Flege Schirru Interaction Between SC 2003

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/220119799
Interaction between the native and second language phonetic

subsystems
Article in Speech Communication · June 2003

DOI: 10.1016/S0167-6393(02)00128-0 · Source: DBLP
CITATIONS READS
429 1,223
3 authors, including:
James Emil Flege Carlo Schirru

University of Alabama at Birmingham Università degli Studi di Sassari
253 PUBLICATIONS 23,551 CITATIONS 14 PUBLICATIONS 636 CITATIONS
SEE PROFILE SEE PROFILE
All content following this page was uploaded by James Emil Flege on 29 July 2021.
The user has requested enhancement of the downloaded file.

Speech Communication 40 (2003) 467–491
www.elsevier.com/locate/specom
Interaction between the native and second language

phonetic subsystems
a,*
James E. Flege , Carlo Schirru b, Ian R.A. MacKay c
a
Division of Speech and Hearing Sciences, University of Alabama at Birmingham, CH20, Room 119, 1530 3rd Avenue,
South Birmingham, AL 35294-2042, USA
b
Department of Linguistics, University of Padua, 35137, Padua, Italy
c
Department of Linguistics, University of Ottawa, 70 Laurier Avenue East, Ottawa, Ont., Canada KIN 6N5
Received 1 April 2001; received in revised form 2 January 2002; accepted 20 May 2002
Abstract
The underlying premise of this study was that the two phonetic subsystems of a bilingual interact. The study tested
the hypothesis that the vowels a bilingual produces in a second language (L2) may differ from vowels produced by
monolingual native speakers of the L2 as the result of either of two mechanisms: phonetic category assimilation or
phonetic category dissimilation. Earlier work revealed that native speakers of Italian identify English /ei / tokens as
instances of the Italian /e/ category even though English /ei / is produced with more tongue movement than Italian /e/ is.
Acoustic analyses in the present study examined /ei /s produced by four groups of Italian–English bilinguals who differed
according to their age of arrival in Canada from Italy (early versus late) and frequency of continued Italian use (low-L1-
use versus high-L1-use). Early bilinguals who seldom used Italian (Early-low) were found to produce English /ei / with
significantly more movement than native English speakers. However, both groups of late bilinguals (Late-low, Late-
high) tended to produced /ei / with less movement than NE speakers. The exaggerated movement in /ei /s produced by the
Early-low group participants was attributed to the dissimilation of a phonetic category they formed for English /ei /
from Italian /e/. The undershoot of movement in /ei /s produced by late bilinguals, on the other hand, was attributed to
their failure to establish a new category for English /ei /, which led to the merger of the phonetic properties of English /ei /
and Italian /e/.
Ó 2002 Elsevier Science B.V. All rights reserved.
Keywords: Bilingualism; Second language acquisition; Vowel production; Language interaction; Tongue movement; English; Italian
1. Introduction one) the age of first exposure to the L2 has been

