Contextual effects in vowel perception II evidence for two processing mechanisms

In the present studies, anchored ABX discrimination functions and signal detection analyses of identification data Ibefore and after anchoring} for an [i]-[I] vowel series were used to d

1980, Vol 27 (5), 421-434

Contextual effects in vowel perception II:

Evidence for two processing mechanisms

JAMES R SAWUSCH, HOWARD C NUSBAUM and EILEEN C SCHWAB

State University of New York, Buffalo, New York 14226

Recent experiments have indicated that contrast effects can be obtained with vowels by

anchoring a test series with one of the endpoint vowels These contextual effects cannot

be attributed to feature detector fatigue or to the induction of an overt response bias In

the present studies, anchored ABX discrimination functions and signal detection analyses

of identification data Ibefore and after anchoring} for an [i]-[I] vowel series were used to

demonstrate that [i] and [I] anchoring produce contrast effects by affecting different

per-ceptual mechanisms The effects of [i] anchoring were to increase within-[’:] category

sensitiv-ity, while [I] anchoring shifted criterion placements When vowels were placed in CVC

syllables to reduce available auditory memory, there was a significant decrease in the size

of the [I]-anchor contrast effects The magnitude of the Ill-anchor effect was unaffected by

the reduction in vowel information available in auditory memory These results suggest that

[i] and [I] anchors affect mechanisms at different levels of processing The [i] anchoring

results may reflect normalization processes in speech perception that operate at an early

level of perceptual processing, while the [I] anchoring results represent changes in response

criterion mediated by auditory memory for vowel information

Previous research in speech perception has

consis-tently revealed differences between the perception of

stop-consonants and vowels In experiments on

dichotic listening, stop-consonants consistently yield

a right-ear advantage (Shankweiler &

Studdert-Kennedy, 1967; Studdert-Kennedy & Shankweiler,

1970), while a right-ear advantage is not found for

steady-state vowels under similar circumstances

(Darwin, 1971 ; Haggard, 1971; Studdert-Kennedy &

Shankweiler, 1970) Experiments on categorical

per-ception also yield consistent differences between

stop-consonants and vowels For the stop consonants,

discrimination is typically categorical That is, the

ability to determine whether two stimuli are different

is limited by the ability to identify the same stimuli

(Liberman, Harris, Hoffman, & Griffith, 1957;

Pisoni, 1971, 1973) The discrimination of vowels,

however, is typically much better than would be

pre-dicted from identification data (Pisoni, 1973, 1975;

Fujisaki & Kawashima, Note 1, Note 2) Finally,

stop-consonants tend to show little or no influence of

contextual information on their identification (Eimas,

This work was supported by NINCDS Grant NS-12179 to

Indiana University (which supported development of the speech

synthesizer used in Experiment 3), NIMH Grant MH31468-01 to

SUNY/Buffalo, NSF Grant BNS7817068 to SUNY/Buffalo, and

SUNY Research Foundation and University Awards grants The

authors would like to thank Dr David B Pisoni for making the

facilities of the Speech Perception Laboratory at Indiana

Univer-sity available for stimulus preparation and Jerry C Forshee for

his assistance in constructing the tapes for Experiments 1 and 2.

Reprint requests should be sent to the first author at the Department

of Psychology, 4230 Ridge Lea Road, Buffalo, New York 14226.

1963; Fry, Abramson, Eimas, & Liberman, 1962; Simon & Studdert-Kennedy, 1978; Sawusch & Pisoni, Note 3), while vowels show large changes in identifi-cation as a function of context and surrounding vowels (Fry et al., 1962; Ladefoged & Broadbent, 1957; Repp, Healy, & Crowder, 1979; Sawusch & Nusbaum, 1979)

The question of contextual influences in speech perception has been addressed by a number of proce-dures Eimas (1963), Fry et al (1962), and more recently, Repp et al (1979) have requested subjects

to identify stimuli presented in a discrimination task format One general finding for all these experiments

is that identification of any individual stimulus item tends to migrate toward categories other than those

of the items it is presented with This contrastive effect is especially pronounced for ambiguous syl-lables (those near a phonetic category boundary) Furthermore, the contrastive effects in isolated, steady-state vowel identification are substantially larger than the effects found with stop-consonants (see Eimas, 1963) Using a similar procedure in which stimuli were presented in groups of four, Diehl, Elman, and McCusker (1978; Diehl, Lang, & Parker,

in press) have also reported small contrastive changes

in stop-consonant identification for near boundary stimuli

A different procedure was employed by Ladefoged and Broadbent (1957; Broadbent & Loadefoged, 1960) in which an ambiguous word was placed at the end of a sentence The first formant frequencies of the carrier sentence were systematically varied The

Copyright 1980 Psychonomic Society, Inc 421 0031-5117/80/050421-14501.65/0

422 SAWUSCH, NUSBAU~M, AND SCHWAB

effect of the various carrier sentences on the ambiguous

test items (which differed in their vowel) was one of

contrast That is, with a carrier sentence that was

synthesized with first-formant frequencies low in

their range, a word that was ambiguous between

"bit" and "bet" would be heard as "bet" (higher F1)

With a high first-formant carrier sentence, the same

word would be heard as "bit" (lower F1)

