INTRODUCTION
Research in the fields of first- and second-language acquisition have
repeatedly revealed the important relationship between speech
perception and early linguistic experience (Bohn,
2000; Strange, 1995). Studies of
infant speech perception have shown that children under six months of
age are able to discriminate phonemic contrasts that are not present in
their ambient language (Jusczyk, 1997). For
example, 6-month-old infants who were raised in English-speaking
environments could discriminate Czech alveolar trills and
palato-alveolar fricatives (Eilers, Gavin, &
Oller, 1982), Swedish front rounded and unrounded vowels (Kuhl, Williams, Lacerda, Stevens, & Lindblom,
1992), and Salish glottalized velar and uvular stops (Werker & Tees, 1984). Native English-speaking
adults, however, have more difficulty making these discriminations
(Polka, 1992, 1995)
and infants appear to lose their ability to discriminate many nonnative
contrasts in the first year of life (Aslin &
Pisoni, 1980; Kuhl, 1993; Werker & Tees, 1984).
As a child is exposed to her native language, her perceptual
abilities become tuned to the relevant phonological contrasts in that
language. According to the attunement theory of language development,
children are born with the basic sensory capacities to discriminate all
possible phonemic contrasts in any language, but early linguistic
experience shapes their perceptual abilities to enhance the relevant
contrasts in their native language and attenuate the irrelevant
contrasts (Aslin & Pisoni, 1980). The
result of this process of attunement is that by one year of age, the
child performs like adults on nonnative phoneme discrimination tasks
and finds it extremely difficult to make phonemic contrasts that are
not present in her first language (Flege,
1995; Polka, 1992, 1995). Kuhl (1993) has
described this process of attunement as the result of a perceptual
magnet effect in which children carve out phonetic categories based on
their native language phonemic inventory and simultaneously become less
sensitive to contrasts outside that inventory.
In contrast to these native-language constraints on the development
of perceptual categories, early exposure to multiple languages has been
shown to have long-lasting effects on adult speech perception. Polka
(1992) conducted a study to assess the
generalization of featural contrasts to nonnative segment contrasts.
She found that early bilingual Farsi speakers were able to generalize
the uvular-velar place distinction found in Farsi voiced stops to the
same place distinction in Salish voiceless glottalized stops. Neither
monolingual English speakers nor late bilingual Farsi speakers could
discriminate the same pair of segments. These results suggest that mere
exposure to multiple languages early in life may produce effects on
speech perception abilities later in adulthood.
Early linguistic experience with a specific nonnative language may
also produce lasting effects on perceptual abilities for phonemic
contrasts in that language. Tees and Werker (1984) investigated the discrimination of the Hindi
voiceless retroflexed stop and voiceless dental stop contrast in a
number of different populations of listeners. They found that native
English-speaking participants who had been exposed to Hindi in the
first two years of life displayed high levels of discrimination
accuracy, like native English infants and native English-speaking
adults with five years of experience speaking Hindi. Native
English-speaking adults with only one year of experience speaking Hindi
and native English-speaking adults who received laboratory training on
the contrast demonstrated performance at floor levels. Tees and Werker
concluded from these results that early exposure to a specific
linguistic contrast leads to the maintenance of the ability to perceive
the contrast, even in the absence of further exposure or reinforcement.
This conclusion is also consistent with Aslin and Pisoni's (1980) theory of perceptual attunement, which
permits the maintenance of a perceptual ability even in the absence of
any feedback.
Most of the research on the effects of linguistic experience on
speech perception has investigated the perception of individual
segmental contrasts. However, Allen (1983)
showed that the effects of linguistic experience are present in
prosodic development, as well. He obtained both cross-sectional and
longitudinal evidence for the loss of the perceptual ability to detect
lexical stress distinctions in certain trisyllabic forms by
French-speaking 4- and 5-year-olds. German- and Swedish-speaking
children of the same age showed no such loss in discrimination ability.
The critical difference between the French children and the German and
Swedish children was that the stress contrast under examination occurs
in German and Swedish, but not in French. Once again, these results
reveal that children lose the ability to make distinctions that are not
relevant in their native language as a normal part of development.