indexed by immigrants’ age of arrival (AOA) in a
Many studies have shown that the age of first predominantly L2-speaking country. Research has
exposure to a second language (L2) exerts a strong demonstrated better performance in the L2 by
and usually persistent influence on overall success individuals who arrived in childhood (‘‘early’’
in acquiring the L2. In some studies (including this bilinguals) than by individuals who arrived in late
adolescence or early adulthood (‘‘late’’ bilinguals).
Age effects have been observed for the production
*
Corresponding author. of L2 consonants (e.g. Flege et al., 1995b; MacKay
E-mail address: jefl[email protected] (J.E. Flege). et al., 2001) and the perception of L2 consonants
0167-6393/02/$ - see front matter Ó 2002 Elsevier Science B.V. All rights reserved.
doi:10.1016/S0167-6393(02)00128-0
468 J.E. Flege et al. / Speech Communication 40 (2003) 467–491
(Yamada, 1995; MacKay et al., 2001). Age effects amination of United States census data suggested
have also been observed for the production of L2 to Stevens (1999) that age effects on immigrants’
vowels (Flege, 1992; Munro et al., 1996; Flege learning of English as an L2 in the United States
et al., 1999a; Piske et al., 2002), L2 vowel percep- might be attributed to ‘‘social and demographic
tion (Flege et al., 1999a), overall degree of foreign considerations’’. For example, child immigrants
accent in L2 sentences (Flege et al., 1995a; Yeni- are usually enrolled in a school where they interact
Komshian et al., 2000; Piske et al., 2001) and the frequently with native speakers of English, whereas
recognition of L2 words presented in noise (Mayo adult immigrants often enter the workplace where
et al., 1997; Meador et al., 2000). they interact frequently with fellow native speakers
The literature has offered three general types of of their L1. Early bilinguals are also more likely to
explanation for age effects on L2 speech perfor- marry L2 native speakers than late bilinguals are.
mance. One explanation is that as the age of first Economic and social factors may well be linked to
exposure to the L2 increases, the mechanisms used or even cause variation in motivation to learn the
in L1 speech acquisition operate less effectively due L2 well (Gardner and Lambert, 1972).
to maturational constraints. For example, some Speech research has suggested that the accuracy
researchers attribute the widespread presence of with which L2 phonetic segments are produced
foreign accent in late bilinguals to the passing of a depends on how much native-speaker input as
critical period for speech acquisition (Scovel, 1988; opposed to foreign-accented L2 input is received
Patkowski, 1989; Mack, In press). 1 (compare Flege and Eefting, 1987, to Flege, 1991).
A second type of explanation offered for age For example, Flege and Liu (2001) examined the
effects is that late bilinguals receive less adequate identification of word-final English consonants by
L2 phonetic input than early bilinguals usually do. groups of Chinese speakers who had lived in the
Grosjean (1982) observed that the contexts in United States for averages of 2 and 7 years. Half of
which languages are learned and used influence a the participants in the 2-year and 7-year residence
bilingual’s performance in both the L1 and the L2. groups were enrolled as full-time students at an
He observed that the L2 may become a bilingual’s American university, whereas the other half had
dominant language if it is used more than the L1 occupations (e.g. laboratory technician) that were
and is needed for a wider range of everyday activ- likely to reduce the frequency of interactions with
ities, and that the dominant language is likely to native English (NE) speakers. The long-residence
develop to a greater extent than the non-dominant students obtained significantly higher identifica-
language does. Work by Jia and Aaronson (1999) tion scores than the short-residence students did,
provided cross-sectional and longitudinal evidence but there was no difference between two groups of
suggesting that child immigrants to the United non-students who differed in length of residence.
States from China received more L2 input from The students and non-students did not differ in
native English speakers than adolescent immi- terms of self-reported percentage use of English
grants did, and so were more likely to become (roughly 50% for both groups). This led to the in-
English-dominant (see also Grenier, 1984). An ex- ference that what differentiated the long-residence
students and non-students was not how frequently
they used English, but with whom.
1
A third type of explanation for age effects is
One might hypothesize that perceptual differences between
L2 learners and monolingual native speakers of the L2 is due to
language interaction. Were completely native-like
the loss of ability to re-weight the features used in decoding performance in an L2 to be observed, it would
phonetic segments following perceptual attunement to the L1 imply that bilinguals can prevent their L1 and L2
phonetic system (see Sebastian-Galles and Soto-Faraco, 1999, sub-systems from influencing one another. How-
p. 112). Or, one might hypothesize that as the L2 phonetic ever, most theorists have rejected the notion that
system develops, cross-language phonetic differences that are
detectable auditorily in some task conditions tend to be filtered
the L1 and L2 phonetic subsystems of a bilingual
out by a kind of cross-language ‘‘categorical perception’’ (see can be separated completely (e.g. Lambert and
Halle et al., 1999, p. 302). Rawlings, 1969; Paradis, 1978; Grosjean, 1989,
J.E. Flege et al. / Speech Communication 40 (2003) 467–491 469
1997, 1999). The speech learning model, or SLM likelihood of a category being formed for an L2
(Flege, 1995, 1999, 2002) posits that the phonic speech sound varies inversely as a function of its
elements making up the L1 and L2 phonetic sub- degree of perceived dissimilarity from the closest
systems of a bilingual exist in a ‘‘common pho- L1 speech sound.
nological space’’, and so will necessarily influence According to the SLM, L1 and L2 speech
one another (see also Mack, 1989). sounds interact through two distinct mechanisms.
The aim of this study was to investigate the third The first mechanism, ‘‘category assimilation’’, is
explanation for age effects on L2 speech acquisi- thought to operate when a new category fails to be
tion, viz. language interaction. It did so by testing established an L2 speech sound despite audible
predictions generated by a language interaction differences between it and the closest L1 speech
model, the SLM, regarding the production of En- sound. By hypothesis, category formation will be
glish vowels. The participants examined in this blocked if instances of an L2 speech category
study were native speakers of Italian who learned continue to be identified as instances of an L1
English when they emigrated from Italy to Canada. category. The SLM predicts that in such cases, a
The Italian–English bilinguals were assigned to one ‘‘merged’’ category will develop over time that
of four groups based on an orthogonal variation in subsumes the phonetic properties of the percep-
AOA and self-reported percentage L1 (Italian) use. tually linked L1 and L2 speech sounds. Consider,
Previous research led to the expectation that the for example, a native speaker of a language in
early bilinguals would produce English vowels which /t/ is implemented as a short-lag stop having
more accurately than the late bilinguals would. an average voice onset time (VOT) value of about
Previous research also suggested that the bilinguals 20 ms (e.g. French or Spanish). Such a person who
who used Italian seldom would produce English later learns an L2 in which /t/ is implemented as a
vowels more accurately than those who continued long-lag stop having an average VOT value of
to use Italian relatively often (Flege et al., 1997; about 80 ms (e.g. English) might develop a merged
Guion et al., 2000; Meador et al., 2000; MacKay L1–L2 /t/ category that specifies an intermediate
et al., 2001; Piske et al., 2001, 2002). VOT value. 2 By hypothesis, the properties speci-
The SLM proposes that the capacity for speech fied by a merged L1–L2 category can be modeled
learning remains intact across the life span. Ac- as a probability–density function (see, e.g. Klu-
cording to the SLM, age effects arise primarily ender et al., 1998) reflecting all tokens of the per-
from age-related changes in how the L1 and L2 ceptually linked L1 and L2 sounds that have been
phonetic subsystems interact. The phonetic cate- experienced (with recent tokens perhaps being
gories used to produce and perceive the phonetic given greater weight than tokens encountered in
segments distinguishing L1 words are hypothe- the distant past; see Sancier and Fowler, 1997). L2
sized to become more powerful attractors of L2 learners often begin by producing L2 words using
vowels and consonants as they develop through unmodified L1 phonetic segments, but they typi-
childhood and into adulthood (e.g. Parnell and cally approximate L2 phonetic norms for certain
Amerman, 1978; Lee et al., 1999; Hazan and L2 phonetic segments more closely over time as
Barrett, 1999; Walley and Flege, 2000; Johnson, they gain experience in the L2 (e.g., Flege and
2000). That is, as L1 vowels and consonants de- Port, 1981). Given that a single, merged L1–L2
velop, they will perceptually assimilate neighbor- category is used to produce corresponding speech
ing L2 vowels and consonants more strongly (see sounds in the L1 and L2 in the absence of category
Baker et al., 2002). This leads to the prediction formation, the SLM predicts that the more a
that, all else being equal, early bilinguals will be
more likely to establish new phonetic categories
for L2 speech sounds than late bilinguals will be. 2
This example greatly oversimplifies the complexity of
Of course, some L2 speech sounds are too similar phonetic category representations, for phonetic categories
to L1 speech sounds for new category formation to specify an array of co-varying, contextually sensitive properties
occur. Another hypothesis of the SLM is that the that differ in perceptual weight and auditory salience.
bilingual approximates the phonetic norm for an It will cause a newly established L2 category and
L2 speech sound, the more her production of the the nearest L1 speech category to shift away from
corresponding L1 speech sound will tend to di- one another in phonetic space. The SLM posits
verge from L1 phonetic norms. that category dissimilation occurs because biling-
Flege (1987) observed the operation of phonetic uals strive to maintain phonetic contrast between
category assimilation in a study examining the all of the elements in their combined L1 þ L2
production of /t/ in French and English words by phonetic space in the same way that monolinguals
late English–French and French–English biling- (or human languages, see Lindblom, 1998) strive
uals. English /t/ is produced with longer VOT val- to maintain phonetic contrast among the elements
ues than French /t/ is. The /t/s produced by the making up their (L1-only) phonetic space.
bilinguals in both of their languages tended to have Mack (1990) obtained evidence for the opera-
VOT values that were intermediate in value to the tion of category dissimilation in a case study ex-
values observed in stops produced by French and amining a bilingual child. Flege and Eefting (1987)
English monolinguals. This suggested that the bil- obtained evidence of category dissimilation in a
inguals’ L1 /t/ had influenced their productions of study examining VOT in the production of Span-
/t/ in the L2, and vice versa. MacKay et al. (2001) ish /p t k/. Both of two groups of early Spanish–
examined the production and perception of voiced English monolinguals produced significantly
stops by Italian–English bilinguals. Italian /b d g/ longer VOT values in English than Spanish /p t k/;
are produced with lead VOT values (i.e. pre-voic- and both groups produced Spanish /p t k/ with
ing) whereas English /b d g/ are typically produced shorter VOT values than did age-matched groups
with short-lag VOT values. The bilinguals tended of Spanish monolinguals. This suggested that the
to misidentify short-lag tokens of English /b d g/ early bilinguals’ Spanish /p t k/ categories dissim-
as /p t k/. They tended to produce English voiced ilated from categories they established somewhat
stops with pre-voicing more often than NE later in life for English /p t k/. A later study (Flege
speakers did, but to pre-voice Italian voiced stops and Eefting, 1988) provided independent evidence
less often than Italian monolinguals did. 3 Impor- that the bilinguals had established new phonetic
tantly, production in the two languages was cor- categories for English /t/. 4
related: The less the bilinguals pre-voiced in However, two other studies did not show a
English (and so approximated L2 phonetic norms), shortening of VOT values in the L1 /p t k/ after
the less they tended to do so in Italian (and thus to English was learned as an L2. Mack et al. (1995)
diverge from L1 phonetic norms). observed no difference in the VOT values pro-
The second mechanism through which L1 and duced in French /p t k/ by monolingual French
L2 phonetic segments are hypothesized to interact children and French–English bilingual children
is called ‘‘phonetic category dissimilation’’. This living in France. This finding does not provide
mechanism is thought to operate when a new cat- counter-evidence to the hypothesized role of pho-
egory has been established for an L2 speech sound. netic category dissimilation because the bilingual
children seemed to have lacked the phonetic input
3
The finding obtained in a recent perceptual experiment also
4
suggested the operation of category assimilation. One might Participants in the Flege and Eefting (1988) study rapidly
expect late Italian–English bilinguals to discriminate Italian imitated the randomly presented members of a VOT continuum
vowels better than early Italian–English bilinguals because their that ranged from a pre-voiced /d2/ to a long-lag/t2/. English
Italian vowel system was better established when they began monolinguals produced stops having the predominantly short-
learning English and because they typically use Italian more lag and long-lag VOT values typical for English /d/ and /t/.
than early bilinguals do. However, Flege et al. (1999a) observed Spanish monolinguals produced stops having the pre-voiced and
a poorer discrimination of Italian vowels by late than early short-lag VOT values that are typical for Spanish /d/ and /t/. The
bilinguals, suggesting that the late bilinguals’ Italian vowels had early bilinguals, on the other hand, produced stops having pre-
changed as the result of category assimilation in the absence of voiced, short-lag, and long-lag VOT values when imitating the
category formation for English vowels. same set of synthetic stimuli.
needed to establish long-lag VOT categories for does not influence the production of L1 vowels, as
English /p t k/. 5 However, counter-evidence was predicted by the SLM. However, the mechanism
obtained by Flege (1991), who observed no might not have operated due to the absence of
shortening of VOT in Spanish /p t k/ by early phonetic category formation for French /u/. All 18
Spanish–English bilinguals. Differences in the na- participants produced French /u/ with English-like
ture of the L2 /p t k/ categories that were formed F2 values that were too high for French.
may explain why VOT shortened in the production The present study evaluated the production of
of Spanish /p t k/ by early Spanish–English bil- English /ei / to determine if its production would
inguals in the Flege and Eefting (1987) study but manifest the predicted effect of phonetic category
not those in the Flege (1991) study. Participants in dissimilation. Several factors led to the selection of
the Flege and Eefting (1987) study were exposed this vowel for analysis. As reported below, English
primarily to Spanish-accented English and pro- /ei / is produced with far more formant movement
duced English /p t k/ with shorter VOT values than than Italian /e/ is. 6 Recent research with Italian
NE speakers did, whereas participants in the Flege adults who were inexperienced in English (Flege and
(1991) study were exposed primarily to native- MacKay, Submitted) revealed that /ei / tokens were
produced English and produced English /p t k/ perceptually assimilated by Italian /e/, but were
with native-like VOT values. If this explanation is nevertheless judged to differ more from Italian /e/
correct, it means that category dissimilation will than English /u/ tokens were judged to differ from
occur only if a new L2 category is relatively close Italian /u/. 7 Finally, Flege et al. (1999a) noted that
in phonetic space to a pre-existing L1 category. just one of four groups of Italian–English bilinguals
Another study failed to provide evidence of examined––early bilinguals who seldom used Ital-
phonetic category dissimilation in the production of ian––were able to discriminate /ei / and /e/ tokens at a
vowels. English /u/ is anterior in the vowel space significantly above-chance rate. This was inter-
with respect to French /u/; that is, it is a ‘‘fronted’’ preted to mean that an early exposure to English
vowel produced with higher F2 values than French
/u/ is. Flege (1987) examined the production of
English /u/ and French /u/ by groups of monoling- 6
The English vowel in ‘‘code’’ (symbolized here as /o/) is
uals and bilinguals. The F2 values in French /u/ probably also produced with more formant movement than its
would not be expected to change as the result of Italian counterpart (/o/), but there is evidence that English /ei / is
phonetic category dissimilation because French /u/ produced with more movement than English /o/ is (Flege, 1989;
see also Flege et al., 1986, Figs. 5 and 7). This suggests that
is maximally posterior due to physiological limita- English /ei / may differ more from Italian /e/ than English /o/
tions. English /u/ might be produced with even differs from Italian /o/. If so, then category formation should be
higher F2 values than is typical for English due to more likely for English /ei / than /o/ according to the SLM.
7
dissimilation. However, the English–French biling- Eleven Italian university students who had lived in Ottawa
uals’ English /u/s had only slightly higher F2 values for just three months classified multiple natural tokens of the
Canadian English vowels /i I e ei æ Z d u o/ in terms of one of
than vowels spoken by English monolinguals. This the seven vowels of standard Italian (/i e e a o u/). The
c
might mean that phonetic category assimilation students also rated each vowel token for goodness (1 ¼ very
different, 5 ¼ very similar) as an instance of the Italian vowel
used to classify it. The modal classifications of the English
vowels /i/, /ei /, /Z/, /o/ and /u/ were unsurprising: Italian /i/, /e/,
5
The children’s primary source of English input was their /a/, /o/ and /u/, respectively. However, the goodness of fit ratings
American or British mothers, who had lived in France for 9–14 suggested that some of the English vowels differed to a greater
years. The results of Flege (1987) suggest that the mothers may extent from the closest Italian vowel than others did. For
have produced English /p t k/ with VOT values that were example, the English /ei / tokens received a lower rating (mean
intermediate to the short-lag and long-lag VOT values typical 3.3) than the English /u/ tokens did (mean 4.2) but a higher
for French and English, respectively. Perhaps because of this, rating than the /d/ tokens did (mean 1.7). The Italian students
four of the seven children produced English /p t k/ with VOT may have given lower goodness ratings to the English /ei /
values that were much closer (range: 27–37 ms) to the mean tokens than to the English /u/ tokens because of /ei /’s lower
value observed for French monolinguals (26 ms) than to the position in vowel space than Italian /e/ or to the greater formant
mean value observed for English monolinguals (78 ms). movement in English /ei / than in Italian /e/.
and a relatively infrequent use of the L1 promoted cluded, bilingual participants were required to
phonetic category formation for English /ei /. have been born in Italy and to have arrived in
These findings, when taken together with the Canada between the ages of 2 and 30 years. They
hypotheses of the SLM, led to the two predictions were also required to indicate either a relatively
tested in the present study. Bilinguals who continue frequent or a relatively infrequent use of Italian
to judge /ei / tokens to be instances of Italian during a telephone pre-screening. The bilinguals
/e/––which is more likely to hold true for late than all lived in English-speaking neighborhoods lo-
early bilinguals––will fail to establish a category for cated Ottawa, Ontario at the time of testing. As
English /ei /. Such bilinguals should produce /ei / with part of a language background questionnaire ad-
less movement than NE monolinguals, but with ministered before testing, the bilinguals were asked
more movement than is typical for Italian /e/ as the to estimate their percentage use of Italian in the
result of category assimilation. Bilinguals who preceding five years, five months, and five weeks.
manage to establish a new category for English The three estimates were strongly correlated, and
/ei /––which should occur most often for early bil- so were averaged. The bilinguals’ mean estimates
inguals who seldom use Italian––should produce /ei / of percentage Italian use, which are shown in
with even more movement than NE monolinguals Table 1, appear to have been valid and reliable. 9
do as the result of phonetic category dissimilation. The bilinguals were assigned to four groups of
These predictions were tested through acoustic 18 each based on an orthogonal variation in AOA
analysis of formant movement patterns in English and percentage Italian use. The 36 bilinguals who
/ei /. The study was organized as follows. Section 2 arrived in Canada between the ages of 2 and 13
describes the procedures used to select participants years have been designated ‘‘early’’ bilinguals. 10
and elicit their production of consonant–vowel– The 36 native Italian speakers who arrived in
consonant (CVC) English words. Section 3 pre- Canada between the ages of 15 and 26 years have
sents the results of analyses examining listeners’ been designated ‘‘late’’ bilinguals. Subgroups of
judgments of the English vowels produced by the early and late bilinguals differed according to self-
four groups of Italian–English bilinguals and by reported percentage Italian use. The individuals
the participants in a NE control group. Section 4
presents the results of acoustic analyses that fo-
9
cused on the production of /ei /. As part of the language background questionnaire, each
bilingual was asked to name the persons with whom they
‘‘typically’’ or ‘‘sometimes’’ spoke Italian. The high-L1-use
bilinguals named significantly more persons than the low-L1-
2. Method use bilinguals did, F ð1; 68Þ ¼ 34:1, p < 0:01. However, the
number of persons named by the two groups of early bilinguals
2.1. Participants and by the two groups of late bilinguals did not differ
significantly, F ð1; 68Þ ¼ 0:3, p > 0:10, nor did the AOA L1
use interaction reach significance in the analysis of the number
The mean characteristics of the five groups of
of named interlocutors, F ð1; 68Þ ¼ 0:27, p > 0:10. The biling-
participants are presented in Table 1. All 90 of the uals were asked how much they used Italian at home, at work,
participants were either members of a predomi- at social events, while shopping, while speaking on the
nantly Italian Roman Catholic parish in Ottawa telephone, with friends, and with family members. The high-
where the testing took place or were socially con- L1-use bilinguals reported a higher percentage use of Italian
nected to it in some way. 8 The participants in four than the low-L1-use bilinguals did in each context, whereas
there was little difference between the early and late bilinguals.
groups were Italian–English bilinguals. To be in- When an average was computed for the contextualized L1 use
estimates, these estimates were strongly correlated with the
average Italian percentage use estimates in Table 1,
8
Many of the participants lived in the vicinity of the Roman F ð1; 70Þ ¼ 0:91, p < 0:01.
10
Catholic parish where the tests were administered when they The terms ‘‘early’’ and ‘‘late’’ bilinguals are used here for
first arrived from Italy in the 1950s and 1960s. However, most convenience, and should not be taken as an implicit claim
of the participants subsequently moved to outlying English- regarding the state of neurological development associated with
speaking suburbs. a particular chronological age.
Table 1
Characteristics (means, SDs, ranges) of the five groups of participants
Gender Age AOA %Use LOR NII EDUC
Native English 9m 50(4) – – – – –
9f 39–57
Early-low 8m 50(4) 7(3) 7%(4) 42(4) 2.8(1.4) 14(3)
10 f 42–58 2–13 1–13 36–50 1–7 10–18
Early-high 8m 49(6) 8(4) 43%(15) 40(4) 5.0(1.7) 11(6)
10 f 35–61 2–13 25–80 33–49 2–10 2–24
Late-low 10 m 51(7) 20(3) 10%(5) 31(8) 2.6(1.0) 2(2)
8f 29–62 15–25 2–15 4–42 0–5 0–6
Late-high 8m 49(8) 20(3) 53%(13) 29(9) 4.4(1.6) 2(2)

