Content uploaded by Charles B. Chang
Author content
All content in this area was uploaded by Charles B. Chang on Aug 30, 2023
Content may be subject to copyright.
INDIVIDUAL DIFFERENCES IN VOWEL COMPACTNESS PERSIST
UNDER INTOXICATION ACROSS FIRST AND SECOND LANGUAGES
Charles B. Chang, Kevin Tang, & Andrew Nevins*
Boston Univ.; Heinrich-Heine-Universität Düsseldorf & Univ. of Florida; Univ. College London
cbchang@post.harvard.edu, Kevin.Tang@hhu.de, a.nevins@ucl.ac.uk
ABSTRACT
Alcohol intoxication facilitates inhibition of one’s
first language (L1) ego, which may lead to reduced
individual differences among second language (L2)
speakers under intoxication. This study examined
whether, compared to speaking while sober, speak-
ing while intoxicated would reduce individual dif-
ferences in the acoustic compactness of vowel cate-
gories in sequential bilinguals exemplifying diverse
L1–L2 pairs (German–English, Korean–English).
Vowel compactness in F1×F2space varied by lan-
guage (German, Korean, English) and by vowel,
and was generally lower in intoxicated compared to
sober speech, both across languages and throughout
a bilingual’s language repertoire. Crucially, how-
ever, there was still a wide range in compactness un-
der intoxication; furthermore, individuals with more
compact vowels while sober also produced more
compact vowels while intoxicated, in both L1 and
L2. Taken together, these findings show patterned
variability of vowel compactness, suggesting that
articulatory precision is an individual-difference di-
mension that persists across speaking conditions and
throughout the repertoire.
Keywords: compactness, individual differences,
vowels, bilingualism, alcohol.
1. INTRODUCTION
How does drinking affect one’s speech? A grow-
ing literature on alcohol intoxication and spoken lan-
guage production suggests that there may not be a
simple answer to this question. On the one hand,
several studies have found, consistent with the gen-
eral decline of motor control under intoxication, that
intoxication degrades speech, in terms of more er-
rors, less precise gestural coordination, more conso-
nant lenition, slower speech rate, greater pitch vari-
*We gratefully acknowledge the data contributions and re-
search assistance of Kai Xin Bao, Marissa Carl, Erik Duch-
nowski, Jenny Geng, Sam Green, Michael Hindley, Young Shin
Kim, Peyton Krinsky, Ji Hye Kwon, Wen Jia Liu, Eliana Mugar,
Yin Wang, Yifan Wu, and Steven Zhang, as well as the helpful
feedback of two anonymous reviewers.
ability, and/or lower intelligibility [1, 2, 3, 4, 5, 6,
7, 8]. On the other hand, studies that have exam-
ined intoxication effects on speech in a bilingual’s
second language (L2), in isolation or in comparison
to speech in their first language (L1), have often ob-
served that intoxication affects L2/nonnative speech
differently from L1 speech [7, 9]. For example, in-
toxication improved L1 German speakers’ pronunci-
ation in Dutch [10], which can be explained in terms
of intoxication making a speaker’s “language ego”
less L1-biased, facilitating nonnative speech [9].
The possibility of facilitative intoxication raises
interesting questions for L2 speech research. For
one, if intoxication improves L2 production, might
it “level the playing field” among L2 learners, who
often show marked individual differences even from
the same L1 background [11, 12, 13, 14, 15]? L2
speech research has been increasingly concerned
with accounting for individual differences, linking
them to factors such as cue-weighting preferences,
memory variation, and perceptual and neural differ-
ences [16, 17, 18, 19, 20]. In this study, we focused
on individual differences in compactness, the acous-
tic consistency of a phonetic category’s production
[21], thought to reflect the “consistency of [speak-
ers’] phonological-motor mapping” [22, p. 826].