In addition to these studies, other researchers have shown that
cultural experience shapes the perception of personality traits in
vocal features (Peng, Zebrowitz, & Lee,
1993). Peng et al. found that American listeners rated American
talkers who spoke faster as more powerful and more competent than
talkers who spoke slower. Korean listeners, on the other hand, rated
Korean talkers with tense voices as more powerful than talkers with
more relaxed voices. Korean listeners in the United States, however,
rated louder American talkers as more competent, and tense Korean
voices as more powerful.
These results reveal that exposure to linguistic and cultural
associations with voice types has an effect on the perception of
indexical properties of speech, such as personality traits. Spoken
language perception and linguistic experience seem to be intimately
linked at various levels of language processing, from linguistic
properties like features (Polka, 1992),
segments (Tees & Werker, 1984), and
suprasegmentals (Allen, 1983), to indexical
properties of the talker (Peng et al., 1993).
The goal of the present study was to examine the relationship between
early linguistic experience and the perception of another indexical
property of speech, dialect variation. In particular, this study
investigated the role of residential history on the perception of
regional dialect variation in American English.
Dialect categorization has been studied by sociolinguists, and more
recently by speech perception researchers, using a number of different
methodologies and stimulus materials. In one recent study, Purnell,
Idsardi, and Baugh (1999) used the
matched-guise technique to investigate the racial and ethnic dialect
identification abilities of landlords in the San Francisco area. They
found that the landlords were able to identify the dialect of the
talker, based only on a short answering machine message over the
telephone. Preston (1993) used a matching
task to explore regional dialect identification by Michigan and Indiana
adults. The stimulus materials in his study were short samples of
spontaneous speech from nine middle-aged male talkers representing nine
different cities on a north-south continuum between Michigan and
Alabama. Preston found that his listeners were able to make only broad
distinctions between northern and southern talkers.
Using an eight-alternative forced-choice categorization task,
Williams, Garrett, and Coupland (1999) asked
Welsh adolescents to categorize other Welsh adolescents by regional
dialect after they listened to samples of short narratives. Williams et
al. found that the adolescents were about 30% accurate in making their
responses. Clopper and Pisoni (in press) conducted
a similar six-alternative forced-choice categorization task in which they
asked undergraduates in Indiana to categorize talkers by regional dialect of
American English using read sentence materials. Like Williams et al.,
we also found that the listeners were about 30% accurate in making
their categorization judgments. The low dialect categorization accuracy
scores reported by Williams et al. and Clopper and Pisoni suggest that
the forced-choice categorization task is difficult for naïve
listeners, despite performance levels that are above chance.
Effects of residential history have been documented in a number of
studies that have explored the perception of dialect variation. For
example, Williams et al. (1999) found that
Welsh school teachers performed much better than the Welsh schoolboys
on the same dialect categorization task. Although they did not carry
out a statistical comparison of the two groups, Williams et al.
presented data that suggested that the teachers performed the task with
approximately 52% accuracy, compared to the adolescents' 30%
accuracy. Williams et al. attribute this difference in performance to
the teachers' travel and teaching experiences, which the
schoolboys lacked. Williams et al. also found that the schoolboys were
more accurate (45%) in categorizing talkers from their own region than
talkers from the other regions (24%). This finding suggests that even
the limited linguistic experience of adolescents is enough to
differentially affect their performance on a dialect categorization
task.
In a recent study using sentences from unfamiliar talkers, Clopper
and Pisoni (in press) did not find any differences in accuracy based on
the residential history of their listeners, but they did find
differences in the dialect perceptual similarity spaces of the
listeners based on whether they were from northern Indiana, southern
Indiana, or had lived out of state. In particular, a clustering
analysis of the stimulus-response confusion matrices revealed that the
three groups of listeners demonstrated consistent differences both in
the perceived structure of the perceptual similarity spaces and in the
perceptual distances between the different dialects. This pattern
suggests that early linguistic experience affects how dialect variation
is encoded and represented in memory and used in explicit
categorization tasks.
Similarly, in his study of dialect identification, Preston (1993) found differences between Indiana and
Michigan listeners in the location of where they perceived the major
north–south dialect boundary to be. The Michigan listeners heard
the north–south boundary between Indiana and Kentucky, whereas
the Indiana listeners heard the boundary to be between Kentucky and
Tennessee. Preston (1986, 1989) also reported differences based on
residential history in how undergraduates draw and label dialect maps
of the United States and in their judgments about where
“correct” and “pleasant” English is spoken.