10 f 29–57 15–26 30–75 8–39 2–7 0–8
M 49(6) 14(7) 28%(23) 36(9) 3.7(1.8) 7(6)

Note: Age, chronological age, in years; AOA, age of arrival in Canada, in years; %Use, self-reported percentage use of Italian; LOR,
length of residence in Canada, in years; NII, number of interlocutors with whom Italian was used; EDUC, years of education in
Canada, in years.
who reported using Italian between 1% and 13% of tory of auditory disorder, and all passed a pure-
the time were designated the ‘‘low-L1-use’’ bi- tone hearing screening at octave frequencies be-
linguals; those who reported using Italian between tween 500 and 4000 Hz (re: 35 dB HL) prior to
25% and 85% of the time were designated the participating.
‘‘high-L1-use’’ bilinguals. The four groups formed Characteristics of the four groups of bilinguals
in this way were named ‘‘Early-low’’ (early bi- (Table 1) were assessed in a series of AOA (early,
linguals who seldom used Italian), ‘‘Early-high’’ late) L1 use (low-L1-use, high-L1-use) ANO-
(early bilinguals who used Italian often), ‘‘Late- VAs. As intended, the high-L1-use bilinguals re-
low’’ (late bilinguals who seldom used Italian), and ported a higher percentage use of Italian than the
‘‘Late-high’’ (late bilinguals who used Italian of- low-L1-use bilinguals did, F ð1; 68Þ ¼ 267:7, p <
ten). The bilinguals were born in one of 13 Italian 0:01. The late bilinguals reported using Italian
regions (Abruzzo-24, Calabria-12, Sicilia-8, Ve- more than the early bilinguals did, F ð1; 68Þ ¼ 7:3,
neto-7, Campania-6, Basilicata-4, Lazio-3, Friuli- p < 0:01. However, as intended by the design, the
2, Puglia-2, Lombardia-1, Marche-1, Piemonte-1, AOA and L1 use factors did not interact signifi-
Toscana-1). Place of birth did not vary systemat- cantly in the analysis of self-reported percentage
ically across the four bilingual groups. Italian use, F ð1; 68Þ ¼ 2:45, p > 0:10. The early
The mean age of the bilinguals, 49 years, was bilinguals arrived in Canada at significantly earlier
comparable to that of the participants in the NE ages than the late bilinguals had, F ð1; 68Þ ¼ 227:3,
control group. This group consisted of 18 native p < 0:01. However, the low-L1-use and high-L1-
speakers of English who were born and raised in use bilinguals’ AOAs did not differ significantly,
the Ottawa, Ontario region. The NE speakers were F ð1; 68Þ ¼ 0:7, p > 0:10. As intended by the de-
‘‘monolingual’’ in the sense that they did not use sign, the interaction between AOA and L1 use was
any language other than English in their daily lives non-significant in the analysis of the AOA values,
(Grosjean, 1982). They all had some knowledge of F ð1; 68Þ ¼ 0:9, p > 0:10.
French because this language is usually studied by Nearly all of the bilinguals were highly experi-
Canadian anglophones at school; however, none enced in English. All but two had lived in Canada
of them reported speaking French well or using it for more than 10 years; and all but three had lived
often. None of the 90 participants reported a his- there for at least 20 years. The bilinguals’ length of
residence (LOR) in Canada was confounded with Table 2

AOA because the four groups of bilinguals were The stimuli used to elicit vowel production in two conditions
matched for chronological age. An ANOVA re- Vowel Elicitation condition
vealed that the early bilinguals had lived in Canada 1-Word 3-Word
longer than the late bilinguals had, F ð1; 68Þ ¼ 47:5,
/Z/ cod bad cod heard
p < 0:01. Importantly, however, the low-L1-use /d/ heard cud heard bade
and high-L1-use bilinguals’ LORs did not differ /ei / bade heed bade hid
significantly, F ð1; 68Þ ¼ 1:4, p > 0:10, and AOA /o/ code heard code bad
did not interact with L1 use in the analysis of the /u/ booed bade booed cod
/ / could hid could bed
LOR values, F ð1; 68Þ ¼ 0:0, p > 0:10.
/e/ bed bad bed hid
All of the early bilinguals were enrolled in En- /I / hid cod hid bad
glish-speaking schools in Canada when they first /æ/ bad heard bad bade
arrived in Canada. They remained in school at least / / cud bade cud heard
through the end of high school. However, most of /i/ heed code heed booed
the late bilinguals entered the work force or worked Note: Only the italicized words were analyzed. Filler material at
in the home upon arriving in Canada. Fifteen of the the beginning and end of the lists is not shown.
36 late bilinguals never attended school in Canada.
Not surprisingly, the early bilinguals had received then digitized (at 22.05 kHz with 16-bit resolution)
significantly more education in Canada than the and normalized for peak intensity. The stimulus
late bilinguals had, F ð1; 68Þ ¼ 168:9, p < 0:01. words produced by the female talker were pre-
However, the low-L1-use and high-L1-use biling- sented for repetition following the male talker’s
uals did not differ significantly in years of Canadian stimuli. Only words repeated in response to the
education, F ð1; 68Þ ¼ 2:2, p > 0:10, nor did AOA female talker’s stimuli were analyzed, however.
and L1 use interact significantly in the analysis of The words were elicited in two conditions. In
years of Canadian education, F ð1; 68Þ ¼ 1:78, the ‘‘1-word’’ condition, the stimulus words were
p > 0:10. presented one at a time at the beginning of a
carrier phrase (‘‘. . . is the next word to say’’). To
2.2. Procedures reduce the likelihood of direct imitation, the
participants were required to listen to the entire
The participants were tested by a NE-speaking utterance before repeating the variable utterance-
experimenter in a small, quiet room following the initial stimulus. The likelihood of direct imitation
administration of a pure-tone hearing screening was further reduced in the ‘‘3-word’’ condition,
and a language background questionnaire. The where the stimuli from the 1-word condition were
vowel production experiment reported here was presented as the middle word of 3-word sequences
administered following a sentence repetition task placed at the beginning of a carrier phrase (‘‘. . . are
(all 90 participants) and a sentence translation task the next words to say’’). The participants repeated
(just the bilinguals). all three words in the sequence after listening to
The participants produced /CVd/ words con- the entire utterance, but just the middle word was
taining the vowels /Z d ei o u e I æ i/ (see Table subsequently analyzed.
2). The words were elicited using a delayed repe-
tition task to avoid the influence of orthography
(see, e.g. Piske et al., 2002) and to reduce variation 3. Listener judgments
in speaking rate, which is known to influence de-
gree of formant movement in /ei / (Gay, 1968, The words elicited as just described were later
Table 2). The participants repeated the /CVd/ words digitized and presented to NE-speaking listeners.
after hearing them via a loudspeaker. The stimulus The dependent variable examined in the non-
words to be repeated were spoken by one male and parametric statistical analyses presented in Section
one female native speaker of English from Ottawa, 3.2 was the number of listeners who judged each
token as having been produced accurately, as op- monolingual (as defined above) speakers of Ca-
erationally defined below. In Section 3.3, a single nadian English with a mean age of 31 years (range:
NE-speaking phonetician transcribed subsets of 20–46 years). The listeners had an average of 5
tokens which, according to the listener data in years of post-secondary education (range: 1–8
Section 3.2, had been produced accurately or in- years). They were all born and raised in Ontario
accurately. The aim of the transcriptional analysis (Toronto-6, Ottawa-1, Hamilton-1, Fort Francis-
was to provide insight into why the listeners tended 1, Deep River-1, Brantford-1). All of them re-
to judge certain tokens as being a distorted instance ported normal hearing and passed a pure-tone
of the intended category or an instance of some hearing screening prior to participating.
other, non-target vowel. Finally, the regression The digitized words were presented via loud-
analyses presented in Section 3.4 examined the speakers at a self-selected comfortable level to the
predictive power of the two variables that had been listeners, who were tested individually in a sound
used to select the bilingual participants (AOA, booth. The 180 words containing each target vowel
percentage Italian use) on NE-speaking listeners’ (90 participants 2 conditions) were randomly
judgments of vowels the bilinguals had produced in presented in separate counterbalanced blocks to
two elicitation conditions. each listener. The stimuli comprising each block
were presented one time each in a different ran-
3.1. Method domized order to each listener. Ten practice stimuli
at the beginning of each block were not analyzed.
The participantsÕ repetitions of 1980 /CVd/ The listeners judged the vowel in each word by
words (5 groups 18 participants 11 vowels 2 clicking one of four buttons displayed from left to
elicitation conditions) were digitized at 22.05 kHz right on the screen of a personal computer. The
(16-bit resolution) using a waveform editor (Cool buttons were labeled ‘‘wrong vowel’’ (1), ‘‘distorted’’
Edit, Syntrillium Corp.), then normalized to 50% (2), ‘‘acceptable’’ (3), ‘‘good’’ (4). The listeners were
of full-scale intensity. Prior to analysis, any pre- not trained on the rating task. However, the target
voicing that was present in word-initial /b/ tokens vowel to be judged in each block was illustrated
(in ‘‘bad, bed, bade, booed’’) was edited out. This by three written keywords (e.g. ‘‘bad’’, ‘‘dad’’ and
was done to remove an unwanted disparity be- ‘‘sad’’ for /æ/) before the block began. The listeners
tween the late bilinguals, who tended to pre-voice were told to say the keywords aloud, and to use
often, and the NE speakers, who seldom pre- their own pronunciation of the keywords as a point
voiced (see MacKay et al., 2001). Similarly, the of reference when rating the intended productions
initial /h/ in three other words (‘‘heed, heard, hid’’) of each vowel category. They were told to focus
was edited out because a few bilinguals omitted their attention on the vowel in each stimulus, and
this consonant due to its absence in Italian. Finally, to ignore variation in voice quality or subjective
all portions of the words following constriction of loudness insofar as possible. The interval between
the final /d/ tokens were removed to prevent pos- each response and the next trial was 1.0 s.
sible between-group differences in the final stops
from affecting the listenersÕ vowel judgments. 3.2. Between-group differences
Vowels in the digitized words were evaluated
auditorily by NE adults (six male, five female). It The dependent variable examined in non-para-
would have been ideal to recruit a panel of NE- metric analyses was the number of listeners who
speaking listeners who were matched in age, social judged each vowel token to have been produced
class and education to the 90 participants who accurately. A vowel token was operationally de-
produced the vowels being evaluated, and who had fined as ‘‘accurate’’ if it was judged to be an ‘‘ac-
spent their entire lives in the same communities in ceptable’’ or ‘‘good’’ instance of its intended
Ontario where the 90 participants had lived and category (i.e. received a rating of 3 or 4).
were currently residing. This was not possible, The average numbers of listeners who judged
however. The 11 listeners who participated were vowels elicited in the 1-word condition to have
from the NE group. An alpha level of 0.05 (unad-