A third procedure that has been used to investigate

contextual influences in speech perception is

anchor-ing (Sawusch & Nusbaum, 1979; Simon &

Studdert-Kennedy, 1978; Sawusch & Pisoni, Note 3; Rosen,

Note 4) In this procedure, subjects are presented

with a set of stimuli to be identified under two

condi-tions The first is an equiprobable control in which

each stimulus occurs equally often In the second,

anchor condition, one of the stimuli occurs more

often than the other stimuli This procedure has been

used to investigate the perception of brightness

(Helson, 1964), dots varying in numerosity (Helson

& Kozaki, 1968), heaviness of lifted weights (Parducci,

1963, 1965), and tones varying in frequency or

inten-sity (Cuddy, Pinn, & Simons, 1973; Sawusch &

Pisoni, Note 3), as well as vowels and

stop-con-sonants In general, small contrast effects or no

effects at all have been found for stop-consonants as

a function of anchoring (Simon & Studdert-Kennedy,

1978; Sawusch & Pisoni, Note 3) However, large,

consistent effects have been found in anchoring

experiments with vowels (Sawusch & Nusbaum,

1979; Simon & Studdert-Kennedy, 1978; Sawusch &

Pisoni, Note 3)

A number of possible processing mechanisms have

been considered in connection with these contrast

effects with vowels Sawusch and Nusbaum grouped

these into three classes: feature detector fatigue,

changes in auditory ground (adaptation level), and

changes in response bias The feature detector fatigue

explanation seems to be implausible because the extra

occurrences of a stimulus in the anchor condition are

usually widely separated in time and are interspersed

with presentations of other stimuli Thus, although

virtually identical patterns of results are found for

vowels with adaptation (Morse, Kass, & Turkienicz,

1976) and anchoring procedures (see Sawusch &

Nusbaum, 1979), both results probably reflect

pro-cesses other than feature detector fatigue

The response bias explanation was explored in an

experiment by Sawusch and Nusbaum (1979), who

found identical contrast effects for subjects who were

informed of the extra occurrences of the anchoring

vowel and subjects who were not This would seem to

eliminate any overt response bias explanation of the

vowel anchoring results in which the subjects simply

tried to use the available response categories equally

often (cf Parducci, 1975)

The third possibility concerns changes to an

audi-tory ground (see Sawusch & Nusbaum, 1979; Simon

& Studdert-Kennedy, 1978) The auditory ground represents a standard against which incoming stimuli are compared The composition of this auditory ground could include information from long-term memory about auditory characteristics, pattern.,;, or features for various items as well as information from auditory memory concerning the immediately preceding stimuli During baseline identification test.-ing, no one stimulus would dominate the auditory ground, since each stimulus is equally likely How.-ever, when in an anchoring procedure, one stimulus occurs more often than any of the other stimuli, it

is more likely that auditory memory information about this stimulus will be available for comparison with subsequent stimuli Thus, the auditory ground is more likely to contain information about the anchor-ing stimulus than any other stimulus This will cause ambiguous stimuli to be mapped on to categories other than that of the anchoring stimulus

The influence of the more frequently occurring stimulus could come about in one of two ways One possibility is that the presentation of any particular stimulus now has a higher probability of being pre-ceded by the anchor stimulus If the subject retains some trace of the quality of the preceding stimulus and uses this in evaluating the current stimulus, the effect of anchoring would be to give the anchored stimulus the largest weight in this comparison The second possibility is that a cumulative adaptation level, as suggested by adaptation level theory (Helson, 1964), is the basis of vowel anchoring results Since the anchored vowel occurs more often than any other vowel, it would have a disproportionate weight in determining this adaptation level Recent results reported by Nusbaum and Sawusch (Note 5) support the adaptation level description In their experiment,

a target vowel from the middle of an [i]-[II vowel series was preceded by either an [i] or an [I] endpoint vowel and the interval between the two vowels was varied At very short ISis, both endpoint vowels caused contrast effects in the identification of ambig-uous test vowels However, this influence decreased substantially as ISI was increased to 500 msec Given that the ISI in previous anchoring studies was 4 sec, the contrast effects found in vowel anchoring can not

be adequately explained simply on the basis of audi-tory memory for the immediately preceding stimulus Rather, vowel anchoring seems to involve the

build-up of information about the anchoring vowel, pos-sibly in the form of an adaptation level If this adap-tation level is, indeed, auditory in nature, it may be

in a form similar to Massaro’s (1972) synthesized auditory memory

The auditory ground explanation is consistent with the vowel anchoring results previously reported It is also consistent with the anchoring data for stop.-consonants Since stops show less evidence of auditory memory than do vowels (Pisoni, 1971, 1973; Fujisaki

& Kawashima, Note 2), they should have a smaller

auditory memory component in determining their

auditory ground This should lead to little or no

effect of anchoring upon stop-consonant

identifica-tion, which has, in general, been the case (see Simon

& Studdert-Kennedy, 1978; Sawusch & Pisoni, Note 3,

for review) The experiments described below were

conducted as a further test of the auditory ground

and response bias explanations of anchoring effects

with vowels.

EXPERIMENT 1

The first experiment was designed to test the

effects of anchoring upon ABX discrimination of

vowels Previous experiments that have investigated

the relationship between isolated steady-state vowel

identification and ABX discrimination have found

that listener’s discrimination is not categorical (Pisoni,

1971, 1973, 1975; Fujisaki & Kawashima, Note 1,

Note 2) Rather, the discrimination of vowels from

within a phonetic category is typically well above

chance Thus, listeners can discriminate vowels that

are identified as belonging to the same category.

Pisoni (1973, 1975) and Fujisaki & Kawashima

(Note 2) have attributed this within-category

dis-crimination to the use of information in auditory

memory about the three stimuli in an ABX triad.

Thus, to the extent that changes in auditory memory

or other early perceptual processes underlie the

contrast effects found with the anchoring procedure,

we would expect systematic changes in ABX

discrim-ination within a phonetic category as a function of

anchoring That is, discrimination should change

sys-tematically at the anchored end of the vowel series.