Taken together, all of these results suggest that early linguistic
experience based on residential history may produce large effects on
how listeners perceive and categorize dialect variation.
In a related study, Niedzielski (1999)
reported the results of an experiment in which listeners'
perception of vowel quality was manipulated based on their beliefs
about the talker. In particular, she asked participants from Detroit to
listen to sentences and then try to match the vowel quality of the
final word in the sentence to one of six synthetic vowel tokens. One
group of listeners was told that the talker was from Detroit; the other
group was told that the talker was from Canada. The listeners in the
Canadian group selected the actual matching vowel tokens, whereas the
Detroit group selected canonical vowels as the best match.
Niedzielski's results suggest that perception is affected not only
by actual linguistic experience with dialect variation, but also by the
beliefs a listener has about the talker and where the talker may come
from.
The goal of the present study was to further investigate the role of
early linguistic experience on dialect categorization performance by
explicitly manipulating the residential history of the listeners. In
particular, two groups of listeners with very different residential
histories were recruited from undergraduate psychology classes. The
listeners were asked to carry out the same six-alternative
forced-choice categorization task used in the earlier study by Clopper
and Pisoni (in press). One group of listeners (i.e.,
“homebodies”) had lived exclusively in the state of
Indiana; the other group of listeners (i.e., “army brats”)
had lived in three or more different states by the time they were 18
years old.
We predicted that the listeners who had lived in different states and
been exposed to greater linguistic variation would perform better on
the categorization task than the listeners who had lived in only one
state. Based on the findings reported by Williams et al. (1999), we also predicted that army brats who had
lived in a given region would be more likely to correctly categorize
the talkers from that region than army brats who had not lived in that
region. Finally, we expected to replicate the pattern of results
revealed in the earlier clustering analysis reported by Clopper and
Pisoni (in press), which showed that listeners with different
residential histories have structurally different perceptual similarity
spaces of dialect variation. Unlike the listener groups in Clopper and
Pisoni's study that were defined post hoc after the study was
completed, the listeners in the current study were specifically
recruited to meet criteria for two objectively defined groups based on
residential history.
METHODS
Talkers
Sixty-six talkers were selected from the TIMIT Acoustic-Phonetic
Continuous Speech Corpus (Fisher, Doddington, &
Goudie-Marshall, 1986; Zue, Seneff, &
Glass, 1990). The TIMIT corpus contains recordings of 630
talkers reading 10 sentences each. The documentation accompanying the
TIMIT corpus provides the age, gender, and ethnicity of each talker, as
well as a regional dialect label. The eight dialect labels used to
identify talkers on the TIMIT corpus were: New England, North, North
Midland, South Midland, South, West, New York City, or Army Brat. The
66 talkers in the present study represented six of these regions, with
11 talkers from: New England, North, North Midland, South Midland,
South, and West. To control for as many other variables as possible,
all of the talkers in this study were white males who were between the
ages of 20 and 29 at the time of recording.
Stimulus materials
The stimulus materials consisted of three meaningful English
sentences for each talker. Two of the sentences were those read by all
630 talkers on the TIMIT corpus. These two “calibration
sentences” were originally designed to elicit dialect variation
(Fisher et al., 1986; Zue
et al., 1990). They are shown in (1) below. In addition, a third
sentence was selected for each talker to ensure that a different
sentence was read by each talker and that no sentence was ever repeated
during the course of the experiment. Examples of these novel sentences
are shown in (2).
(1) a. She had your dark suit in greasy wash water all
year.
b. Don't ask me to carry an oily rag like that.
(2) a. Beg that guard for one gallon of gas.
b. A huge tapestry hung in her hallway.
Each sentence for each talker was copied into an individual sound
file which was segmented to include only speech materials using
Syntrillium's CoolEdit 96. The sound files were then leveled to 55
dB using Level16 (Tice & Carrell, 1998)
and stored digitally for later playback to listeners in the
experiment.