justed) was used for the pair-wise comparisons.
The effect of Group was significant for 10 of the
vowels elicited in the 1-word condition, H ð4Þ ¼
13:3 to 35.3, Bonferroni adjusted p < 0:05. Only
the effect for /i/ was found to be non-significant,
H ð4Þ ¼ 11:7, Bonferroni p > 0:10. Participants in
the Late-high group were found to have produced
10 vowels in the 1-word condition (/I e d o Z u ei
æ/) less accurately than the NE speakers did. The
pair-wise tests revealed that the Late-low partici-
pants produced seven vowels (/I e d o Z u/) less
accurately than the NE speakers did. However,
neither group of early bilinguals (Early-low, Early-
high) differed significantly from the NE group for
any of the 11 vowels examined.
As shown in Fig. 1(b), the stratification between
the two groups of late bilinguals (Late-low, Late-
high) was less clear for vowels spoken in the
3-word condition than it was in the 1-word condi-
tion. However, when the statistical procedures just
described for the 1-word condition were applied,
they revealed that both groups of late bilinguals
Fig. 1. The mean number of listeners who judged vowels elic-
ited in (a) the 1-word condition and (b) the 3-word condition to produced vowels in the 3-word condition less ac-
have been produced accurately (see text). curately than either group of early bilinguals or the
NE group. The effect of Group was significant for
all vowels in the 3-word condition except /i/,
been produced accurately are shown in Fig. 1(a). H ð4Þ ¼ 21:9 to 37.9, Bonferroni adjusted p < 0:05.
The data in this figure have been arranged in order Between-group tests revealed that participants in
of increasing accuracy. Of the 11 vowels, /e/ was the Late-high and Late-low groups produced 10
judged to have been produced accurately by the vowels (/e ei o I æ d Z u/) less accurately than
fewest listeners (mean ¼ 7:11 out of a maximum of the NE speakers did (p < 0:05). Participants in the
11) and /Z/ was judged to have been produced Early-high group differed from the NE group for
accurately by the most listeners (mean ¼ 9:10). On five vowels (/e ei o I /), whereas those in the Early-
average, vowels spoken in the 1-word condition by low group differed only for /e/.
participants in the NE and Early-low groups were Mann–Whitney tests were carried out to assess
judged to have been produced accurately by more the effects of AOA (early versus late) and amount
listeners (mean ¼ 9:5 for both) than vowels spoken of L1 use (low-use versus high-use). The test of
by participants in the other three groups (Early- AOA involved the comparison of vowels spoken
high 8.9, Late-low 7.5, Late-high 6.1). by the 36 early and 36 late bilinguals. Separate
Vowels produced in the 1-word condition were tests were carried out for vowels elicited in the
evaluated by a Kruskal–Wallis test. A separate test 1-word and 3-word conditions. The Mann–Whit-
was performed for each of the 11 vowels. When the ney U values were standardized to permit the
effect of Group (5 levels) reached significance at a computation of exact p-values, which were then
Bonferroni-adjusted alpha level of 0.05, pair-wise adjusted to compensate for the fact that 11 tests––
tests using the procedure recommended by Conover one for each vowel––were carried out.
(1980) were carried out to determine which of the The early bilinguals were found to have
four bilingual groups, if any, differed significantly produced nine vowels in the 1-word condition
(/ei I e d o Z u/) more accurately than the late The 218 selected tokens were transcribed by one
bilinguals did, z ¼ 3:20 to 4.74, Bonferroni ad- of the authors who is a native speaker of English
justed p < 0:05. (Differences between the early and with training in phonetics (JEF). The vowel tokens
late bilinguals for the remaining two vowels nar- were randomly presented to this listener in two
rowly missed reaching significance; /æ/ z ¼ 2:76, sessions. Each token was labeled as an instance of
Bonferroni p ¼ 0:06; /i/ z ¼ 2:65, Bonferroni p ¼ one of the 14 vowels and diphthongs of English.
0:08.) The early bilinguals produced all 11 vowels The 14 (6.5%) tokens that were labeled differently
in the 3-word condition except /i/ more accu- in the two sessions were transcribed a third time by
rately than the late bilinguals did, z ¼ 3:37 to 5.20, the same listener. The labeling discrepancies were
Bonferroni p < 0:05. (The difference for /i/ nar- resolved for all but one token, a front rounded
rowly missed reaching significance, z ¼ 2:82, vowel deemed unclassifiable as even a poor in-
Bonferroni p ¼ 0:052.) stance of an English vowel.
The test of L1 use involved comparisons of The transcriptions are summarized in Table 3.
vowels spoken by the 36 low-L1-use and 36 high- As expected, the tokens classified as accurate
L1-use bilinguals. None of the tests examining based on the listener data reported earlier were
vowels elicited in the 1-word condition reached transcribed as instances of their intended cate-
significance, z ¼ 0:24 to 2.21, Bonferroni adjusted gories more often than the tokens classified as
p > 0:10. Tests examining vowels spoken in the 3- inaccurate were (96% versus 35%). No obvious
word condition yielded one significant difference. generalization can be drawn regarding the inac-
More listeners judged the low-L1-use bilinguals’ curate productions of the target vowels /Z o u/,
than the high-L1-use bilinguals’ production of / / which were transcribed as instances of several
to be accurate, z ¼ 2:96, Bonferroni p < 0:05. different non-target vowel categories. However, a
generalization can be drawn regarding errors for
the high and mid front vowels /i I ei e/. When
3.3. Transcriptions produced inaccurately, these vowels tended to be
heard as vowels that were lower in vowel space
A subset of the 1980 vowels were transcribed than the target vowel. The inaccurate /i/ tokens
phonetically to provide insight into the nature of tended to be transcribed as /ei /; the /I / tokens as /ei /,
the vowel production errors made by late biling- /e/ or /æ/; the /ei / tokens as /e/ or /æ/; and the /e/
uals. This analysis focussed on vowels produced by tokens as /æ/. 11 The basis for these misidentifica-
males. (This is because one of the acoustic analyses tions is uncertain. It might have been the result of
to be reported later focused on males’ vowels.) a systematic cross-language difference in vowel
Only ‘‘accurate’’ or ‘‘inaccurate’’ vowels were in- production (see, e.g. the comparison of German
cluded in the transcription analysis. Included and English front vowels by Bohn and Flege, 1992,
among the accurate vowels were tokens that re- Fig. 1) or to a systematic tendency in how listeners
ceived a rating of 3 (acceptable) or 4 (good) by at identify ambiguous vowels. 12
least 10 of the 11 listeners. Five accurate produc-
tions of all 11 vowels in both conditions were se-
11
lected for transcription. In the case of tie scores, A reviewer noted that the word ‘‘bade’’ is pronounced
the token with a higher rating (see below) was /bæd/ rather than /bei d/ in some varieties of English. However,
this was unlikely to have been responsible for the several /æ/ for
used. Of the 110 accurate tokens, 97 were spoken /ei / substitutions observed here. Other front vowels were
by NE speakers and 13 were spoken by Early-low sometimes incorrectly realized as an [æ]-quality vowel; and /ei /
participants. The criterion used in selecting the production was elicited by having the participant’s repeat NE
inaccurate tokens was a rating of 3 or 4 by five or speakers’ productions of /bei d/.
12
fewer listeners. It was not possible to find five in- Peterson and Barney (1952) found that NE-speaking
listeners miss-classified /I / tokens spoken by fellow NE speakers
accurate tokens of all 11 vowels in both condi- as /e/ in 7% of instances, and /e/ tokens as /æ/ in 9% of instances.
tions. The 108 inaccurate tokens identified for this A later replication by Hillenbrand et al. (1995) revealed a
analysis were all produced by late bilinguals. smaller proportion of /e/-for-/I / and /æ/-for-/e/ confusions.
Table 3
Transcriptions of English vowels produced relatively well (‘‘accurate’’) or poorly (‘‘inaccurate’’) by male participants in two elicitation
conditions
Accurate Inaccurate
1-Word 3-Word 1-Word 3-Word
/i/ i–3, e –2
i
i–3, e –2
i
e –5
i
ei –3, I –2
/I / I –5 I –5 e–2, ei –1, æ–1 e–2, ei –1, æ–1
*–1 i–1
/ei / ei –5 ei –5 e–3, æ–2 æ–3, e–1, I –1
/e/ e–5 e–5 æ–4, e–1 æ–4, e–1
/æ/ æ–5 æ–5 æ–3, –1, d–1 æ–3, –1, e–1
/Z/ Z–4, –1 Z–5 Z–4 –2, Z–1, d–1
o–1
/ / –5 –5 Z–2, –1, æ–1 Z–2, d–2, –1
–1
/d/ d–5 d–5 d–5 d–4, e–1
/o/ o–5 o–5 o–3, –1 o–3, Z–2
/ / –5 –5 –4, d–1 –4, o–1
/u/ u–5 u–5 –3, u–1, o–1 o–4, d–1
Note: The numbers indicate how many tokens were labeled using each phonetic symbol (maximum ¼ 5 in all but two instances). One
token that was unclassifiable as an English vowel has been designated by an asterisk.
3.4. Regression analyses ratings obtained for vowels spoken in the two
elicitation conditions. The variables used in se-
The non-parametric analyses presented earlier lecting the bilingual participants served as predic-
suggested that variation in the bilinguals’ AOA tor variables in both analyses. As required by the
exerted a stronger influence on vowel production design of the study, these variables––AOA and
than percentage Italian use did. The early biling- percentage L1 use––were uncorrelated (see Table
uals’ vowels were judged to have been produced 4). The model developed for vowels elicited in the
accurately by more listeners than the late biling- 1-word condition accounted for 63.4% of the
uals’ vowels were in 19 of 22 possible instances (11 variance in the vowel ratings, F ð1; 70Þ ¼ 81:9,
vowels 2 conditions). However, vowels spoken p < 0:01. AOA accounted for 53.9% of the vari-
by low-L1-use and high-L1-use bilinguals differed ance at Step 1 and percentage Italian use ac-
in just one of 22 possible instances, suggesting that
the effect of L1 use on L2 vowel production was
Table 4
negligible. The aim of the analyses presented here Pearson correlations between the participant variables shown in
was to provide a more precise assessment of the Table 1
relative effects of AOA and L1 use. Following the
AOA % Ital- LOR NII EDUC
practice of previous research (Munro, 1993; Piske ian use
et al., 2002), the listener ratings described in Sec-
Age 0.18 0.04 0.61 )0.15 )0.26
tion 3.1 were treated as an interval scale and AOA 0.18 )0.66 )0.14 )0.86
subjected to parametric statistical analyses. More % Italian use )0.12 0.42 )0.25
specifically, the dependent variable examined in LOR 0.01 0.49
this section were average ratings based on each NII 0.10
participants’ production of all 11 vowels in the 1- EDUC
word and 3-word conditions. The mean ratings Note: Age, chronological age, in years; AOA, age of arrival in
obtained in this way are presented in Appendix A Canada, in years; % Italian use, self-reported percentage use of
Italian; LOR, length of residence in Canada, in years; NII,
as a function of group and vowel.
number of interlocutors with whom Italian was used; EDUC,
Separate step-wise multiple linear regression years of education in Canada, in years. One and two asterisks
analyses were carried out to examine the mean indicate significance at the 0.01 and 0.05 levels, respectively.
counted for an additional 9.5% of the variance at ceived relatively few years of education in Canada
Step 2. The model developed for vowels in the 3- (and also to have arrived in Canada relatively late
word condition accounted for 71.7% of the vari- in life and to have lived there a relatively short
ance, F ð1; 70Þ ¼ 113:2, p < 0:01. AOA accounted time) tended to use Italian relatively often.
for 61.8% of the variance at Step 1 and percentage Partial correlation analyses were carried out to
Italian use accounted for 9.9% of the variance at determine if AOA was correlated with the average
Step 2. AOA and percentage Italian use were also vowel ratings after confounded variables had been
found to account for significant amounts of vari- controlled statistically. The simple correlation be-
ance in the early and late bilinguals’ vowels when tween AOA and the ratings for vowels elicited in
these groups were examined separately. 13 the 1-word condition, rð70Þ ¼ 0:73, p < 0:01,
A well-known problem in L2 speech and lan- remained significant when the influence of LOR
guage research is that AOA is often confounded and years of Canadian education were partialled
with variables that might reasonably be expected out, rð68Þ ¼ 0:32, p < 0:01. The simple correla-
to influence L2 performance (see, e.g. Bahrick tion between AOA and the ratings obtained for
et al., 1994; Flege, 1998; Flege et al., 1999b; Yeni- vowels spoken in the 3-word condition, rð70Þ ¼
Komshian et al., 2000). Pearson correlations 0:79, p < 0:01, also remained significant when
among the participant variables shown in Table 1 LOR and years of Canadian education were par-
have been summarized in Table 4. AOA was cor- tialled out, rð68Þ ¼ 0:38, p < 0:01.
related with length of residence (LOR) in Canada Similar analyses were carried out to determine if
and the number of years of formal education in the bilinguals’ self-reported percentage Italian use
English-speaking Canadian schools. The later in was correlated with the average vowel ratings when
life the bilinguals had arrived in Canada, the the influence of a confounded variable was statis-
shorter was their LOR in Canada at the time of tically controlled. The simple correlation between
testing and the fewer years of education they had L1 use and the ratings obtained for vowels spoken
received in English-speaking Canadian schools. in the 1-word and 3-word conditions, rð70Þ ¼
Two variables were also correlated with the bil- 0:44 and )0.45, p < 0:01, indicated that the more
inguals’ percentage use of Italian: the number of the bilinguals continued to use Italian, the less ac-
specific interlocutors in Italian the bilinguals were curately they tended to produce English vowels.
able to name and years of education in Canadian These correlations remained significant when
schools. The weak correlation with the latter variation in years of Canadian education was
variable indicated that the bilinguals who had re- partialled out, rð70Þ ¼ 0:38 and )0.42, p < 0:01.
3.5. Summary
13
Separate step-wise multiple regression analyses were
carried out to examine the effect of AOA and percentage The NE listeners’ ratings of English vowels in-
Italian use on the mean ratings accorded vowels spoken by the
36 early bilinguals. Two other analyses examined the 36 late dicated that AOA influenced the Italian–English
bilinguals’ ratings. All four analyses accounted for a significant bilinguals’ production of English vowels. Non-
amount of variance in the mean ratings (F-values ranging from parametric analyses revealed that the early bil-
8.5 to 21.6 with 1,33/34 dfs, p < 0:01). Percentage Italian use inguals produced most of the English vowels
accounted for 20.0% of the variance in the early bilinguals’ examined more accurately than the late bilinguals
production of vowels in the 1-word condition. AOA accounted
for 43.9% of the variance in the early bilinguals’ production of
did. This agrees with the findings of previous
vowels in the 3-word condition at Step 1, and percentage Italian studies examining the production of English
use accounted for an additional 12.8% of the variance at Step 2. vowels by Italian–English bilinguals (e.g. Munro
For the late bilinguals, AOA accounted for 24.1% of the et al., 1996; Flege et al., 1999a; Piske et al., 2002).
variance for vowels in the 1-word condition at Step 1, and
L1 use was also found to influence the bilinguals’
percentage Italian use accounted for 13.2% of the variance at
Step 2. Percentage Italian use accounted for 26.1% of the production of English vowels. Non-parametric
variance for vowels in the 3-word condition at Step 1, and AOA tests indicated that the low-L1-use bilinguals pro-
accounted for 15.9% of the variance at Step 2. duced just one English vowel (/ /, in the 3-word
condition) more accurately than the high-L1-use ment in tokens of /ei / and /e/ that were classified by
bilinguals did. However, regression analyses ex- listeners as accurate or inaccurate in the last sec-
amining the average ratings obtained for all 11 tion. Section 4.3 compared degree of movement in
vowels revealed that percentage Italian use ac- English /ei / and Italian /e/. Section 4.4 assessed
counted for a significant amount of variance for amount of movement in the /ei / tokens produced
vowels spoken by both early and late bilinguals in by NE speakers and the four groups of Italian–
both elicitation conditions. The L1 use effect ob- English bilinguals. Finally, regression analyses in
tained here agrees with the findings obtained by Section 4.5 examined the effect of variation in
Piske et al. (2002) for groups of early Italian– movement on listeners’ judgments of /ei / produc-
English bilinguals that were matched for AOA (7 tion accuracy.
years) but differed according to percentage Italian
use (Early-low ¼ 8%, Early-high ¼ 32%). 4.1. Method
The lack of between-group differences for /i/ may
have been due to the fact that the difference between Commercially available software (the Multi-
English /i/ and Italian /i/ is too small for NE lis- speech program of Kay Elemetrics, Inc.) was used
teners to detect (even assuming that the bilinguals to make spectral and temporal measurements.
examined here used Italian /i/ without modification Three measurement locations in the ‘‘vowel’’ por-
in English words). Two possible explanations exist tion of each digitized word were identified from
for between-group differences in the production of time domain waveforms and spectrographic repre-
the remaining 10 vowels (/Z d ei o u e I æ /), sentations. Using procedures similar to those de-
which were likely to have differed from Italian scribed by Hillenbrand et al. (1995), points located
vowels that the bilinguals had acquired as young 20%, 50%, and 80% into the vowel interval were
children (see Munro et al., 1996, Fig. 4). The groups identified in each token. These points were desig-
of participants may have differed in how effectively nated the ‘‘beginning’’, ‘‘midpoint’’, and ‘‘ending’’
they adapted their production of an Italian vowel locations. A 20-ms Blackman window was centered
for use in English words. Alternatively, they may at each location and the auto-correlation method of
have differed in their ability to establish new English linear predictive coding (LPC) analysis was used to
vowel categories. The aim of the acoustic analyses estimate the frequency of the first two vowel for-
presented in the next section was to evaluate the mants (F1 , F2 ). Twenty-four LPC coefficients were
second hypothesis. calculated in most instances. The estimates ob-
tained in the LPC analyses were confirmed through
FFT analyses and, when necessary, visual inspec-
4. Acoustic analyses tion of spectrographic representations. The same
software was also used to obtain estimates of fun-
As mentioned in the Introduction, the SLM damental frequency (F0 ) at each location. Finally,
(Flege, 1995) hypothesizes that learning will pro- the duration of each vowel token was measured
ceed differently for an L2 vowel depending on from the onset to the offset of periodicity in the
whether or not a new category is established for it. vocalic portion of each waveform.
Predictions generated by the SLM were tested here
through acoustic analyses of the vowel /ei /. The 4.2. Accurate versus inaccurate vowels
first prediction was that more early than late bi-
linguals would produce English /ei / with a greater The first question of interest was whether the
amount of movement than the NE speakers. The accurate and inaccurate tokens of /ei / (see Section
second prediction was that more late than early 3.3) were produced with differing amounts of for-
bilinguals would produce English /ei / with less mant movement. To address this question, the first
movement than the NE speakers. and second formant frequency values obtained for
This section is organized as follows. Section 4.2 accurate and inaccurate /ei / tokens produced by
compared the direction and magnitude of move- males were converted from Hertz to Bark units
(Syrdal and Gopal, 1986), then plotted in a two-

