well get their articulators in position to say the word before they aasked to utter it. If so, delayed naming wounot erase articulatory biases, and may introdua new source of problems, through hyperarticlation. A more insidious case is if the delayenaming task itself is biased with respect to sopsycholinguistic variable, perhaps the very othat is being studied. Goldinger, AzumaAbramson, and Jain (1997) showed that lo
omword frequency inhibits delayed naming. Using

tbe

rtu

i r

v

raw

,ot

x

m

t he

ta

er

isnl

cte isbe- of be oftsve-notorrtic-ble ag ang

de-ita-estiontalsesis

en-r.

italig-ay

ig-ses

tchngononif-megh aasb-ra-ofgowre-ddi-


frequency-biased measurements as a correcto one’s primary data would introduce those ases into the data. Admittedly, the frequency fect in delayed naming is weak and may not,practice, do much harm. However, until we asure that the delayed naming task does not induce any measurement bias of its own, the of the delayed naming procedure in a test of stistical significance may be questioned.

Another solution available for many experments, especially those with a small numbertrials or a large amount of funding, is to recospoken responses and analyze them using mern digital processing techniques such as waform inspection. There is no question but thsuch techniques permit a much more accudetermination of response time. The main draback of the voice key is that the experimenterorder to eliminate false readings caused by nspeech noise, must set the key to ignore a cerlevel of sound. Transient noise is not as bigproblem when using a digital recording: The eperimenter can generally differentiate speefrom noise by visual inspection of the waveforor by playing back parts of the recording that ain doubt. The use of visual waveforms greareduces the variance and therefore greatlycreases the power of studies. Some researchave abandoned voice keys, especially in recyears (e.g., Fushimi et al., 1999; KawamotoKello, 1999), although a large majority of researchers still uses voice keys. A dividend waveforms and digital signal processing is ththese techniques permit one to study variabother than time to initiate response, such asduration of the response and of its individuparts.

Despite their advantages, digital techniqudo not solve all problems. For one thing, diffeentiating noise from speech is not always easy as it would seem in theory. The beginninof utterances tend to be very soft: If there is abackground noise at all, it can be challengingpinpoint exactly where speech begins. Also, dital techniques do not address articulatory biaat all. Fricatives are produced faster than sorants, for example (Sakuma et al., 1997). Psives and affricates (typically including eve
voiced ones in English) begin with periods of s
, AND MULLENNIX

ioni-f-inero-seta-

-ofdod-e-

atte-

inn-ain a-

ch

relyin-ersnt

&-ofatleshel

s-

asgsnytog-eso-o-n

lence, of varying durations, as the vocal trabuilds up pressure for a noisy release. Therno way that digital techniques can detect the ginning of a silent articulatory gesture, unlesscourse the default state of the speaker is tonoisy, as in the postvocalic naming techniqueKawamoto et al. (1998), where participandrone a schwa sound until they utter a plosiinitial response; but such a technique could be used uniformly with all phonemes. FJapanese, at least, Sakuma et al. report an aulatory bias range of about 40 ms, comparato the voice key phonetic bias in SB. Fromqualitative standpoint, we are simply tradinone set of biases for another: It is not much ofimprovement to have fricatives go from beindetected later than other phonemes to beingtected earlier than other phonemes. Quanttively, it must be admitted that digital techniquare superior: Voice key biases are a combinaof the articulatory biases that also affect digitechniques, plus another set of acoustic biaunique to voice keys. Thus the variance greater, and the magnitude of the bias is pottially (and, in most studies, actually) greateHowever, it should not be assumed that digtechniques would allow the experimenter to nore the effect that different phonemes mhave on measured response time.

Even if one uses the most accurate digital snal processing techniques, measurement biawill still exist, for which one must apply stricstatistical controls. The most reliable of sucontrols is to block the data by the leadiphonemes. For example, in an experiment the effect of spelling-to-sound consistency naming latency, one would want to measure dferences only between words that have the sabeginnings. This is already best practice, thouour survey of the literature showed that onlyslim majority of the studies (56%) matched much as the initial phoneme, even in the asence of controls like the delayed naming padigm. We have confirmed here the wisdom controlling for the initial phoneme by showinhow pervasive the acoustic differences are, hgreat their magnitude can be, and how unpdictable they can be across experiments. In a
i-tion, we have demonstrated that the second

C

t

g

u o

an

o

a

bti

t

tr

81)e-ete andhen,arengustle

aourkeyari-s is

that there are several correctives available to

gypel:

.tic

i- of

n

in

ns.-

D.ic

.d

ui-tic

).g.