Listeners
One hundred and six Indiana University undergraduates (44 males and
62 females) were recruited to serve as listeners for this study. The
listeners were recruited to meet one of two residential history
requirements: either the listener must have lived his or her entire
life in Indiana or he or she must have lived in at least three
different states. All of the listeners received partial course credit
for their participation. Prior to the final analyses, data from 43
listeners were removed: three were bilingual, 14 were below chance on
all three phases of the experiment, 22 did not meet the residential
history requirements for either listener group, three had a history of
a speech or hearing disorder, one was outside the age distribution of
the rest of the population, one had a nonnative English-speaking
parent, and one set of data was lost because of technical
difficulties.
The remaining 61 listeners (27 males and 34 females) were all
monolingual native speakers of American English whose parents were also
native speakers of American English. None of the listeners reported a
history of a hearing or speech disorder at the time of testing and all
performed statistically above chance on all three phases of the
experiment. The listeners ranged in age from 17 to 22, with a mean age
of 19.0 years.
Thirty-one of the 61 listeners had lived exclusively in the state of
Indiana. They comprised the “homebodies” group. The
remaining 30 listeners had lived in at least three different states and
comprised the “army brats” group. The army brats did not
need to have lived in three different states by any specific age to
participate. However, given that all of the participants in this study
were undergraduates at a large midwestern university, the army brats
had all lived in three or more states before the age of 18.
To analyze the effects of residency1
Throughout this article, the term “residency”
is used to refer to whether or not a given participant lived in a
particular region. The residency variable, therefore, distinguishes
between those army brats who have lived in a certain dialect region
from those who have not lived there. The term “residential
history” is used to refer to all of the places that a given
participant lived. Residential history is the variable that
distinguishes the army brats from the homebodies.
on
categorization performance, the army brat group was divided up into two
groups based on residency for the New England, North, South, and West
regions. In particular, using the maps shown in
Figure 1, if a listener had lived in any of the
states included in a given dialect region, he or she was considered to
have been a resident of that region. If a listener had not lived in any
of the states in a given region, she or he was considered to be a
nonresident of that region. All of the listeners had lived in Indiana
and were therefore considered residents of the North Midland and South
Midland. Comparisons based on residency were therefore not made for
these two dialect regions. Given that these divisions were made post
hoc, the number of army brats in each residency group for each region
varied considerably. These numbers are shown in
Table 1.
The six response alternatives in the categorization task (from
Clopper & Pisoni, in press).
Number of army brats in each residency group for New England, North,
South, and West dialect regions
Procedure
The listeners were seated in front of personal computers equipped
with a mouse and a pair of Beyerdynamic DT100 headphones. The six
response alternatives were displayed on the computer screen as 2″
× 2″ icons depicting the six dialect regions. The icons
were partial maps of the United States in which the dialect region was
shaded and each icon was labeled with the name of the region. Figure 1 shows the six icons as they were arranged
on the screen. The response icons were displayed on the screen prior to
beginning the experiment and the listeners were encouraged to
familiarize themselves with the six regions.
The experiment consisted of three phases. In the first phase, the
listeners heard all 66 talkers reading the first calibration sentence.
The second phase was identical to the first, except that the listeners
heard all of the talkers reading the second calibration sentence. In
the third phase, the listeners heard each of the talkers reading a
different, novel sentence. The talkers were presented in a different
random order for each listener in each of the three phases. On each
trial, the listeners heard a sentence one time and were asked to listen
to it carefully and then to select the region on the screen that they
thought the talker was from by clicking on the icon with the mouse. No
feedback was given about the accuracy of the responses during any phase
of the experiment.
RESULTS
Categorization performance
Overall performance on the categorization task was, as expected,
barely above chance for both groups of listeners on all three
sentences. Figure 2 shows the proportion of
correct responses for each listener group on each of the sentences. A 2
× 3 repeated measures ANOVA on listener group (army brats or
homebodies) and sentence (first, second, or novel) revealed significant
main effects of listener group and sentence (F(1,181) = 5.32,
p < .05 and F(2,180) = 10.54, p <
.001, respectively). The listener group × sentence interaction
was not significant. The army brats performed better than the
homebodies overall, but planned post hoc t-tests revealed a
significant listener group difference only for the first calibration
sentence (t(59) = −2.04, p < .05). Post hoc
Tukey tests showed that performance on the first sentence and the novel
sentences was significantly better than performance on the second
sentence (p < .01 for both). Scores on the first sentence
and the novel sentences were not significantly different.