dimensional (B1 versus B2 ) space. The beginning
value for each token (marked by a cross) and the
ending value (marked by a filled or unfilled circle,
depending on accuracy) was connected by a
straight line to provide a rough indication of the
change in vowel quality that occurred as each
vowel token was produced. Such changes were
likely to have occurred primarily as the result of
tongue movement (Flege et al., 1986).
The values for /ei / tokens produced in the 1-
word and 3-word conditions are shown in the two
panels of Fig. 2. Relatively low B1 and B2 values in
this figure indicates relatively high and posterior
vowel qualities, respectively. The ending values of
the accurate /ei / tokens are marked by unfilled
circles. These tokens consistently showed move-
ment toward a higher, fronter vowel quality
whereas the inaccurate /ei / tokens (whose ending
values are marked by filled circles) did not. The
ending values for the accurate /ei / tokens from both
elicitation conditions specified vowel qualities that
were higher and fronter in the vowel space than did
the ending values for inaccurate /ei / tokens. This
suggested that the tongue moved up and forward
during the production of the accurate but not the
inaccurate /ei / tokens. It can be seen that move-
ments for the inaccurate /ei / tokens were smaller in
magnitude than those for the accurate tokens, went
in the wrong direction, or both. The midpoint
values (not shown) for all but one of the inaccurate
/ei / tokens were lower in the vowel space than those Fig. 2. Frequency values obtained for tokens of /ei / elicited in
for the accurate /ei / tokens. This finding is consis- (a) the 1-word condition and (b) the 3-word condition. A
straight line connects the beginning values (crosses) and the
tent with the fact that the inaccurate /ei / tokens
ending values (filled or unfilled circle) for each token.
were transcribed as /e/ or /æ/ in Section 3.3.
Results obtained by applying the same proce-
the vowel space did the values obtained for the
dures to the accurate and inaccurate /e/ tokens
accurate /e/ tokens. This is consistent with the fact
have been plotted in Fig. 3 for the purpose of
that most of the inaccurate /e/ tokens were labeled
comparison. As expected from more direct mea-
as /æ/ in Section 3.3.
sures of tongue movement (Flege et al., 1986, Fig.
5), the accurate /e/ tokens were produced with far
less formant movement than the accurate /ei / to- 4.3. Italian /e/ versus English /ei /
kens were. The accurate and inaccurate /e/ tokens
were not distinguished by different movement The results presented in the last section suggests
patterns as was the case for the accurate and in- that some Italian–English bilinguals produced
accurate /ei / tokens considered earlier. The mid- English /ei / with too little movement due to cross-
point values (not shown) of the inaccurate /e/ language phonetic interference (Flege and Port,
tokens specified vowel qualities that were lower in 1981). This explanation assumes that Italian /e/ is
eight females) had a mean age of 26 years and were

living in Padua, Italy when recorded. The partici-
pants in the two groups began by repeating four
words in their native language following a native-
speaker model. Each word contained the vowel of
interest (/ei / for English, /e/ for Italian). After
hearing the four words again, the participants in-
serted the vowel found in all four words into a
/b_do/ context, yielding a non-word. After saying a
non-word (/bei do/ or /bedo/) in isolation, the par-
ticipants produced it two times in a carrier phrase.
This yielded 36 tokens of English /ei / and 36 tokens
of Italian /e/, all in a /b_do/ context.
The F0 , F1 and F2 frequencies of each /ei / and
/e/ token were measured at the beginning and
ending locations as described earlier. Given that
both groups contained males and females, values
in Hertz were converted to Bark values, and Bark
difference values were then computed. As dis-
cussed by Syrdal and Gopal (1986), this procedure
substantially reduces gender-based differences
arising from differences in vocal tract size. It pro-
vided estimates of vowel quality in two dimen-
sions. Subtracting the F0 value from F1 values
(B1 –B0 ) provided estimates of vowel quality in a
high–low dimension. Subtracting the F1 values
from the F2 values (B2 –B1 ) provided estimates of
vowel quality in a front–back dimension.
The mean beginning and ending values for
English /ei / and Italian /e/ are plotted in a two-
Fig. 3. Frequency values obtained for tokens of /e/ elicited in dimensional (high–low versus front–back) pho-
(a) the 1-word condition and (b) the 3-word condition. A netic vowel space in Fig. 4. The beginning and
straight line connects the beginning values (crosses) and the ending values have been connected with a straight
ending values (filled or unfilled circle) for each token. line to provide a rough indication of tongue
movement. As expected, the English /ei / tokens
produced with less movement than English /ei /, as were produced with more movement than the
is the case for /e/ in other Romance languages (e.g. Italian /e/ tokens were. Although the ending values
Delattre, 1964). The acoustic results obtained by of the two vowels were similar, their beginning
Flege et al. (1999a) supported this assumption, but values differed considerably.
this analysis was based on just five English /ei / A Language ðItalian versus EnglishÞ measure-
tokens and five Italian /e/ tokens. ment location ðbeginning versus endingÞ dimen-
The aim of the analysis presented here was to sion ðhigh–low versus front–backÞ ANOVA
compare English /ei / and Italian /e/. Speech sam- examining the Bark difference values yielded a
ples were elicited from two groups for this pur- significant three-way interaction, F ð1; 70Þ ¼ 37:3,
pose. The English monolinguals (four males, eight p < 0:01. Tests of simple main effects revealed that
females) had a mean age of 28 years. They were the beginning B1 –B0 values were smaller for Italian
residents of Columbus, Ohio or Birmingham, Al- /e/ than English /ei / (p < 0:01) indicating a higher
abama. The Italian monolinguals (four males, vowel quality at the onset of /e/ than /ei /. The be-
duced with less movement than the bilinguals’ ac-