PHONETIC VOI

phoneme can have a significant effect. In special case when the first two phonemes both obstruents (/s/ followed by a plosive), thechoes and corroborates the recent findingRastle and Davis (2002). However, our findingreatly broaden our awareness of the scopethe problem. All sorts of second-position consnants can cause differential voice key effecand, surprisingly, even the identity of a seconposition vowel makes a difference. Thereforeis necessary to equate words at least throtheir first vowel. This is currently done in onlysmall minority of studies. It is not sufficient tmatch by broad class of phonemes, becausehave seen that virtually all the main featuresphonemes (manner, place, and voicing of connants; height and frontness of vowels) haveeffect; balancing words depending on whethor not they were fricatives, for example, woustill leave open the possibility of biases causby voicing or place of articulation. Besidesome of the details vary from study to studyunpredictable ways, so that phonemes that have the same in one study may behave cpletely differently in another.

Our exhortation to “match” the entire head stimuli has been intentionally vague, becauthere are many ways to do so. In a small stuwhere the stimuli are carefully selected for bance, one will want to select words in advanso that sets vary on a property of interest,have the same beginnings, and match on oconfounding variables as well. In an exhaustmegastudy, the matching of items would occafter data gathering, in the analysis phase. Oapproach then would be to compute significanof the variable of interest by Monte Carlo tesblocking by the head of the stimuli along wiother covariables. Another approach, alreaemployed by some researchers for the iniphoneme (e.g., Connine et al., 1990; SpieleBalota, 1997; Treiman et al., 1995), is to analythe data by regression analyses, first factorout variance that can be accounted for by head of the stimuli.

We do not wish to seem blithe about our reommendations. Frankly, blocking stimuli so thentire heads match can be difficult, especia
given the need to match stimuli on other va
E KEY BIASES 169

heareis ofs ofo-ts,d- itgh

a

weofso-

ables as well. Some 20 years ago, Cutler (19lamented that it was extremely difficult to dsign experiments with well-matched stimuli; thaddition of a new variable on which to equathem makes the task even harder. The onsetvowel carry a disproportionate amount of tdifference between words (Kessler & Treima1997); there are often very few words that shboth. Finding suitable words can be frustratiat best or impossible at worst when one malso manipulate the psycholinguistic variabbeing studied, not to mention controlling forhost of potential covariables. Nevertheless,study of megastudies has shown that voice bias is so large, pervasive, systematic, and vable that it cannot be ignored. The good new

erldeds,inbe-m-

ofsedyl-

ceutherveurnece

ts,hdyial &zeingthe

c-atlly

cautious researchers.

REFERENCES

Andrews, S., & Scarratt, D. R. (1998). Rule and analomechanisms in reading nonwords: Hough dou pearede gnew wirds? Journal of Experimental PsychologyHuman Perception and Performance, 24,1052–1086.

Bachoud-Lévi, A.-C., Dupoux, E., Cohen, L., & Mehler, J(1998). Where is the length effect? A cross-linguisstudy of speech production. Journal of Memory andLanguage, 39,331–346.

Balota, D. A., & Chumbley, J. I. (1984). Are lexical decsions a good measure of lexical access? The roleword frequency in the neglected decision stage. Jour-nal of Experimental Psychology: Human Perceptioand Performance, 10,340–357.

Bates, E., Devescovi, A., Pizzamiglio, L., D’Amico, S., &Hernandez, A. (1995). Gender and lexical accessItalian. Perception & Psychophysics, 57,847–862.

Cedrus Corporation (2000). SuperLab Pro Input Optio[On-line]. Available: http://www.superlab.com/pro/inputoptions.htm.

Coltheart, M., Davelaar, E., Jonasson, J. T., & Besner,(1977). Access to the internal lexicon. In S. Dorn(Ed.), Attention and performance VI(pp. 535–555).Hillsdale, NJ: Erlbaum.

Connine, C. M., Mullennix, J., Shernoff, E., & Yelen, J(1990). Word familiarity and frequency in visual anauditory word recognition. Journal of ExperimentalPsychology: Learning, Memory, and Cognition, 16,1084–1096.

Cutler, A. (1981). Making up materials is a confounded nsance; or, Will we be able to run any psycholinguisexperiments at all in 1990? Cognition, 10,65–70.

Eriksen, C. W, Pollack, M. D., & Montague, W. E. (1970Implicit speech: Mechanism in perceptual encodin

ri- Journal of Experimental Psychology, 84,502–507.

l

).m-

P

-i

:

s

se:

t

n

in -

f

ns

/

e

.