Proportion of correct responses on each phase of the categorization
task by the two listener groups. Chance performance (17%) is indicated
by the solid line. Performance significantly above chance (25%) is
indicated by the dashed line.
To determine the effects of talker dialect on categorization
performance, a 2 × 3 × 6 ANOVA on listener group (army
brats or homebodies), sentence (first, second, or novel), and talker
group (New England, North, North Midland, South Midland, South, or
West) was conducted. All three main effects were significant:
F(1,1096) = 5.71, p < .05 for listener group,
F(2,1095) = 11.76, p < .001 for sentence, and
F(5,1092) = 88.17, p < .001 for talker group.
Again, the army brats performed significantly better than the
homebodies overall. Post hoc Tukey tests on the sentence variable again
revealed that performance on the first sentence and the novel sentences
was significantly better than performance on the second sentence
(p < .01 for both). Performance on the first sentence and
novel sentences did not differ from each other.
Post hoc Tukey tests on talker group revealed that performance on the
sentences from the New England talkers was significantly better than
performance on the other groups of talkers (p < .01 for all
comparisons). Performance on the sentences from the Southern talkers
was significantly better than performance on all of the groups except
New England (p < .001 for all comparisons). Performance on
the sentences from the North Midland and South Midland talkers was
significantly better than performance on the Northern and Western
groups (p < .05 for all comparisons). Finally, performance
on the sentences from the Northern talkers was significantly better
than performance on the Western group (p < .01).
Performance on the North Midland and South Midland talkers was not
significantly different. The only significant interaction was the
sentence × talker group interaction (F(10,1086) = 6.02,
p < .001). This interaction was mainly a result of the
increased performance on the sentences from the New England talkers on
the first and novel sentences over the second sentence for both
listener groups. Figure 3 shows the
performance of each listener group for each talker group on each
sentence.
Proportion of correct responses on each phase of the categorization
task for each talker group by the two listener groups. Chance
performance (17%) is indicated by the solid line. Performance
significantly above chance (25%) is indicated by the dashed
line.
Effects of residency on categorization performance
A 2 × 3 × 4 ANOVA on residency (resident or nonresident),
sentence (first, second, or novel), and talker group (New England,
North, South, or West) was conducted on the army brat listeners only to
evaluate the effects of residency on dialect categorization. Overall,
residents performed better than nonresidents, as revealed by a
significant main effect of residency (F(1,358) = 13.57,
p < .001). The main effects of sentence and talker group
were also significant (F(2,357) = 1.38, p < .001
and F(3,356) = 55.45, p < .001, respectively).
Post hoc Tukey tests on the sentences again revealed that
categorization performance on the first sentence and the novel
sentences was better than performance on the second sentence
(p < .05 for both). The difference between the first and
novel sentences was also not significant. Post hoc Tukey tests on the
talker groups revealed that performance on New England and Southern
talkers was significantly better than performance on Northern and
Western talkers (p < .001 for all comparisons), but that
differences between performance on New England and Southern talkers and
between Northern and Western talkers were not significant. The only
significant interaction was the sentence × talker group
interaction (F(6,352) = 6.33, p < .001), again
because of the large differences in performance for both residency
groups on the New England talkers across the three sentence conditions.
Figure 4 shows the proportion of correct
responses for residents and nonresidents for the four talker groups not
found in Indiana on each of the three sentences.
Proportion of correct responses on each phase of the categorization
task for each talker group by the two residency groups. Data reflect
only the responses of the army brat listener group. Chance performance
(17%) is indicated by the solid line. Performance significantly above
chance (25%) is indicated by the dashed line.
Perceptual similarity spaces
Stimulus-response matrices were constructed for each listener group
for each sentence based on the responses collected in the perceptual
categorization task. To determine the structure of these confusions,
the 6 × 6 matrices were submitted to the Similarity Choice Model
(SCM; Nosofsky, 1985) which returned
similarity and bias parameters for each confusion matrix. The
similarity parameters indicate the degree of perceptual similarity
between any two dialect regions. That is, they provide a measure of how
similar two dialects sound to a given listener group. The bias
parameters indicate the degree of listener response bias for each of
the dialect regions and provide some indication of whether or not
listeners were selecting one response alternative more frequently than
the others.