On the other hand, if the contrast effects due to

anchoring are a result of a criterion shift in the

label-ing (identification) process, then only

between-category changes in discriminability would be expected

as a function of anchoring The change in the

phonetic category boundary caused by anchoring

would lead to a shift in the peak of the ABX

dis-crimination function (since disdis-crimination across the

category boundary seems to rely on identification

labels in STM; see Pisoni, 1973, 1975; Repp et al.,

1979; Fujisaki & Kawashima, Note 2) However,

unless changes were found in identification within a

phonetic category as a function of anchoring, no

change in within-category discriminability would be

expected That is, no changes in discriminability

would be expected for the anchored end of the series.

Thus, the nature of changes in ABX discrimination

as a function of anchoring will provide evidence of

the relative involvement of early perceptual processes

vs later criterion shifts in vowel contrast effects.

Method

Subjects The subjects in this experiment were 12

undergrad-uate and gradundergrad-uate students at the State University of New York

at Buffalo All subjects were right-handed, native speakers of English with no reported histories of any speech or hearing dis-orders The subjects were paid $3/h for their participation.

Stimuli The stimuli consisted of a set of seven isolated, steady-state vowels which ranged perceptually from [i] as in

beet to [I] as in bit These vowels were originally generated by

Pisoni (1971) using the vocal tract analogue synthesizer at the Research Laboratory of Electronics, Massachusetts Institute of Technology All of the stimuli were 300 msec in duration and con-tained five formants These stimuli varied in their formant fre-quencies for their first three formants from 270 Hz (F1), 2,300 Hz (F2), and 3,019 Hz (F3) for the [i] end of the series to 374 Hz (F1), 2,070 Hz (F2), and 2,666 Hz (F3) for the [I] end of the series

in six logarithmic steps A more complete description of these stimuli can be found in Pisoni (1971) These seven vowels were recorded on audiotape and then digitized using the PDP-11 com-puter in the Speech Perception Laboratory at Indiana University These stimuli were then reconverted to analogue form to make six test tapes Three of these tapes were identification tapes In the baseline identification tape, each of the seven stimuli occurred

10 times in random order In the [i]-anchor tape, Stimulus 1 occurred 40 time and each of the other six stimuli occurred

10 times In the [l]-anchor tape, Stimuli 1 through 6 each oc-curred 10 times and Stimulus 7 ococ-curred 40 times In each of the anchor tapes, the order of stimuli was randomized, with the restriction that no single stimulus could occur more than three times in succession All three tapes were recorded with 4 sec between stimuli.

The other three tapes were ABX discrimination tapes All dis-crimination tapes were composed of the six one-step ABX triads.

In any given triad, the first two stimuli (A and B) were adja-cent stimuli from the series The third stimulus (X) was identical

to either the first or the second stimulus For any given pair

of stimuli, this allowed four distinct triads: ABA, ABB, BAA, and BAB Each of these four compositions for each of the six adjacent pairs occurred four times (96 total triads) in the base-line ABX tape In the [i]-anchor ABX tape, the Stimulus 1,2 triads (121, 122, 211, 212) each occurred 16 times and the other triads occurred 4 times each In the [l]-anchor ABX tape, the Stimulus 6,7 triads occurred 16 times each and the other triads occurred four times In all three ABX tapes, the order of triads was random, with the restriction that no more than three triads from one particular stimulus pair could occur in succession In all tapes, there was 1 sec between items within a triad and 4 sec between triads.

Procedure The subjects were divided into two groups of six subjects each They were run in small groups of from two to four subjects each Each subject participated in two 1-h sessions

on successive days The stimulus tapes were reproduced on a Revox A-700 tape deck and presented binaurally to subjects via Telephonics TDH-39 matched and calibrated headphones The intensity of the stimuli was set to 80 dB SPL for a steady-state calibration vowel ([i]) for all tapes Each group listened to the baseline identification and ABX tapes at the beginning of each session The subjects were informed that they would be listening

to synthetic syllables that would sound like the vowels [i] and [I] They were asked to make two responses to each item on the identification tape First, they were requested to identify each vowel as either [i] or [I] Their second response was to be a rating indicating how sure they were that they had identified the stimu-lus correctly A 4-point scale was used, with a 1 indicating that the subject was positive her (his) identification was correct, a 2 indicating a probable correct, a 3 indicating a possible correct, and a 4 indicating a guess For the ABX tapes, the subjects were informed that they would be hearing groups of three stimuli.

In these groups, the first two stimuli would always be different while the third would be identical to either the first or the second They were to indicate whether the third item sounded most like the first stimulus or most like the second.

Following the control tapes, each of the two groups listened to

a different set of anchor tapes The [i] group heard [i]

identifi-424 SAWUSCH, NUSBAUM, AND SCHWAB

0 7¸

STIMULUS VALUE

Figure 1 Rating functions for the control (solid circles) and

anchor (open circles) conditions with data for the [i]-anchor group

on the left and the [i]-anchor group on the right.

cation and [i] ABX anchor tapes, while the other group heard

the corresponding [I]-anchor tapes The subjects were not given

any new instructions regarding these tapes They used the same

response procedures for the two types of anchor tapes that they

had used for the baseline tapes By the end of the experiment,

each subject had provided at least 20 identification responses

to each stimulus under both baseline and anchoring conditions.

They had also provided at "least 32 discrimination responses to

each pair of stimuli under each condition.