-

n

en

f

.

fg-

,

d

a

r-e

n in

?o-

f


Forster, K. I., & Chambers, S. M. (1973). Lexical access anaming time. Journal of Verbal Learning and VerbaBehavior, 12,627–635.

Fry, D. B. (1979). The physics of speech. Cambridge: Cam-bridge University Press.

Fushimi, T., Ijuin, M., Patterson, K., & Tatsumi, I. F. (1999Consistency, frequency, and lexicality effects in naing Japanese kanji.Journal of Experimental Psychology: Human Perception and Performance, 25, 382–407.

Goldinger, S. D., Azuma, T., Abramson, M., & Jain,(1997). Open wide and say “Blah!” Attentional dynamics of delayed naming. Journal of Memory and Lan-guage, 137,190–216.

Good, P. (1995). Permutation tests, 2nd corrected printing.New York: Springer.

Griffin, Z. M., & Bock, K. (1998). Constraint, word frequency, and the relationship between lexical processlevels in spoken word production. Journal of Memoryand Language, 38,313–338.

Hutzler, F. (1999). Voice-key [On-line]. Availablehttp://www.sbg.ac.at/psy/people/hutzler/voicekey.htm

International Phonetic Association (1996). Reproduction ofthe international phonetic alphabet[On-line]. Avail-able: http//www2.arts.gla.ac.uk/IPA/ipachart.html.

International Phonetic Association (1999). Handbook of theInternational Phonetic Association: A guide to the uof the International Phonetic Alphabet. Cambridge,England, Cambridge University Press.

Kawamoto, A. H. & Kello, C. T. (1999). Effect of onset cluter complexity in speeded naming: A test of rule-basapproaches. Journal of Experimental PsychologyHuman Perception and Performance, 25,361–375.

Kawamoto, A. H., Kello, C. T., Jones, R., & Bame, K(1998). Initial phoneme versus whole-word criterioninitiate pronunciation: Evidence based on responsetency and initial phoneme duration. Journal of Experi-mental Psychology: Learning, Memory, and Cognitio,24,862–885.

Kello, C. T., & Plaut, David C. (2000). Strategic control word reading: Evidence from speeded respondingthe tempo-naming task. Journal of Experimental Psychology: Learning, Memory, and Cognition, 26, 719–750.

Kent, R. D., & Read, C. (1992). The acoustic analysis ospeech. San Diego, CA: Singular.

Kessler, B., & Treiman, R. (1997). Syllable structure athe distribution of phonemes in English syllableJournal of Memory and Language, 37, 295–311.Available on-line: http://www.idealibrary.com/linksdoi/10.1006/jmla.1997.2522.

Kessler, B., & Treiman, R. (2001). Relationships betwesounds and letters in English monosyllables. Journal of Memory and Language, 44, 592–617. Availableon-line: http://www.idealibrary.com/doi/10.1006/jmla2000.2745.

Klatt, D. H. (1974). The duration of [s] in English wordsJournal of Speech and Hearing Research, 17,51–63.

, AND MULLENNIX

nd

-

.-

ng

.

e

-d

.ola-

in

d.

n

.

Klatt, D. H. (1975). Voice onset time, frication, and aspiration in word-initial consonant clusters. Journal ofSpeech and Hearing Research, 18,686–706.

Klatt, D. H. (1976). Linguistic uses of segmental duration iEnglish: Acoustic and perceptual evidence. Journal ofthe Acoustical Society of America, 59,1208–1221.

Kondo, T., & Wydell, T. N. (1997, December). Nature ofnaming latency for Japanese kanji words. Paper pre-sented at the International Conference on CognitivProcessing of Asian Languages & Symposium oBrain, Cognition, and Communication (ICCPAL97),Nagoya, Japan. Available on-line: http//www.brl.ntt.co.jp/people/kondo/wydell/index.html.

Lange, M., & Content, A. (1999, June).The grapho-phono-logical system of written French: Statistical analysis andempirical validation. Paper presented at the meeting othe Association for Computational Linguistics, Univer-sity of Maryland. Available on-line: http//homepagesulb.ac.be/,mlange/acl/LangeContentACL99Submitted.html.

Laxon, V., Masterson, J., & Moran, R. (1994). Are chil-dren’s representations of words distributed? Effects oorthographic neighbourhood size, consistency and reularity of naming.Language and Cognitive Processes,9, 1–27.