Examination of the bias parameters revealed that neither of the two
listener groups was biased to select one response alternative over the
others for any of the three phases of the task. The goodness-of-fit of
the full SCM solution was compared to the goodness-of-fit of a
restricted SCM solution in which the similarity parameters were held
constant across both listener groups for each sentence and across all
three sentences for each listener group. The goodness-of-fit values of
the restricted models were all significantly worse than the
goodness-of-fit values of the full models, suggesting that the
structure of the similarity spaces for the two listener groups was
significantly different.
The matrices resulting from the full SCM solutions for each of the
two listener groups for each of the three sentences were then submitted
to an additive clustering scheme,2
An
additive clustering scheme was selected because of the high degree of
reciprocity between the six regions. For example, South was most often
confused with South Midland and vice versa. This kind of reciprocity is
well-modeled by clustering analyses. In addition, other spatial
analyses, such as multidimensional scaling, were inappropriate for this
data given the small size (6 × 6) of the data matrices.
ADDTREE (
Corter, 1995), to obtain a spatial
representation of the perceptual similarities between the dialect
regions. The clustering solutions for the army brats and the homebodies
are shown in
Figure 5 for the first, second,
and novel sentences. In this figure, perceptual dissimilarity is
depicted as a function of vertical distance. The perceptual distance
between any two dialects is the sum of the distances of the least
number of vertical branches connecting them.
Clustering solutions for the army brats and homebodies on each phase:
Sentence #1 (A), Sentence #2 (B), and Novel Sentences
(C).
Two basic structural patterns are displayed in these six solutions.
In one pattern, New England is by itself, South and South Midland
cluster together, and the remaining three regions (North, North
Midland, and West) form the third cluster. This clustering pattern is
found in the solutions for the army brats for the first sentence and
for both listener groups for the novel sentences. In the other pattern,
New England and North cluster together, South and South Midland cluster
together, and North Midland and West form the final cluster. This set
of clusters is found in the solution for the homebodies for the first
sentence and for both groups of listeners for the second sentence.
Therefore, for the first sentence, the two groups of listeners reveal
perceptual similarity spaces that differ in structure. For the army
brats, Northern talkers cluster with North Midland and Western talkers
on Sentence 1, but for the homebodies, the Northern talkers cluster
with the New England talkers. For the second and novel sentences,
however, the differences between the army brats and homebodies are
represented in the perceptual distances between the different dialects
and not the structure of the clusters themselves.
These results suggest that regardless of prior residential history,
both groups of listeners use three broad dialect categories for
American English regional varieties: New England; South and South
Midland; North Midland and West, with Northern talkers falling into a
cluster with New England or with North Midland and Western talkers,
depending on the linguistic content of the utterance. The difference
between the perceptual similarity spaces of the army brats and the
homebodies lies in the discriminability of the different dialects.
Given that the army brats performed better in terms of accuracy than
the homebodies, we can conclude that the army brats were better able to
discriminate the six different dialects than the homebodies. This
difference in discrimination is also revealed in the subtle differences
in the dialect similarity spaces uncovered by the clustering analysis
of the perceptual confusions.
DISCUSSION
All three of our original predictions were confirmed by this study.
First, the army brats demonstrated better overall performance on the
dialect categorization task than the homebodies. Like the school
teachers in the Williams et al. (1999) study,
the army brats in the present study were able to benefit from their
greater exposure to linguistic variation in making their categorization
responses in this forced-choice categorization task. These results are
consistent with the findings in the language acquisition literature
that more diverse early linguistic experience leads to better
performance on a wide range of speech perception tasks (Polka, 1992; Tees & Werker,
1984). General exposure to language variation either in the form
of different languages (as in Polka, 1992) or
different dialects (as in the present study) seems to have lasting
effects on the perception of linguistic and indexical properties of
spoken language.
It should be noted here that most of the second-language acquisition
work with adults on the perception of nonnative segmental contexts has
typically involved the use of both identification and discrimination
tasks (e.g., Polka, 1992, 1995). The present study only required participants
to make perceptual categorization judgments. The listeners were never
directly asked to discriminate differences between talkers of different
dialects. A dialect discrimination task using paired-comparison methods
is an essential next step in this research on the effects of early
linguistic experience on the perception of indexical properties of
speech. Converging evidence from such a discrimination task for the
relationship between exposure to variation and the perception of
variation would provide further support for the claim made here that
early exposure to multiple regional dialects affects a listener's
ability to perceive linguistic variation.