Results

The identification and rating responses for the

identification tapes were converted to an 8-point

scale A rating of 1 indicated an extremely confident

[i] response, ratings of 4 and 5 indicated [i] and [I]

guesses, while a rating of 8 indicated a positive [I]

response The results for the two groups are shown in

Figure 1 Both groups showed significant shifts in

their category boundaries toward the category of the

anchoring stimulus [t(5) = 3.80, p < 02, for the

[i]-anchor group and t(5) = 4.90, p < 01, for the

[I]-anchor group].’ These contrast effects are essentially

identical to those previously reported by Sawusch

and Nusbaum (1979) and Sawusch and Pisoni

(Note 3)

The ABX discrimination results are shown in

Figure 2 The category boundaries for the

cor-responding identification functions are shown by the

arrows The peak in the [i]-anchor group

discrimina-tion funcdiscrimina-tion is shifted toward the [i] end of the series,

relative to the baseline condition (Figure 2, left side)

A corresponding shift in the [I]-anchor ABX peak

was found for the [I] group (right side, Figure 2)

Of the 12 subjects, 11 show this pattern of a shift in

the discrimination function peak toward the anchored

category (which is significant, p = 012, using a

two-tailed sign test) The one subject who did not show

the expected shift was in the [I]-anchor group This

subject showed no evidence of a discrimination peak

shift

In addition to a shift in peak discriminability, the [i]-anchor group showed a marked increase in the dis-criminability of the Stimulus 1,2 pair (the anchoring pair) Each of the six [i]-anchor-group subjects showed this increase in discriminability, which was significant [t(5) = 5.72, p < 01, for the 16.8°70 mean increase in discriminability] No comparable increase,

in discriminability for the Stimulus 6,7 pair was found for the [I]-anchor group (Three subjects showed increases in discrimination and three showed decreases following the [I] anchor.)

Discussion

For the [I]-anchor group, no changes in within- category discriminability were found as a function of anchoring The discriminability of the Stimulus 6,7 pair showed no change due to anchoring Thus, for the [I] anchor, the contrast effects found could be due to criterion shifts Criterion shifts in the identifi-cation of stimuli would lead to a shift in the ABX discrimination peak if, as is usually assumed, implicit identification (categorization) underlies the between- category discrimination of subjects (see Pisoni, 1973; Repp et al., 1979; Fujisaki & Kawashima, Note 2) The criterion shift explanation of anchoring predicts

no change in within-category ABX discrimination because no change in identification performance within either category was found

The criterion shift explanation does not, however,, appear to be an adequate explanation of the [i] anchor results In the [i]-anchor group, a large increase

in discriminability was found within the [i] category, for the anchored, Stimulus 1,2 pair This increase can not be accounted for by the small and inconsistent change in identification (rating) for Stimulus 2 fol.-lowing [i] anchoring (see Figure l, left side)? The [ill anchoring results seem to be due, in part, to a change

100

50

z w w

ABX DISCRIMINATION

~ Control o -0 Anchor

2 3 4 5 ~ 7 1 2 3 4 5 6

STIMULUS VALUE

Figure 2 Percent correct discrimination for the [i]-anchor group (left) and the [D-anchor group (right) Rating category boundaries (see text) are marked by arrows.

in perceptual processing prior to identification of the

stimulus Thus, these results indicate that two

distinct types of processing changes may be involved

in the contrast effects found with vowels The next

experiment was conducted as a further test of

whether [i] and [I] anchoring effects reflect the

involvement of distinct perceptual processing

mech-anisms.

EXPERIMENT 2

If two distinct processes are involved in anchoring

effects with vowels, and one of these represents an

early perceptual change while the other represents a

higher level change, then we might expect them to

show up as sensitivity changes and criterion shifts,

respectively, in a signal detection analysis However,

our previous experiments have collected far too few

judgments per stimulus to allow the use of this type

of data analysis The present experiment was

de-signed to collect a sufficient number of subject

responses to each stimulus for a signal detection

analysis of individual subject data The variation of

signal detection theory proposed by Durlach and

Braida (1969; Braida & Durlach, 1972) will be used

to evaluate the data If the within-category

discrim-inability increase for [i] vowels that was found in

Experiment 1 reflects an early perceptual change, we

would expect an increase in sensitivity for the

Stimu-lus 1,2 pair as a result of [i] anchoring However, to

the extent that anchoring induces changes at a later

stage in perceptual processing, we would expect to

find criterion shifts as a result of both [i] and [I]

anchoring.

Method

Subjects The subjects in this experiment were 14 undergraduates

at the State University of New York at Buffalo who

partici-pated for course credit These subjects met the same

require-ments as those in Experiment 1.

Stimuli The same seven vowel series used in Experiment 1

was also used here These stimuli were recorded to make two

additional baseline, two [i]-anchor and two[l]-anchor tapes

(yielding three tapes of each type) As before, all stimuli were

recorded in random order, with no more than three occurrences

of any stimulus in succession.

Procedure The experimental tapes were reproduced and

presented to subjects in a manner similar to that of Experiment 1.

The subjects were divided into two groups of seven All seven

subjects in a group were run simultaneously Each subject

par-ticipated for a total of 5 h spread over 4 days In each session,

the subjects listened to three presentations of baseline tapes (all

using different stimulus orders on any given day) and then, after

a short break, listened to three presentations of an anchoring tape

(again, all different on any given day) One group listened to

[i]-anchor tapes and the other listened to [I]-anchor tapes In

addition, on Day 1, the subjects listened to the seven vowel

stimuli in order and one extra presentation of a baseline tape for

practice purposes.

The subjects were informed that the tapes contained random

orders of seven different vowels varying from [i] to [11 The

subjects then listened to the seven stimuli, in order, from the [i] endpoint to the [I] endpoint They were asked to use a 7-point response scale and to attempt to uniquely identify each of the seven vowels A response of 1 was to denote the [i] endpoint vowel, while a 7 was to denote the [I] vowel The values in between were to identify the intermediate vowels Subjects then listened to the vowels in order a second time Following this, they used the seven responses to identify the stimuli from a base-line tape for practice The results of the practice tape were not included in the data analysis By the end of the experiment, each subject had provided at least 120 responses (4 sessions × 3 tapes

× 10 occurrences) to each of the seven stimuli in both baseline and one of the anchoring conditions, exclusive of the practice tape.