Part of a larger project examining intoxication ef-
fects on bilingual speech, the current study asked
how individual differences among bilinguals in
vowel compactness would be affected by intoxica-
tion, across L1 and L2. Based on findings of intox-
ication degrading L1 but facilitating L2 speech, we
hypothesized that intoxication would result in less
compact L1 vowels, yet—by enhancing access to L2
phonological-motor mappings—more compact L2
vowels. Given the persistence of variation in com-
pactness across L1 and L2 [21, 22], intoxication-
induced changes in compactness were not predicted
to reduce individual differences in L1; that is, ha-
bitually “less compact” speakers were predicted to
become even less compact under intoxication along
with “more compact” speakers. In contrast, if intox-
ication does indeed have an L2-specific facilitation
effect (and there is an upper limit on compactness),
it could allow less-compact speakers to “catch up”
3. Speech Production and Speech Physiology ID: 1035
1182
to more-compact speakers in terms of compactness,
leading to reduced individual differences in L2. To
explore these predictions, we compared the sober
and intoxicated speech of bilinguals exemplifying
diverse L1–L2 pairs.
2. METHODS
2.1. Participants
Participants comprised two groups of sequential
bilinguals: Germans (L1 German, L2 English; N=
8; 4f, 4m; Mage =27.1 yr) and Koreans (L1 Ko-
rean, L2 English; N=8; 8f, 0m; Mage =27.1 yr).
The groups were born and raised in Germany or
South Korea, respectively, and were tested in the
UK, where most participants had been living for 1–
2 years for study abroad. They reported no history
of hearing or speech problems, and their average
scores on the IELTS [23] (7.0 band) and LexTALE
[24] (60s), which did not differ significantly between
groups, suggested that their English proficiency was
generally “upper intermediate” (B2) or higher.
2.2. Materials
Speech materials were compiled for each language
on the basis of published dialogues (e.g., from a
play), which were edited for emotional and gender
neutrality, turn length, and archaisms. The German
dialogue contained a total of 55 turns; the Korean di-
alogue 123 turns; and the English dialogue 95 turns.
The full materials with source references are avail-
able open-access at osf.io/y2r87/.
2.3. Procedure
The main task was a dialogue-reading task, which
participants completed in a sound-insulated room
four times (2 languages ×2 speaking conditions)
across two test sessions spaced no more than 14 days
apart. Each dialogue was uttered in two speaking
conditions (sober, intoxicated), completed on sepa-
rate days in the order sober–intoxicated in the Ko-
rean group and in counterbalanced order in the Ger-
man group. Participants were told not to eat, drink,
or use mouthwash starting 2 h before each session
and not to smoke starting 0.5 h before.
Before beginning the dialogue-reading task, par-
ticipants were instructed to read the dialogues nat-
urally. Seated in front of a microphone facing the
experimenter, the participant went through each di-
alogue with the experimenter, each person read-
ing one character’s lines. The session was audio-
recorded at 44.1 kHz with 16-bit resolution in stereo;
the recordings were then converted to mono.
Participants’ BAC was tested in both conditions
using a breathalyzer (AlcoMate Premium AL-7000).
First, BAC was measured to ensure that participants
had no alcohol in their system. In the intoxicated
condition, the target BAC was 0.12%. To reach this
BAC, participants consumed, at their own pace, an
amount of alcohol (vodka or rum, mixed with juice)
determined on the basis of their self-reported weight
and BAC charts [25]. BAC was tested 15 min after
about 75% of the mixture was consumed and then
every 3–5 min until it went over 0.12% and dropped
back down to 0.12%; if needed, a small top-up
was given from the remaining alcohol. Once BAC
had reached 0.12%, participants were taken into the
sound-insulated room for the dialogue-reading task.
2.4. Analysis
Forced-aligned TextGrids (produced using the Mon-
treal Forced Aligner [26]) were hand-corrected and
then submitted, along with the audio recordings, to
acoustic analysis in Praat [27]. For each of the vowel
intervals marked in the TextGrids, the analysis ex-
tracted both the first and second formants (F1,F2)
over the middle 50% of the vowel’s duration using
the default settings of Praat’s linear predictive cod-
ing method; the maximum formant parameter was
set to 5,000 Hz for males and 5,500 Hz for females.
Because German and English, in contrast to Ko-
rean, are described as having unstressed vowel qual-
ity reduction, the final dataset for statistical analy-
sis included only primary-stressed monophthongs in
German and English, as well as the monophthongs
of Korean. The F1and F2measures on this set of
vowels were checked for errors separately by partic-
ipant and by language, using the 3-SD criterion for
outliers; tokens with one or more errors (comprising
1.2% of the dataset) were removed. Formant mea-
sures on the remaining tokens (N=64,382 total;
9,081 in German, 27,450 in Korean, 27,851 in En-
glish) were then Nearey-normalized [28] using the
vowels package [29] in R [30]. The full dataset is
available open-access at osf.io/s9vkw/.