Lewellen, M. J., Goldinger, S. D., Pisoni, D. B., & GreeneB. G. (1993). Lexical familiarity and processing effi-ciency: Individual differences in naming, lexical deci-sion, and semantic categorization. Journal of Experi-mental Psychology: General, 122,316–330.

Morrison, C. M., & Ellis, A. W. (1995). Roles of word frequency and age of acquisition in word naming anlexical decision. Journal of Experimental Psychology:Learning, Memory, Cognition, 21,116–133.

Nusbaum, H. C., Pisoni, D. B., & Davis, C. K. (1984).Sizing up the Hoosier mental lexicon: Measuring thefamiliarity of 20,000 words, Research on Speech Per-ception Progress Report 10. Bloomington, IndianUniversity.

Pechmann, T., Reetz, H., & Zerbst, D. (1989). Kritik eineMeßmethode: Zur Ungenauigkeit von Voice-key Messungen. [Critique of a method of measurement: On thunreliability of voice-key measurements.] Sprache &Kognition, 8, 65–71.

Peereman, R., Content, A., & Bonin, P. (1998). Is perceptioa two-way street? The case of feedback consistencyvisual word recognition. Journal of Memory and Lan-guage, 39,151–174.

Psychology Software Tools (2000).PST serial responsebox[On-line]. Available: http://www.pstnet.com/srbox/srb.htm.

Rastle, K., & Davis, M. H. (2000, November). On the com-plexities of naming: What do voice keys measurePaper presented at the meeting of the Psychonomic Sciety, New Orleans.

Rastle, K., & Davis, M. H. (2002). On the complexities omeasuring naming. Journal of Experimental Child Psy-chology: Human Perception and Performance,28,
307–314.


Sacia, C. F., & Beck, C. J. (1926). The power of fundamen-tal speech sounds. Bell System Technical Journal, 5,

n

scio

c

A

d

edt

Solso, R. L., & Juel, C. L. (1980). Positional frequency andversatility of bigrams for two- through nine-letter Eng-

n-

-el.

n-:

- ing-l

ng-

e-se-

.

393–403.Sakuma, N., Fushimi, T., & Tatsumi, I. (1997). Onseiha

shisatsu ni yoru kana no ondokusenji no sokutei: Odokusenji wa gotoon no choonho ni yori okiku ko-tonaru. [Naming latency measurements of kana baon inspection of voice waveforms: Naming latenvaries greatly depending on the manner of articulatof the initial phoneme.]Shinkeishinrigaku, 13, 126–136.

Scarborough, D. L., Cortese, C., & Scarborough, H.(1977). Frequency and repetition effects in leximemory. Journal of Experimental Psychology: HumaPerception and Performance, 3, 1–17.

Schneider, W. (1988). Micro Experimental Laboratory:integrated system for IBM PC compatibles. BehaviorResearch Methods, Instruments, and Computers, 20,206–217.

Schwartz, M. F. (1970). Duration of /s/ in /s/-plosive blenJournal of the Acoustical Society of America, 47,1143–1144.

Seidenberg, M. S., & Waters, G. S. (1989, NovembReading words aloud: A mega study. Paper presenteat the meeting of the Psychonomic Society, AtlanGA.

on-

edyn

S.aln

n

s.

r).

a,

lish words.Behavior Research Methods & Instrumetation, 12, 297–343.

Spieler, D. H., & Balota, D. A. (1997). Bringing computational models of word naming down to the item levPsychological Science, 8, 411–416.

Stanovich, K. E., & West, R. F. (1983). On priming by a setence context. Journal of Experimental PsychologyGeneral, 112,1–36.

Treiman, R., Mullennix, J., Bijeljac-Babic, R., & Richmond-Welty, E. D. (1995). The special role of rimesthe description, use, and acquisition of English orthoraphy. Journal of Experimental Psychology: Genera,124,107–136.

Umeda, N. (1977). Consonant duration in American Elish. Journal of the Acoustical Society of America, 61,846–858.

Wurm, L. H., & Ross, S. E. (2001). Conditional root uniquness points: Psychological validity and perceptual conquences.Journal of Memory and Language, 45,39–57.

Zeno, S. M., Ivenz, S. H., Millard, R. T., & Duvvuri, R(1995). The educator’s word frequency guide. Brew-ster, NY: Touchstone Applied Science Associates.

(Received February 8, 2001)(Revision Received May 18, 2001)

Documents

Phonetic Biases in Voice Key Response Time Measurements€¦ · A voice key is a device that determines voice onset in real time. The voice key is connected to a microphone, which