curate /ei / tokens were. This may have been due to
the fact that the Italian vowel that most often
perceptually assimilates English /ei /, viz. Italian /e/,
is produced with less movement than English /ei / is.
The purpose of this section was to assess
amount of movement in the /ei /s produced by all
90 participants. The 180 /ei / tokens (5 groups 18
participants 2 elicitation conditions) presented
to listeners in Section 3 were measured acoustically
using the procedures described earlier.
Fig. 5 shows the mean beginning and ending
values for /ei /s spoken in both elicitation condi-
tions by the five groups in a 2-dimensional (high–
low versus front–back) formant space. The
beginning quality of the vowels spoken by all five
groups were similar, but the ending vowel quality
values of the five groups differed considerably.
Fig. 4. The mean beginning and ending frequency values for
Bilinguals in the Late-high group seem to have
productions of English /ei / and Italian /e/ by 12 monolingual
speakers of each language. produced /ei / with ending qualities that were lower
and farther back in the vowel space than the NE
ginning B2 –B1 values were larger for Italian /e/ speakers did. Conversely, the Early-low partici-
than English /ei / (p < 0:01), indicating a more an- pants seem to have produced /ei / with ending
terior vowel quality at the onset of /e/ than /ei /. vowel qualities that were higher and farther for-
However neither the ending B1 –B0 nor the ending ward in the vowel space than the NE speakers did.
B2 –B1 values obtained for Italian /e/ and English Analyses of the beginning and ending Bark
/ei / differed significantly. differences values revealed that the Early-low
The Euclidean distances between the beginning groups’ ending values were higher in vowel space
B1 –B0 and B2 –B1 values and the ending B1 –B0 and than the NE and Late-high groups’ values
B2 –B1 values were computed for each token. This (p < 0:05), and that the Early-low and Early-high
provided an estimate of amount tongue of move- groups’ ending values were fronter than the Late-
ment in a 2-dimensional vowel space. An ANOVA high groups’ ending values (p < 0:05). 14 The
confirmed that the distance values were signifi-
cantly greater for English /ei / than Italian /e/,
F ð1; 70Þ ¼ 36:8, p < 0:01. It is not certain, of 14
The Bark difference values in Fig. 5 were examined in
course, that this finding will generalize to the a group ð5 levelsÞ measurement location ðbeginning versus
vowels of all English and Italian varieties or dia- endingÞ dimension ðfront–back versus high–lowÞ ANOVA.
lects, or that the measures obtained here for Italian The three-way interaction it yielded, F ð4; 82Þ ¼ 5:9, p < 0:01,
/e/ typified the /e/s spoken by the Italian–English was explored by simple effect tests. The effect of group was non-
bilinguals when they first arrived in Canada. significant for the beginning high–low (B1 –B0 ) values, F ð4;
82Þ ¼ 0:2, n.s., and the beginning front–back (B2 –B1 ) values,
F ð4; 82Þ ¼ 0:2, n.s. However, the simple effect of group was
4.4. Formant movement in /ei / significant for both the ending high–low values, F ð4; 82Þ ¼ 5:0,
p < 0:01, and the ending front–back values, F ð4; 82Þ ¼ 4:6,
Previous sections revealed that NE-speaking p < 0:01. Tukey’s tests revealed that the Early-low groups’
vowels had significantly smaller B1 –B0 ending values than the
listeners judged the /ei /s produced by some Italian–
NE and Late-high groups’ vowels (p < 0:05). Vowels spoken by
English bilinguals––especially late bilinguals––to the Early-low and Early-high groups had significantly larger B2 –
be inaccurate. Acoustic analyses suggested that the B1 values than the Late-high groups’ vowels (p < 0:05). No
bilinguals’ inaccurate /ei / tokens were often pro- other between-group differences reached significance.
cely longer than the late bilinguals’ vowels. A

Tukey’s test using duration as a co-variate again
revealed that the Early-low groups’ distance values
to be significantly greater than the NE, Late-Low,
and Late-high groups’ values (p < 0:01).
A second possible objection to the interpreta-
tion offered earlier is that the analysis included
vowel tokens that were not identifiable either as
English /ei / or Italian /e/. Perhaps the participants
in certain groups produced vowels with less
movement than participants in the Early-low
group did because they were attempting to pro-
duce some vowel other than /ei / or its Italian
counterpart /e/. A second analysis was, therefore,
Fig. 5. The mean beginning values (crosses) and ending values carried out which excluded any token that was not
(various symbols) for English /ei /s produced in two conditions
an instance of the /ei / category, as judged by an
by five groups of participants.
author who is a NE speaker, or an instance of the
Italian /e/ category, as judged by an author who is
amount of movement exhibited by each token was a native Italian speaker.
estimated by computing the Euclidean distance The beginning and ending values for this subset
between its beginning and ending values. The dis- of tokens are shown in Fig. 6. The effect of group
tance scores obtained for the five groups (NE 1.53, (NE mean ¼ 1:52, 35 tokens; Early-low mean ¼
Early-low 2.14, Early-high 1.81, Late-low 1.53, 2:23, 33 tokens; Early-high mean ¼ 1:85, 34 to-
Late-high 1.33) were found to differ significantly, kens; Late-low mean ¼ 1:75, 30 tokens; Late-high
F ð4; 175Þ ¼ 5:9, p < 0:01. A Tukey’s post-hoc test mean ¼ 1:34, 22 tokens) was significant in an
revealed that the Early-low groups’ values were ANOVA examining distance values, F ð4; 149Þ ¼
significantly larger than those obtained for the NE, 5:7, p < 0:01. A post-hoc test indicated that vowels
Late-low and Late-high groups (p < 0:05). No spoken by the Early-low group had larger distance
other between-group differences reached signifi- values than the vowels spoken by participants in
cance.
One possible interpretation of the between-
group difference just reported is that more partic-
ipants in the Early-low group than in the two late
bilingual groups established a category for English
/ei /, and this new category dissimilated from a pre-
existing Italian /e/ category in order to preserve
phonetic contrast (Flege, 1995, 2002). Several po-
tential objections might be raised concerning this
interpretation, however. Participants in the Early-
low group might have produced /ei / with more
movement than the late bilinguals did because
their /ei /s were longer. Variation in vowel duration
was probably not responsible for the observed
differences, however. The vowels produced by the
five groups (NE 284 ms, Early-low 282, Early-high
Fig. 6. The mean beginning values (crosses) and ending values
296, Late-low 280, Late-high 260) did differ sig- (various symbols) for the subset of English /ei / tokens produced
nificantly in duration, F ð4; 175Þ ¼ 3:2, p < 0:05, by the participants in five groups that were judged to be in-
but the Early-low participants’ vowels were scar- stances of either /ei / or /e/ (see text).
the NE and Late-high groups (p < 0:05). These Table 5

results suggest that the difference between the The number of /ei / tokens (maximum 36) produced with dis-
tance values that were at least 1 SD greater than the mean
Early-low group and other groups cannot be at- distance values observed for the NE group (overshoot) or 1 SD
tributed to the inclusion of non-target vowels. less than the NE group’s mean (undershoot)
Still another potential objection to consider is
Undershoot Overshoot
that the Late-low groups’ beginning and ending
NE 4 6
values (and, by extension, amount of tongue
Early-low 3 21
movement) were virtually identical to the values Early-high 4 12
obtained for the NE speakers, at least when all Late-low 12 11
tokens were considered (see Fig. 5). The theoretical Late-high 14 9
expectation generated by the SLM is that expe- Note: The frequencies marked by an asterisk exceed the fre-
rienced Italian–English bilinguals will either pro- quency observed for the NE group at the 0.05 level by a Chi-
duce /ei / with less movement than NE speakers (if square test (p < 0:05).
they continue to identify English /ei /s as instances
of the Italian /e/ category) or they will produce /ei / bution of /ei / movement values obtained for the
with more movement than NE speakers (if they Late-low and NE groups differed. Values obtained
establish a category for /ei / which dissimilates from for the Late-low participants in both dimensions
Italian /e/). Fig. 7(a) shows the distribution of tended to be larger or smaller than the values
differences between the beginning and ending high- obtained for the NE speakers, which clustered in a
low (B1 –B0 ) values. (The values for the Early-low narrow region. The fact that the Late-low partic-
group are also shown for the purpose of compar- ipants’ average values were virtually identical to
ison). Fig. 7(b) shows the distribution of values those of the NE speakers in Fig. 5 is, therefore, a
obtained for movement in the front–back (B2 –B1 ) statistical artifact.
dimension. As expected theoretically, the distri- Another question of interest was how many
tokens the bilinguals produced with more or less
movement than was typical for the NE speakers.
To address this question, each /ei / token having a
distance value exceeding the NE speakers’ mean
by more than 1.0 standard deviation (SD) was
identified. These were designated the ‘‘overshoot’’
tokens. The /ei / tokens having distance values that
were more than 1.0 SD smaller than the NE
speakers’ mean distance value were also identified.
These were designated the ‘‘undershoot’’ tokens.
The numbers of overshoot and undershoot tokens
produced by the five groups of participants are
shown in Table 5. Of the four bilingual groups,
only the Early-low group produced more over-
shoot tokens than the NE group did, X ð1Þ ¼ 7:53,
p < 0:01. Conversely, both groups of late biling-
uals produced more undershoot tokens then the
NE group did, Late-low X ð1Þ ¼ 4:0, Late-high
X ð1Þ ¼ 5:55, p < 0:05.
4.5. Multiple regression analyses