Results The data for 2 of the 14 subjects were dropped from the experiment because these subjects did not use all seven responses One subject in the [i] group did not use Responses 3 and 4 and one subject in the [I] group did not use Response 4 The average rating functions for the remaining six subjects in each group are shown in Figure 3 In both groups, a significant shift in the category boundary toward the category of the anchor was found [t(5) = 4.71,

p < 01, and t(5) = 5.21, p < 01, for the [i]-and [I]-anchor groups, respectively) Each of the 12 subjects showed the expected shift in their category boundary The confusion matrices (seven stimuli by seven responses) for both baseline and anchored con-ditions for each subject were submitted to a signal detection analysis The version of signal detection analysis proposed by Durlach and Braida (1969; Braida & Durlach, 1972) was used The individual confusion matrices were converted to cumulative probability matrices with entries accumulated over responses for each stimulus These cumulative proba-bilities were converted to z scores, with the restric-tion that only probabilities between 008 and 992 were converted Cells with probabilities outside this range were considered indefinite and were not used

[i] Anchor [I] Anchor

7

Z 6

4

STIMULUS VALUE

Figure 3 Rating functions for the [i]-anchor (left) and [l]-anchor groups of Experiment 2 in both control (solid circles) and anchored (open circles) conditions.

426 SAWUSCH, NUSBAUM, AND SCHWAB

2,0’

0.0

I

I II

1,2 2,3 3,4 4,5 5,6 6,7 1,2 2,3 3.4 4,5 5,6 6,7

STIMULUS PAIR

Figure 4 Paired d’ values for the [i]-anchor (left) and

Ill-anchor (right) groups in both control (solid circles) and Ill-anchored

(open circles) conditions.

in determining either sensitivities (d’) or criterions.3

The d’ for each adjacent pair of stimuli was computed

by taking the mean difference between z scores for

the two stimuli over the seven response alternatives.

The paired d’ values for baseline and [i]-anchored

conditions are shown in Figure 4 on the left A

number of aspects of the data should be noted.

First, every subject showed an increase in d’ for the

Stimulus 1,2 pair and every subject showed a

decrease in d’ for the Stimulus 4,5 pair Second,

for three of the subjects, the category boundary fell

between Stimuli 4 and 5, while for the other three

subjects, the boundary fell between Stimuli 3 and 4.

Each of the six subjects showed a decrease in d’

for the stimulus pair which spanned (one stimulus

on either side of) the baseline category boundary.

The paired d’ values for the baseline and [I]-anchor

conditions are shown in the right-hand side of Figure 4.

In contrast to the [i]-anchor results, there was no

sig-nificant change in d’ for the stimulus pair spanning

the baseline category boundary [t(5) = 1.64, p > 1].

Separate two-way ANOVAs were used to evaluate

the d’ results for [i] and [I] anchoring For the

[i] anchoring condition, the main effect of stimuli

was significant [F(5,25) = 7.50, p < 001], but the

main effect of anchoring was not [F(1,5) = 32,

p > 25] The interaction between anchoring and

stimuli was significant [F(5,25) = 2.71, p < 05].

Post hoc Newman-Keuls tests revealed that for the

Stimulus 1,2 pair, anchoring caused a significant

increase in d’, while for the Stimulus 4,5 pair,

anchoring led to a significant decrease (both p < 05).

For the comparable [I]-anchor analysis, the main

effect of stimuli was significant [F(5,25) = 6.01,

p < 001], but the effect of anchoring was not [F(I,5)

= 1.04, p > 25] The interaction was marginally

significant [F(5,25) = 2.42, 05 < p < 1] Post hoc

tests showed that only the increase in d’ for the Stimulus 4,5 pair was significant (p < 05) Thus, the d’ results from the present experiment are gener-ally consistent with the discriminability changes in Experiment 1 Consistent d’ changes were found both within the [i] category, for the Stimulus 1,2 pair and across the category boundary for [i] anchor-ing, but little consistent change in d’ was found for the [I]-anchor condition.

Criterion cut points were determined for each response pair, based on the previously computed values for d’ Cumulative criterion placements were calculated by determining the z score that cor-responded to each criterion based on a cumulative d’ scale The six cumulative criterion placements for the seven response categories for both baseline and [i] anchoring are shown on the left side of Figure 5 Although there does appear to be an overall shift

in the criteria toward the [i] end of the series, this was not consistent across subjects Rather, four of the subjects showed a shift in all of their criteria toward the [i] end of the series, while two showed

a shift in all criteria toward the [I] end of the series Thus, the direction of criterion shift for two of the subjects is opposite the shift in the category bound-ary However, these two subjects also showed the smallest category boundary shifts due to [i] anchoring Thus, the criterion shifts may account for part of the category boundary shifts for the [i] anchor How.-ever, all of the [i]-anchor shift for two of the sub-jects was due to sensitivity changes, while at least part of the [i]-anchor effect for the other four sub-jects was due to sensitivity changes The cumulative criteria for the baseline and [I]-anchored functions are shown on the right side of Figure 5 Every one

1,2 23 3,4 45 5,6 67 1,2 2,3 3,4 4,5 56 6

RESPONSE PAIR

Figure 5 Cumulative criterion cutpoints (in z units) for the

[i]-anchor (left) and [I]-[i]-anchor (fight) groups in both control (solid

circles) and anchored (open circles) conditions.

[i] Anchor [I] Anchor

1.0¸

.75

.5O

1,2 2,3 3,4 4,5 5,6 6,7 1,2 2,3 :3,4 4,5 5,6 6,7

STIMULUS PAIR

Figure 6 Paired values of the sensitivity index P(A) for the

[i]-anchor (left) and [l]-anchor (right) groups in both control

(solid circles) and anchored (open circles) conditions.

of the six subjects exhibited a shift in all six of their

criteria toward the [I] end of series as a result of

[I1 anchoring.