Second, specific experience with a given variety also produces
effects on speech perception, as demonstrated by the results of the
residency comparisons within the army brat group. We found that a
history of residence in a given region led to better categorization
performance on talkers from that region. This result supports the
finding by Williams et al. (1999) that Welsh
schoolboys could better categorize talkers from their own region than
those from other regions. The division of the army brats into resident
and nonresident groups was based only on whether or not the listeners
had lived for any period of time in New England, North, South, or West.
Factors such as how long they lived there or at what age they lived
there were not taken into consideration. The robust finding that
residents performed better than nonresidents on categorizing talkers
from a given region suggests that exposure to variability matters and
that prior experience with linguistic variation shapes speech
perception. These results are consistent with the finding that early
linguistic experience with a specific segmental contrast has lasting
effects on speech perception (Tees & Werker,
1984). Early experience with linguistic variation appears to
have lasting effects on perceptual dialect categorization abilities
even in young adults like those studied here.
Although age and length of residency were not controlled in
developing our criteria for residency within the army brat group, the
fact that all of our participants were younger than 23 years old means
that the relevant linguistic experience with dialect variation had to
have occurred at a fairly early age. In his study of dialect
acquisition, Chambers (1992) found that
children as old as 17 could acquire some new dialect features in
production, suggesting that they were also able to perceive dialect
differences between native and nonnative dialects. In addition, Bohn
and Flege (1997) found that differences in
nonnative perceptual skills were based more on the amount of experience
than on the age of acquisition. However, Bond and Adamescu (1979) found a monotonic relationship between age
and discrimination ability of a nonnative contrast in a study on
English-speaking children (4-year-olds), adolescents
(11–13-year-olds), and adults, although the difference between
the adults and the adolescents was not statistically significant.
Future studies of the effects of exposure to linguistic variation
should consider such issues as critical periods and length of residency
more carefully and obtain quantitative measures of these variables,
because they have been shown to be relevant in related work on the
contribution of linguistic experience, age of acquisition, neural
development, and perceptual abilities to performance in these
behavioral tasks (Bailey, Bruer, Symons, &
Lichtman, 2001).
Finally, in addition to our findings on differences in categorization
performance based on residential history, the responses obtained in
this study were also submitted to a clustering analysis, which revealed
structural differences in the perceptual similarity spaces of the army
brats and homebodies. The SCM and ADDTREE analyses revealed spatial
differences between the army brats and the homebodies in how they
perceive and represent the similarities between the dialect regions.
These differences in perceptual similarity were found mainly in the
degree of perceptual distance between the different dialects. Overall,
both the army brats and the homebodies appeared to use three major
dialect clusters: New England; South and South Midland; North Midland
and West. Northern talkers were grouped perceptually with the New
England talkers or with the North Midland and Western talkers,
depending on sentence context.
These three broad dialect clusters correspond closely to the three
major regional dialects of American English that Labov (1998) defined based on his acoustic analyses of
vowel productions: North, South, and West. These three clusters are
also similar to the three major perceptual clusters reported in Clopper
and Pisoni (in press) and the major north–south perceptual
boundary reported by Preston (1993).
Although the main differences in categorization performance between
the two listener groups were related to perceptual distance, two other
findings were consistent across all of the clustering solutions. First,
the South branch of the South/South Midland node was always longer
than the South Midland branch of this node, suggesting that the
sentences from the Southern talkers were more discriminable and
perceptually distinct from the other dialects than those from the South
Midland talkers. Second, the New England branch was always the longest
individual branch on the trees, suggesting that the New England dialect
was also very perceptually distinct. The fact that our listeners
perceived New England and Southern talkers as being the most distinct
is not surprising. Preston (1986, 1989) gave groups of undergraduates maps of the
United States with state boundaries and asked them to draw and label
the places where they thought people speak differently. He found that
his participants, who were undergraduates in Michigan, Indiana, and
Hawaii, almost always indicated an area that they labeled as South and
an area in the northeast labeled variably as New York City, East Coast,
or New England. Thus, the Southern and Northeastern varieties of
American English are clearly both perceptually and culturally salient
for naïve listeners, regardless of what experimental methodology
is used to assess performance.