Fig. 7. The distribution of change values observed in two di-
mensions for /ei / tokens produced by three groups of partici-
pants: (a) movement in the high–low (B1 –B0 ) dimension, (b) The data just presented suggest the following
movement the front–back (B2 –B1 ) dimension. interpretation. Participants in the Early-low group
may have produced a large number of /ei / tokens can be detected by listeners. The possibility existed,
with exaggerated movement (the ‘‘overshoot’’ to- however, that the amount of variance accounted
kens) because many of them established a new for by the regression model was artificially aug-
category for English /ei / that dissimilated from mented by the inclusion of tokens not identifiable
Italian /e/. This is consistent with the finding that, as /ei / or /e/. A second regression analysis was,
of the four groups examined by Flege et al. therefore, carried out. It examined /ei /s produced
(1999a), only participants in an Early-low group by the 36 early bilinguals, whose /ei /s were nearly
were able to discriminate English /ei / and Italian /e/ always heard as intended by the NE listeners. This
tokens at a significantly above-chance rate. Par- analysis accounted for 71.2% of the variance,
ticipants in the two late bilingual groups, on the F ð2; 69Þ ¼ 85:1, p < 0:01. Movement in the front–
other hand, may have produced English /ei / with back dimension accounted for 62.2% of the vari-
less movement than the NE speakers did because ance at Step 1, and the front–back (B2 –B1 ) values
many of them continued to treat /ei / tokens as at the beginning of the vowel accounted for an
instances of the Italian /e/ category, and so merged additional 8.9% of the variance at Step 2. These
the properties of English /ei / and Italian /e/. results also suggested that the NE listeners’ judg-
The interpretation just offered assumes that ments of /ei / depended importantly on the amount
listeners can perceptually distinguish variations in of movement.
amount of movement in /ei /-quality and /e/-quality One further objection might be raised regarding
vowels. This assumption agrees with the findings the conclusion that the listeners’ ratings were in-
of previous research (e.g. Nearey, 1989; Strange, fluenced by amount of movement. It is that several
1989; Strange and Bohn, 1998; Hillenbrand et al., predictor variables were correlated. Movement in
1995) showing that listeners make use of spectral the front–back dimension was modestly correlated
information distributed over the entire vowel when with duration, rð70Þ ¼ 0:23, p ¼ 0:052, and the
perceiving vowels. The analyses presented here beginning high–low values, rð70Þ ¼ 0:31, p < 0:01.
evaluated the role of movement on listeners’ rat- Movement in the high–low dimension was corre-
ings of English /ei / by regressing acoustic measure lated with the beginning high–low values, rð70Þ ¼
of /ei / onto the mean ratings obtained from NE- 0:39, p < 0:01. Accordingly, a hierarchical regres-
speaking listeners in Section 3. sion analysis examining the early bilinguals’ pro-
Five acoustic measures were regressed onto the ductions of /ei / was carried out. The beginning
listener ratings of /ei / tokens produced by the 72 high–low values, the beginning front–back values,
Italian–English bilinguals. The predictor variables and vowel duration were entered at Step 1. The
were the front–back (B2 –B1 ) and high–low (B1 –B0 ) acoustic measures of movement in the high–low
values obtained at the beginning location (i.e. the and front–back dimensions were entered at Step 2.
vowel onset), amount of movement in the front– Duration accounted for 10.1% of the variance at
back and high–low dimensions (quantified as the Step 1. Movement in the front–back dimension
amount of B2 –B1 and B1 –B0 change from the be- accounted for an additional 54.1% of the variance
ginning to the end of the vowel), and vowel du- at Step 2, F ð2; 69Þ ¼ 61:9, p < 0:01, independently
ration. The model accounted for 74.6% of the of the confounded variables. This confirmed that
variance, F ð4; 139Þ ¼ 102:1, p < 0:01. Movement variation in tongue movement exerted an influence
in the front–back dimension accounted for 56.2% on the listeners’ ratings of /ei /.
of the variance at Step 1. Front–back (B2 –B1 )
values at the beginning of the vowel accounted for
additional 13.7% of the variance at Step 2, dura- 5. General discussion
tion for 3.6% at Step 3, and movement in the high–
low (B1 –B0 ) dimension for 1.1% of the variance at This study examined Italian–English bilinguals’
Step 4. production of 11 English vowels. The bilinguals
The results of this analysis suggested that were selected on the basis of when in life they ar-
tongue movement differences in productions of /ei / rived in Canada and how much they continued to
use Italian (four groups in all). Both AOA and L1 bilinguals produced more undershoot tokens than
use were found to influence the bilinguals’ pro- the NE group did.
duction of English vowels. Parametric and non- These findings can be interpreted within the
parametric analyses of listeners’ ratings indicated framework of the SLM. Participants in the two
that early bilinguals tended to produce the English late bilingual groups may have tended to produce
vowels more accurately than late bilinguals did English /ei / with less movement than the NE
and low-L1-use bilinguals tended to produce En- speakers did because they continued to treat /ei /
glish vowels more accurately than high-L1-use tokens as instances of the Italian /e/ category. That
bilinguals. These findings agreed with previous is, they may have merged the properties of English
research examining the effect of age of L2 learning /ei / and Italian /e/ as the result of the mechanism of
(Flege, 1992; Yamada, 1995; Munro et al., 1996; category assimilation. As discussed in the Intro-
Flege et al., 1995a,b; Flege et al., 1999a; Meador duction, the effect of category assimilation has
et al., 2000; Yeni-Komshian et al., 2000; MacKay been observed in previous studies examining the
et al., 2001; MacKay et al., 2001; Piske et al., 2001, production of L2 stop consonants (Flege, 1987;
2002) and amount of continued L1 use (Flege et al., MacKay et al., 2001).
1997; Guion et al., 2000; Meador et al., 2000; Participants in the Early-low group, on the
MacKay et al., 2001; Piske et al., 2001, 2002). other hand, may have produced a large number of
The primary purpose of the study, however, /ei / tokens with exaggerated movement (overshoot)
was to examine the production of English /ei / in because many of them established a new category
detail. A language interaction model, the SLM for English /ei / that dissimilated from Italian /e/.
(e.g. Flege, 1995, 1999, 2002) generated two pre- This is consistent with the finding that, of the four
dictions that were tested here. The first prediction groups examined by Flege et al. (1999a), only
was that participants in the Early-low group would participants in an Early-low group discriminated
produce English /ei / with a more movement than English /ei / and Italian /e/ tokens at a significantly
participants in the two late bilinguals groups above-chance level. Category dissimilation has
(Late-low, Late-high) would. The second predic- been observed in previous research in which early
tion was that the late bilinguals would produce bilinguals established a new category for an L2
English /ei / with less movement than the NE stop consonant (Flege and Eefting, 1987, 1988).
speakers. Both hypotheses were supported. The The SLM posits that an L2 phonetic category may
amount of movement in each /ei / token was esti- dissimilate from a neighboring L1 vowel category
mated by computing the distance between the in order to preserve phonetic contrast among the
beginning and ending values in a 2-dimensional elements of the L1 and L2 subsystems, which are
formant space. The distance values were signifi- said to exist in a common phonological space.
cantly larger for vowels spoken by the Early-low That is, Italian–English bilinguals may have pro-
group than by the NE, Late-low and Late-high duced English /ei / with more movement than is
groups. A second analysis restricted to vowel to- typical for English in order to make it distinct
kens identifiable as /ei / or /e/ indicated that the from their Italian /e/.
Early-low group produced /ei / with significantly The inference that more participants in the
more movement than participants in the NE and Early-low group than in either of the two late bi-
Late-high groups did. Another analysis focused on lingual groups established a category for English
the number of individuals who produced /ei / with /ei / is consistent with the findings of a study by
distance values that were at least 1 SD greater than Flege and MacKay (Submitted). These authors
the mean value obtained for the NE group (des- examined the categorial discrimination of nine
ignated ‘‘overshoot’’ tokens) or at least 1 SD less pairs of English vowels by the same four groups of
than the NE mean (‘‘undershoot’’ tokens). Of the Italian–English bilinguals who participated in this
four native Italian groups, only the Early-low study. Both AOA and L1 use affected the biling-
group produced more overshoot tokens than the uals’ discrimination of English vowels. Partici-
NE group did; conversely, both groups of late pants in the Early-low group obtained the highest
discrimination scores whereas participants in the nation task (e.g. Flege et al., 1999a). They should
Late-high obtained the lowest scores. Bilingual also prefer /ei / tokens with much movement to /ei /
participants were credited with native-like per- tokens with little movement in a goodness rating
ception if they obtained a score that fell within 2 task. Conversely, bilinguals who show movement
SDs of the NE speakers’ mean. Significantly more undershoot in producing /ei / should either not
early than late bilinguals were so credited; and show these effects or show them to a significantly
there was a non-significant trend for more low-use lesser extent than the bilinguals who show over-
than high-use participants to be so credited. shoot in production. This is because, by hypothe-
At least one alternative account can be offered sis, the bilinguals who show undershoot do so
for the effects observed here. It could be hypoth- because they continue to treat English /ei / tokens
esized that participants in the Early-low group as instances of Italian /e/ and, as a result, fail to
were more likely than those in the other three establish a new category for English /ei /.
bilingual groups to identify strongly with Cana- In summary, the results obtained here sup-
dian culture and/or the English language. If this ported the prediction that certain Italian–English
were so, more Early-low participants may have bilinguals would produce /ei / with too little tongue
wanted to ‘‘sound Canadian’’ for affective or movement whereas other would produce this vo-
socio-cultural reasons. This might explain why wel with too much movement. The findings were
participants in the Early-low group used Italian taken as support for the hypothesis that the L1
less than those in the Early-high group, and why and L2 phonetic subsystems of bilinguals interact
they tended to produce English /ei / with exagger- through two distinct mechanisms, phonetic cate-
ated movement. That is, the Early-low partici- gory assimilation and phonetic category dissimi-
pants’ productions of English /ei / with exaggerated lation. If additional research supports this view of
formant movement might be seen as evidence of L1–L2 interactions, it will be necessary to deter-
‘‘hypercorrection’’. mine how much of the native versus non-native
Additional research examining Italian–English differences observed in L2 speech research can be
bilinguals’ production of Italian vowels might be attributed to language interaction effects.
useful in helping to choose between the hyper-
correction account and a language interaction
account. If research reveals that only bilinguals
showing undershoot in English /ei / produce Ital- Acknowledgements
ian /e/ with more movement in that vowel than
Italian monolinguals do, it would support the This study was supported by grant DC00257
view that the overshoot in English /ei / seen in this from the National Institute for Deafness and
study arose from the establishment of an /ei / Other Communicative Disorders. The authors
category. thank St. Anthony’s parish in Ottawa, all of the
Additional research will also be needed to participants, and K. Aoyama, A. Højen, S. Imai,
evaluate bilinguals’ perception of English /ei /. If an M. Mack, K. Tsukada and an anonymous re-
overshoot of movement in /ei / is the result of cat- viewer for comments on earlier versions of the
egory dissimilation, whereas undershoot is the re- article. Thanks are also extended to A. Zamboni,
sult of category assimilation by individuals who E. Magno Caldognetto and F. Ferrero of the Is-
fail to establish a category for /ei /, then different tituto di Fonetica e Dialettologia, Padua, Italy, for
perceptual outcomes should be observed for sub- assistance in recording Italian vowels.
groups of experienced Italian–English bilinguals.
Specifically, bilinguals showing overshoot in /ei /
should be able to differentially identify English /ei / Appendix A
and Italian /e/ tokens in a language identification
task, and to discriminate English /ei / and /e/ tokens The mean ratings obtained for vowels spoken
at an above-chance rate in a categorial discrimi- by five groups in two elicitation conditions
EC Vowel
/Z/ /d/ /ei / /o/ /u/ // /e/ /I / // / / /i/
Native 1 3.5 3.4 3.2 3.5 3.4 3.3 3.4 3.3 3.3 3.4 3.3
English (0.3) (0.2) (0.6) (0.3) (0.2) (0.4) (0.2) (0.3) (0.5) (0.2) (0.2)
3 3.3 3.4 3.5 3.6 3.5 3.5 3.5 3.4 3.3 3.5 3.3
(0.6) (0.3) (0.2) (0.3) (0.1) (0.1) (0.2) (0.2) (0.4) (0.3) (0.5)
Early- 1 3.5 3.4 3.4 3.5 3.3 3.4 3.3 3.4 3.5 3.2 3.4
low (0.2) (0.2) (0.6) (0.3) (0.3) (0.2) (0.4) (0.3) (0.2) (0.6) (0.2)
3 3.4 3.3 3.1 3.4 3.5 3.4 3.2 3.3 3.3 3.1 3.4
(0.4) (0.3) (0.8) (0.3) (0.2) (0.2) (0.2) (0.3) (0.3) (0.8) (0.2)
Early- 1 3.4 3.3 3.3 3.4 3.3 3.2 3.0 3.2 3.3 3.1 3.2
high (0.3) (0.4) (0.5) (0.3) (0.3) (0.5) (0.7) (0.5) (0.4) (0.6) (0.6)
3 3.3 3.3 3.2 3.3 3.3 3.2 2.9 3.2 3.3 2.4 3.0
(0.3) (0.3) (0.6) (0.3) (0.2) (0.6) (0.7) (0.3) (0.3) (0.9) (0.8)
Late- 1 3.2 3.0 3.1 3.2 3.0 3.1 2.4 2.9 3.1 2.5 3.2
low (0.3) (0.2) (0.7) (0.3) (0.3) (0.4) (0.6) (0.5) (0.6) (0.6) (0.4)
3 2.9 3.0 2.8 3.0 2.9 2.8 2.6 2.9 2.7 2.2 2.9
(0.5) (0.4) (0.9) (0.5) (0.4) (0.4) (0.6) (0.5) (0.5) (0.8) (0.8)
Late- 1 2.9 2.7 2.2 2.8 2.6 2.9 2.5 2.8 2.9 2.5 2.7
High (0.7) (0.5) (0.9) (0.7) (0.5) (0.4) (0.8) (0.5) (0.6) (0.7) (0.7)
3 2.5 2.4 2.2 2.5 2.8 2.8 2.6 2.3 2.8 1.7 2.5
(0.9) (0.6) (0.9) (0.7) (0.5) (0.5) (0.8) (0.8) (0.5) (0.4) (0.9)
Note: EC, elicitation condition (1-word versus 3-word). Standard deviations are in parentheses.
References Flege, J.E., 1989. Differences in inventory size affects the