As a check on the validity of the d’ results, a

nonparametric measure of sensitivity was also

com-puted for each pair of stimuli from the cumulative

probability matrix for each subject The area under

the ROC curve [P(A); see Green & Swets, 1974]

was computed as our alternative measure of paired

sensitivity since it does not depend upon the equal

variance, normal distribution assumptions of the

Durlach and Braida (1969) model.3 The mean values

of P(A) (across subjects) for the baseline and

[i]-anchor conditions are shown in Figure 6 (left side).

As with the d’ results, a significant decrease in

sensi-tivity was found for the stimulus pair spanning the

baseline category boundary for each subject It(5) =

4.92, p < 01] A similar P(A) analysis was done for

the [I]-anchor subjects, and the mean results are

shown in Figure 6 on the right As with the

para-metric analysis, no significant change in sensitivity

was found for the stimulus pair spanning the

base-line category boundary [t(5) = 1.23, p > 2] As

with the d’ data, separate two-way ANOVAs were

run on the P(A) data for the [i]- and [I]-anchor

conditions For the [i]-anchor group, the main effect

of stimuli was significant, while the effect of

anchor-ing was not [F(5,25) = 5.62, p < 01, and F(1,5)

= 33, p > 25, respectively] The Stimulus by

Anchoring interaction was significant [F(5,25) =

6.62, p < 001] and post hoc tests revealed that

both the Stimulus 1,2 and Stimulus 2,3 pairs showed

significant increases in sensitivity as a result of

anchoring, while both the Stimulus 3,4 and

Stimu-lus 4,5 pairs showed significant decreases (all p <

.05) For the [I]-anchor group, the main effect of

stimuli was significant and the effect of anchoring

was not [F(5,25) = 6.183, p < 001, and F(1,5)

= 0.21, p > 25, respectively] The Stimulus by

Anchor interaction was marginally significant [F(5,25) = 2.23, 05 < p < 1], and none of the stimulus pairs showed a significant change using post hoc tests No consistent differences were found

in P(A) as a function of [I] anchoring Thus, the results for P(A) and d’ for the [I]-anchor subjects are virtually identical and show little or no influence

of the [I] anchor on sensitivity measures.

Discussion

The increases in sensitivity found with both P(A) (area under the ROC curve) and d’ for stimuli within the [i] category following [i] anchoring mirror the increase in ABX discriminability for the [i] category found in Experiment I The Stimulus 1,2 pair showed increases in discriminability in both experiments In addition, decreases in sensitivity were found for both measures at the category boundary after [i] anchoring When coupled with the inconsistent changes in criterion placement across subjects for the [i| anchor, these results indicate that the contrast effects found with [i] anchoring on an [i]-[I] series are predom-inantly due to changes in sensitivity By comparison, consistent changes in sensitivity were not found as

a consequence of [I] anchoring Rather, systematic shifts in criteria toward the [I] category would seem

to be responsible for the contrast effects found for

an [I] anchor

Although both Experiments 1 and 2 demonstrate that different processing changes underlie [i] and [I] vowel anchoring, these experiments do not isolate the processing changes themselves However, since one of the major differences between these vowel stimuli (which show anchoring effects) and stop con-sonants (which do not consistently show anchoring effects) is the degree of available auditory memory (Pisoni, 1973; Fujisaki & Kawashima, Note 2), some form of auditory memory, as outlined earlier, may underlie part of the anchoring effects found with vowels Experiment 3 was designed to investigate the role of auditory memory in anchoring

EXPERIMENT 3

As noted earlier, Nusbaum and Sawusch (Note 5) have provided evidence that if some form of audi-tory memory is involved in vowel-anchoring effects,

it is not the auditory memory trace of only the immediately preceding vowel This would seem to indicate that the auditory memory explored by Crowder (1971, 1973; Crowder & Morton, 1969), termed precategorical acoustic storage, or PAS, is probably not responsible for our vowel-anchoring effects The duration of PAS is typically around

2 sec (Darwin, Turvey, & Crowder, 1972), while our ISis were 4 sec Thus, if auditory memory is involved

in our vowel-anchoring results, it is probably not the

428 SAWUSCH, NUSBAUqVl, AND SCHWAB

short-lived PAS However, there is evidence that

some form of precategorical information does persist

at durations of 2 sec or more Repp et al (1979)

found that even with a 2-sec interval filled with

repetitions of an extraneous semivowel, AX vowel

discrimination was well above chance Thus, some

vowel information does seem to persist in a form

where it could influence the identification of a

fol-lowing vowel ’This vowel information may be

involved in our anchoring results The

discrimin-ability of vowels, using either an AX procedure

(Repp et al., 1979) or an ABX procedure (Pisoni,

1973, 1975) can thus serve as an index of the strength

of this memory trace in determining the role of

auditory memory in vowel anchoring results.

The vowel stimuli used in Experiments 1 and 2 are

perceived more nearly continuously than

categori-cally (see Pisoni, 1973) In the present experiment,

we embedded vowels in CVC contexts in an effort

to reduce the influence of auditory memory upon

identification (see Stevens, 1968; Sachs, Note 6) To

the extent that this is successful, ABX discrimination

results for the CVCs should be less continuous than

the isolated vowel results of Pisoni (1973) and more

categorical The results of Experiments 1 and 2

showed that two distinct processes are involved in

vowel anchoring effects Reducing the information

in auditory memory could have its primary influence

on the magnitude of either the [i]-anchor contrast

effects or the [I]-anchor effects.