location but not the precision of tongue positioning in vowel
Bahrick, H., Hall, L., Goggin, J., Bahrick, L., Berger, S., 1994. production. Lang. Speech 32, 123–147.
Fifty years of language maintainance in bilingual Hispanic Flege, J.E., 1991. Age of learning affects the authenticity of
immigrants. J. Exp. Psych. Gen. 123, 264–283. voice onset time (VOT) in stop consonants produced in a
Baker, W., Trofimovich, P., Mack, M., Flege, J.E., 2002. The second language. J. Acoust. Soc. Am. 89, 395–411.
effect of perceived phonetic similarity on non-native sound Flege, J.E., 1992. Speech learning in a second language. In:
learning by children and adults. In: Do, A., Domınguez, L., Ferguson, C., Menn, L., Stoel-Gammon, C. (Eds.), Phono-
Johansen, A. (Eds.), BUCLD 26: Proceedings of the 26th logical Development: Models, Research, and Implications.
Annual Boston University Conference on Language Devel- York Press, Timonium, MD, pp. 565–604.
opment. Cascadilla Press, Somerville, MA. Flege, J.E., 1995. Second-language speech learning: theory,
Bohn, O.-S., Flege, J.E., 1992. The production of new and findings, and problems. In: Strange, W. (Ed.), Speech
similar vowels by adult German learners of English. Stud. Perception and Linguistic Experience, Issues in Cross-
Second Lang. Acquis. 14, 131–158. linguistic research. York Press, Timonium, MD, pp. 233–
Conover, W., 1980. Practical Nonparametric Statistics, second 277.
ed. John Wiley & Sons, New York. Flege, J.E., 1998. The role of subject and phonetic variables in
Delattre, P., 1964. Comparing the vocalic features of English, L2 speech acquisition. In: Gruber, M., Higgins, D., Olsen,
German, Spanish and French. Int. Rev. Appl. Ling. 2, 71– K., Wysocki, T. (Eds.), Papers from the 34th Annual
97. Meeting of the Chicago Linguistic Society, Vol. II, The
Flege, J.E., 1987. The production of ‘new’ and ‘similar’ phones Panels. Chicago, Chicago Linguistic Society, pp. 213–232.
in a foreign language: evidence for the effect of equivalence Flege, J.E., 1999. Age of learning and second-language speech.
classification. J. Phonetics 15, 47–65. In: Birdsong, D.P. (Ed.), Second Language Acquisition and
the Critical Period Hypothesis. Lawrence Erlbaum, Hills- Guion, S.G., Flege, J.E., Loftin, J.D., 2000. The effect of L1 use
dale, NJ, pp. 101–132. on pronunciation in Quichua-Spanish bilinguals. J. Phonet-
Flege, J.E., 2002. Interactions between the native and second- ics 28, 27–42.
language phonetic systems. In: Burmeister, P., Piske, T., Halle, P., Best, C., Levitt, A., 1999. Phonetic versus phonolog-
Rohde, A. (Eds.), An Integrated View of Language Devel- ical influences on French listeners’ perception of American
opment: Papers in Honor of Henning Wode. Wissenschaf- English approximants. J. Phonetics 27, 281–306.
tlicher Verlag Trier, Trier. Hazan, V., Barrett, S., 1999. The development of phoneme
Flege, J.E., Eefting, W., 1987. The production and perception categorisation in children aged 6 to 12 years. J. Phonetics
of English stops by Spanish speakers of English. J. 28, 377–396.
Phonetics 15, 67–83. Hillenbrand, J., Getty, L., Clark, M., Wheeler, K., 1995.
Flege, J.E., Eefting, W., 1988. Imitation of a VOT continuum Acoustic characteristics of American English vowels. J.
by native speakers of English and Spanish: evidence for Acoust. Soc. Am. 97, 3099–3111.
phonetic category formation. J. Acoust. Soc. Am. 83, 729– Jia, G., Aaronson, D., 1999. Age differences in second language
740. acquisition the dominant language switch and maintainance
Flege, J.E., Fletcher, S., McCutcheon, M., Smith, S., 1986. The hypothesis. In: Greenhill, A., Littlefield, H., Tano, C. (Eds.),
physiological specification of American English vowels. Proceedings of the 23rd Annual Boston University Confer-
Lang. Speech 29, 361–388. ence on Language Development. Cascadilla Press, Somer-
Flege, J.E., Frieda, E.M., Nozawa, T., 1997. Amount of native- ville, MA, pp. 301–312.
language (L1) use affects the pronunciation of an L2. J. Johnson, C., 2000. Children’s phoneme identification in rever-
Phonetics 25, 169–186. beration and noise. J. Speech Lang. Hear. Res. 43, 129–143.
Flege, J.E., Liu, S., 2001. The effect of experience on adults’ Kluender, K., Lotto, A., Holt, L., Bloedel, S., 1998. Role of
acquisition of a second language. Stud. Second Lang. experience for language-specific functional mappings of
Acquis. 23, 527–552. vowel sounds. J. Acoust. Soc. Am. 104, 3568–3582.
Flege, J.E., MacKay, I.R.A., Meador, D., 1999a. Native Italian Lambert, W.E., Rawlings, C., 1969. Bilingual processing of
speakers’ production and perception of English vowels. J. mixed-language associative networks. J. Verbal Learning
Acoust. Soc. Am. 106, 2973–2987. Verb. Behav. 8, 604–609.
Flege, J.E., MacKay, I.R.A. Constraints on the perception of Lee, S., Potamianos, A., Narayanan, S., 1999. Acoustics of
vowels in a second language, submitted. children’s speech: developmental changes of temporal and
Flege, J.E., Munro, M.J., MacKay, I.R.A., 1995a. Factors spectral parameters. J. Acoust. Soc. Am. 105, 1455–1468.
affecting degree of perceived foreign accent in a second Lindblom, B., 1998. Systemic constraints and adaptive change
language. J. Acoust. Soc. Am. 97, 3125–3134. in the formation of sound structures. In: Hurford, J.,
Flege, J.E., Munro, M.J., MacKay, I.R.A., 1995b. The effect of Stuttert-Kennedy, M., Knight, C. (Eds.), Approaches to the
age of second language learning on the production of Evolution of Language: Social and Cognitive Bases. Cam-
English consonants. Speech Commun. 16, 1–26. bridge University Press, Cambridge.
Flege, J.E., Port, R., 1981. Cross-language phonetic interfer- Mack, M., 1989. Consonant and vowel perception and
ence: Arabic to English. Lang. Speech 24, 125–146. production: early English–French bilinguals and English
Flege, J.E., Yeni-Komshian, G.H., Liu, H., 1999b. Age con- monolinguals. Percep. Psychophys. 46, 187–200.
straints on second language acquisition. J. Mem. Lang. 41, Mack, M., 1990. Phonetic transfer in a French–English
78–104. bilingual child. In: Nelde, P.H. (Ed.), Language Attitudes
Gardner, R.C., Lambert, W.E., 1972. Attitudes and Motivation and Language Conflict. D€ ummler, Bonn, Germany.
in Second-language Learning. Newbury House, Rowley, MA. Mack, M., 2003. The phonetic systems of bilinguals. In: Banich,
Gay, T., 1968. Effect of speaking rate on diphthong formant M.T., Mack, M. (Eds.), Mind, Brain, and Language:
movements. J. Acoust. Soc. Am. 44, 1570–1573. Multidisciplinary Perspectives. Lawrence Erlbaum Press,
Grenier, G., 1984. Shifts to English as usual language by Mahwah, NJ, in press.
Americans of Spanish mother tongue. Social Sci. Quart. 65, Mack, M., Bott, S., Boronat, C., 1995. Mother, I’d rather do it
537–550. myself, maybe: an analysis of voice-onset time produced by
Grosjean, F., 1982. Life with Two Languages. Harvard early French–English bilinguals. Issues Develop. English
University Press, Cambridge, MA. Appl. Ling. 8, 23–55.
Grosjean, F., 1989. Neurolinguists, beware! The bilingual is not MacKay, I.R.A, Meador, D., Flege, J.E., 2001. The identifica-
two monolinguals in one person. Brain Lang. 36, 3–15. tion of English consonants by native speakers of Italian.
Grosjean, F., 1997. Processing mixed language: issues, findings Phonetica 58, 103–125.
and models. In: de Groot, A., Kroll, J. (Eds.), Tutorials in MacKay, I.R.A., Flege, J.E., Piske, T., Schirru, C., 2001.
Bilingualism: Psycholinguistic Perspectives. Lawrence Erl- Category restructuring during second-language (L2) speech
baum, Mahwah, NJ, pp. 225–254. acquisition. J. Acoust. Soc. Am. 110, 516–528.
Grosjean, F., 1999. Studying bilinguals: Methodological and Mayo, L., Florentine, M., Buus, S., 1997. Age of second-
conceptual issues. Bilingualism: Lang. Cognition 1, 117– language acquisition and perception of speech in noise. J.
130. Speech Hear. Res. 40, 686–693.
Meador, D., Flege, J.E., MacKay, I.R.A., 2000. Factors Scovel, T., 1988. A Time to Speak. A Psycholinguistic Inquiry
affecting the recognition of words in a second language. into the Critical Period for Human Speech. Newbury House
Bilingualism: Lang. Cognition 3, 55–67. Publishers, Cambridge, MA.
Munro, M., 1993. Productions of English vowels by native Sebastian-Galles, N., Soto-Faraco, S., 1999. On-line processing
speakers of Arabic: acoustic measurements and accented- of native and non-native phonemic contrasts in early
ness ratings. Lang. Speech 36, 39–66. bilinguals. Cognition 72, 111–123.
Munro, M.J., Flege, J.E., MacKay, I.R.A., 1996. The effects of Stevens, G., 1999. Age at immigration and second language
age of second-language learning on the production of proficiency among foreign-born adults. Lang. Soc. 28, 555–
English vowels. Appl. Psycholing. 17, 313–334. 578.
Nearey, T., 1989. Static, dynamic, and relational properties in Strange, W., 1989. Dynamic specification of coarticulated
vowel perception. J. Acoust. Soc. Am. 85, 2088–2113. vowels in sentence contexts. J. Acoust. Soc. Am. 85, 2135–
Paradis, M., 1978. The stratification of bilingualism. In: 2153.
Paradis, M. (Ed.), Aspects of Bilingualism. Hornbeam Strange, W., Bohn, O.-S., 1998. Dynamic specification of
Press, Columbia, SC, pp. 165–176. coarticulated German vowels: perceptual and acoustical
Parnell, M., Amerman, J., 1978. Maturational influences on the studies. J. Acoust. Soc. Am. 104, 488–504.
perception of coarticulatory effects. J. Speech Hear. Res. 21, Syrdal, A., Gopal, H., 1986. A perceptual model of vowel
682–701. recognition based on the auditory representation of Amer-
Patkowski, M., 1989. Age and accent in a second language: a ican English vowels. J. Acoust. Soc. Am. 79, 1086–1100.
reply to James Emil Flege. Appl. Ling. 11, 73–89. Walley, A.C., Flege, J.E., 2000. Effects of lexical status on
Peterson, G., Barney, H., 1952. Control methods used in a children’s and adults’ perception of native and non-native
study of the vowels. J. Acoust. Soc. Am. 24, 175–184. vowels. J. Phonetics 27, 307–332.
Piske, T., Flege, J.E., MacKay, I.R.A., Meador, D., 2002. The Yamada, R., 1995. Age and acquisition of second language
production of English vowels by fluent early and late speech sounds: perception of American English / / and /l/ by
Italian–English bilinguals. Phonetica 59, 49–71. native speakers of Japanese. In: Strange, W. (Ed.), Speech
Piske, T., MacKay, I.R.A., Flege, J.E., 2001. Factors affecting Perception and Linguistic Experience, Issues in Cross-
degree of foreign accent in an L2: a review. J. Phonetics 29, language Research. York Press, Timonium, MD, pp. 305–
191–215. 320.
Sancier, M., Fowler, C., 1997. Gestural drift in a bilingual Yeni-Komshian, G.H., Flege, J.E., Liu, S., 2000. Pronunciation
speaker of Brazilian Portuguese and English. J. Phonetics proficiency in the first and second languages of Korean–
25, 421–437. English bilinguals. Bilingualism: Lang. Cognition 3, 131–150.
View publication stats

Flege Schirru Interaction Between SC 2003

Uploaded by

Copyright:

Available Formats

Flege Schirru Interaction Between SC 2003

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Flege Schirru Interaction Between SC 2003

Uploaded by

Copyright:

Available Formats

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

Interaction between the native and second language phonetic

Article in Speech Communication · June 2003

James Emil Flege Carlo Schirru

SEE PROFILE SEE PROFILE

The user has requested enhancement of the downloaded file.

Interaction between the native and second language

1. Introduction one) the age of ﬁrst exposure to the L2 has been

Late-high 8m 49(8) 20(3) 53%(13) 29(9) 4.4(1.6) 2(2)

M 49(6) 14(7) 28%(23) 36(9) 3.7(1.8) 7(6)

residence (LOR) in Canada was confounded with Table 2

from the NE group. An alpha level of 0.05 (unad-

(Syrdal and Gopal, 1986), then plotted in a two-

eight females) had a mean age of 26 years and were

duced with less movement than the bilinguals’ ac-

cely longer than the late bilinguals’ vowels. A

the NE and Late-high groups (p < 0:05). These Table 5

4.5. Multiple regression analyses

References Flege, J.E., 1989. Diﬀerences in inventory size aﬀects the

View publication stats

You might also like