If auditory memory underlies the influence of the

[i] anchor and the [I] anchor induces changes at a

later, response, stage, then reducing available auditory

memory should also reduce the size of the contrast

effects found for [i] In addition, we might also

expect the size o.~ the [I]-induced contrast effects to

decrease, since reducing available auditory memory

renders perception more categorical (Pisoni, 1973,

1975) and previous results have shown that

categori-cally perceived speech stimuli usually show smaller

contrast effects than noncategorically perceived

stimuli (Eimas, 1963; Sawusch & Pisoni, Note 3).

Thus, if the CVC series show less influence of

audi-tory memory in ABX discrimination and both the

CiC and CIC anchor effects are drastically reduced

in magnitude, we will have evidence that some form

of auditory memory underlies the [i| (and [I])

vowel-anchoring effects.

On the other hand, if auditory memory is not a

major factor in [i] anchoring, then a different pattern

of results would be expected According to this

explanation, reducing the available auditory memory

should reduce the contrast effects only for a CIC

anchor (as outlined above) The effects of CiC

anchoring would be relatively unaffected by auditory

memory and should remain despite decreases in

available auditory memory.

A third possibility is that reducing the informa-tion available in auditory memory will have no effect on the contrast effects caused by either (SiC

or CIC anchors This results is predicted by a model recently proposed by Fujisaki and Shigeno (1979),

in which auditory memory underlies assimilation effects (not contrast) in identification According to this model, contrast effects in identification are mediated by categorical (phonetic) short-term memory, while auditory memory underlies assimilation effects Thus, reducing the available auditory mem-ory information should either increase the size of the contrast effects found as a function of vowel anchoring (due to a larger reliance on categor~cal STM) or leave these contrast effects unchanged The following experiment, in which vowels were placed

in CVC syllables to reduce the available auditory memory for vowels, should allow us to discern the role of auditory memory in vowel anchoring.

Method

Subjects The subjects were 20 undergraduates at the State University of New York at Buffalo, who participated to fulfill

a course requirement They met the same requirements as subjects

in the previous experiments.

Stimuli The stimuli consisted of two sets of seven CVC syl-lables One set ranged perceptually from [sis] (as in cease) to [sis] (as in sister), while the other ranged from [bit] (beetj to [blt[ (bit) In the [sVs] series, 250-msec steady-state vowels were

embedded between initial and final Is] fricatives The Is] frica-tives were both 200 msec in duration and consisted of band-limited noise between 3,400 and 5,000 Hz The seven syllables varied only in the first three formants for the vowel The actual formant frequencies, bandwidths, and fundamental frequency were identical to those of the isolated vowels used in Experi-ments 1 and 2 (see also Pisoni, 1971) Thus, the [sis]-[slsl stimuli represent the embedding of vowels similar to our original isolated vowels in an [sVs] context.

The [bit]-[blt] series varied both the initial consonantal transi-tions and the formant frequencies for the vowel in equal logarithmic steps All seven stimuli were 200 msec in duration and consisted of an initial 30-msec consonantal transitional followed by a 90-msec dynamic vowel During the vowel, the first three formants gradually changed from their values at the end

of the consonant to "target" values The target values were attained 60 msec into the vowel and held for the last 30 msec

of the vowel The formant frequencies for the first three formants

of the [bit] and [blt] endpoints at onset (0 msec); end of conso-nantal transitions (30 msec); and end of the vowel transit:ons (90 msec) are shown in Table 1 The values between these p~3ints were determined by linear interpolation The first 35 msec of each vowel contained vocalic excitation, while the last 55 nasec

Table 1 Formant Frequency Values for the [bit] (Stimulus 1) and [bit] (Stimulus 7) Endpoints of the [bit] -[bit] Series at Onset

(0 msec), 30 msec, and 90 msec

Time

[bit] [blt]

3

TIME (msec)

Figure 7 Sound spectrograms of the [bit] and [bit] endpoints of the [bit]-[blt] series used in Experiment 3.

were aspirated in preparation for the voiceless final stop In all

seven stimuli, the vowels were followed by a 60-msec silent period

and then a 20-msec burst appropriate for the voiceless stop It].

The endpoints of the [bVt] series were patterned after sound

spectrograms of utterance of one of the authors Spectrograms of

these two endpoints are shown in Figure 7 All stimuli were

gener-ated using a software cascade synthesizer (Klatt, Note 7, or see

Kewley-Port, Note 8) in the Speech Perception Laboratory at the

State University of New York at Buffalo.

The stimuli were converted to analogue form and recorded to

make four test tapes for each of the two stimulus series The

baseline identification tape contained 10 occurrences of each

stimulus from a series in random order The [sis] and [bit]

end-point anchor tapes each contained 40 occurrences of the [sis]

or [bit] endpoint stimulus and 10 occurrences of each of the other

six stimuli in random order The [sls] and [blt] anchor tapes were

constructed in a similar fashion The final tape for each series

was an ABX discrimination tape This tape consisted of two-step

comparisons (i.e., Stimuli 1 and 3, 2 and 4, etc.) with 500 msec

between stimuli within a triad This tape was constructed in a

fashion similar to the baseline ABX tape for Experiment 1.

In all four tapes, there were 4 sec between trials (triads) All

stimuli were presented in random order, with the restriction that

no stimulus (or stimulus pair for ABX) could occur more than

three times in succession.

Procedure The tapes were reproduced and played to the

sub-jects on the same equipment and in a fashion similar to that used

in Experiments 1 and 2 Eight of the subjects listened to the

[bit]-[blt] series and 12 listened to the [sis]-[sls] series All

sub-jects were run in small groups of from two to six at a time.

Each subject participated in two l-h sessions on separate days.

On each day, the subjects listened to one presentation of the

baseline identification tape followed by the ABX discrimination

tape Following these, two presentations of the [bit] ([sis]) anchoring tape were presented on one day and two presentations

of the [bit] ([sis]) anchoring tape on the other The order of these tapes was counterbalanced across subjects The subjects used the same identification plus rating response for the identification tapes that was used in Experiment 1 For the ABX tapes, the subjects also used the same response procedure that was used in Experiment 1 By the end of the experiment, each subject had provided at least 20 responses to each stimulus in each of the three identification conditions and 32 responses to each stimulus pair for the ABX discrimination tapes.

Results The baseline identification and ABX discrimina-tion results for the [sis]-[sls] group are shown in Figure 8 on the right For comparison purposes, the ABX results for these same vowels in isolation are shown (from Pisotfi, 11973), as are the predictions of the Haskins categorical perception model for our [sis]-[sls] series.4 The ABX results for this series are clearly not categorical [X2(4) = 28.7, p < 001].s However, the obtained [sis]-[sls] discrimination is not as continuous as the long, isolated vowel data

of Pisoni [X2(4) = 7.79, 05 < p < 1] Thus, the [sis]-[sls] series seems to allow some use of auditory memory but less than that for the vowels used in Experiments 1 and 2.

The identification plus rating data was also con-verted into the same 8-point rating scale used in

430 SAWUSCH, NUSBAUM, AND SCHWAB

[sis]-[sls] Series

r~ 5

la.I

4-2 3 4 5 6 7 2 3 4

STIMULUS VALUE

Ioo

5o

Figure 8 Rating functions (left) and ABX discrimination

(right) data for the [sis]-[sls] series in Experiment 3 Control

(solid triangles), [sis]-anchor (open circles) and [sls]-anchor (open

squares) rating functions are shown on the left Control rating

data (solid triangles), obtained ABX discrimination data (solid

circles), Haskins predicted ABX discrimination (open circles), and

obtained ABX discrimination for isolated vowels (open squares)

are shown on the right.

Experiment 1 The baseline rating, [sis] anchor, and

Isis] anchor functions appear on the left in Figure 8

each of the anchoring conditions caused a significant

shift in the category boundary toward the anchored

end of the series [t(ll) = 7.72, p < 001, for the

[sis] anchor and t(ll) = 2.93, p < 02, for the

[sis] anchor] Furthermore, the [sis]-anchor-induced

shift was significantly larger than the Isis] anchor

shift [t(ll) = 3.73, p < 01] Thus, since the

[sis]-Isis] series shows less use of auditory memory than

the isolated vowels and contrast effects at the [I]

end of this series seems to have been reduced, it

appears that auditory memory is the mediating factor

in the anchoring effects of [I] but not of [i]

A similar pattern of results was found for the

[bit]-[bit] series The ABX results are shown on the right

side of Figure 9 Again, the predictions of the

Haskins model and the data of Pisoni (1971) are

shown for comparison The [bit]-[bIt] data were

sig-nificantly less categorical than the Haskins

predic-tions [X2(4) = 26.7, p < 001] Also, as before, the

obtained [bit]-[blt] discrimination data were not as

continuous as the isolated vowels [X2(4) = 6.01, 1

< p < 2] The baseline, [bit]-anchor, and [bIt]-anchor

rating functions are shown on the left side of Figure 9

Both anchors produced significant shifts in the

cate-gory boundary toward the catecate-gory of the anchor

It(7) = 7.04, p < 001, and t(7) = 3.12, p < 02,

for the [bit] and [bit] anchors, respectively] As with

the [sis]-|sIs] series, the [bit] anchor produced a

sig-nificantly larger shift than the [bit] anchor [t(7)

= 3.45, p < 02]

Discussion

The reduction in available auditory memory for

both of these two test series when compared to the

long, isolated vowels clearly indicates that this series

is intermediate between categorical and continuous perception Furthermore, both the [i] and the [I] ends of the series appear to have equal amounts of auditory memory information available, since they have equal within-category discriminability (approxi-mately 65% correct) However, only the anchoring at the [I] end of the series appears to have been affected by this reduction in auditory memory

In Experiments 1 and 2, the [i] and [I] anchor-ing effects were approximately equal, with a slightly larger [i] effect in Experiment 1 and a slightly larger [I] effect in Experiment 2 Although di-rect comparison of the magnitudes of the shifts across experiments is inappropriate because the stimuli are different and hence the category bound-ary shifts (measured in stimulus units) are not strictly comparable,6 the results of Experiment 3 are clearly different from those of Experiments 1 and 2 and from those of Sawusch and Nusbaum (1979) The results with both the [bit]-[bIt] series and the [sis]-[sIs] series indicate that reducing the available auditory memory for vowel information reduces the contrast effects for the [I]-anchor stimuli Thus, the contrast effects found with [I]-vowel anchoring appear to be mediated by auditory memory for vowel information As auditory memory was reduced, the contrast effects for [I] anchors were reduced cor-respondingly These results are inconsistent with the model proposed by Fujisaki and Shigeno (1979) as outlined earlier The effect of the [i] anchors however, was not related to the use of auditory memory Instead, the [i] effect appears to represent a change

in early perceptual processing, separate from audi-tory memory, possibly the rerunning of a prototype from long-term memory

[bit]-[blt] Series

1 2 3 4 5 6 7 1 2 3 4

STIMULUS VALUE

IOO -~

m m

0

5 6 7

Figure 9 Rating functions (left) and ABX discrimination

(right) data for the [bit]-[bit] series in Experiment 3 Control

(solid triangles), [bitl-anchor (open circles) and [blt]-anchor (olden squares) rating functions are shown on the left Control rating

data (solid triangles), obtained ABX discrimination data (solid

circles), Haskins predicted ABX discrmination (olden circles), and obtained ABX discrimination for isolated vowels (open squares) are shown on the right.