The normalized formant measures were analyzed
both qualitatively and quantitatively. First, we in-
spected vowel plots to observe variation in vowel
compactness among participants and languages and
within the repertoire (L1, L2). Then we calculated
a metric of vowel area (i.e., the inverse of compact-
ness) in terms of the area of the 95% confidence el-
lipse around tokens. Finally, we built a linear mixed-
effects model on the vowel area data, using the
lmerTest package [31]. The model included fixed
effects for Group (treatment-coded; ref = German),
3. Speech Production and Speech Physiology ID: 1035
1183
Condition (treatment-coded; ref = sober), RepLang
(treatment-coded; ref = L1), Vowel (sum-coded; all
monophthongs including German /i I e E a A A: u U o
y Y ø œ/, Korean /i e E u o 2 1/, and English /iIeEæ
uUoAO2Ä/), all interactions among Group, Condi-
tion, and RepLang, a Condition ×Vowel interaction,
as well as random intercepts by Participant.1
3. RESULTS
Vowel area varied significantly by language, with
tokens of a vowel category generally more dif-
fusely (i.e., less compactly) distributed in the Korean
monophthongs than the German monophthongs in
sober speech [β=0.226,t=4.575,p< .001]. On
the other hand, vowel area was smaller in the Ger-
man group’s L2 English monophthongs as compared
to their L1 German [β=−0.122,t=−4.019,p<
.001], and this L2 effect was even stronger in the
Korean group [β=−0.138,t=−3.161,p=.002].2
Examination of individual differences in vowel
area revealed substantial variation in both groups.
Table 1 summarizes the ranges in average vowel area
by group, language, and condition. Although the
minimum and maximum values within each group
are different (consistent with the language effect
above), there is a wide range in every case.
Group Language, Condition Min Max
German German, sober 0.303 0.500
German German, intoxicated 0.316 0.637
German English, sober 0.204 0.487
German English, intoxicated 0.308 0.497
Korean Korean, sober 0.556 0.885
Korean Korean, intoxicated 0.565 0.953
Korean English, sober 0.332 0.483
Korean English, intoxicated 0.308 0.515
Table 1: Range in average vowel area (min, max
values) in Nearey-normalized F1×F2space.
Examples of individual differences are depicted in
Fig. 1–4, which show the vowel spaces for some of
the most-compact (G005, K004) and least-compact
(G006, K006) participants by language and condi-
tion on the same scale, with 95% confidence ellipses
around tokens of the same vowel (plots for all par-
ticipants are at osf.io/s9vkw/). Note the marked
difference in compactness between G005 and G006
and between K004 and K006, as well as the broad
similarity among the four panels in each figure.
As seen in Fig. 1–4, the area for a given vowel
tended to be larger in intoxicated than in sober
speech for all languages. Thus, instead of showing
an L2-specific effect (in the interaction coefficients),
Figure 1: Vowel spaces by language (L1 German,
L2 English) and condition in participant G005.
Figure 2: Vowel spaces by language (L1 German,
L2 English) and condition in participant G006.
the model showed that the marginal intoxication ef-
fect in German [β=0.049,t=1.743,p=.082] did
not differ in Korean [β=0.030,t=0.625,p=.532]
or L2 English [|β|s<0.099, |t|s<1.586, ps> .1].
That is, we found no evidence of L2-specific facili-
tation of vowel production. Rather, the intoxication
effect was general, resulting in expanded vowel ar-
eas at the group mean level in L1 (Germans: Mint =
0.460, Msob =0.391; Koreans: Mint =0.743, Msob =
0.677) and L2 (Germans: Mint =0.395, Msob =
0.316; Koreans: Mint =0.414, Msob =0.404).
In order to further examine the relationship be-
tween vowel areas in intoxicated and sober speech
at the individual level, we correlated participants’
by-language average vowel areas for the intoxi-
cated condition against those for the sober condition.
The two sets of values were highly correlated over-
all [r(30) = 0.926,t=13.448,p< .001], in Ger-
man [r(6) = 0.747,t=2.752,p=.033], in Korean
[r(6) = 0.902,t=5.120,p=.002], and in the Ger-
man group’s English [r(6) = 0.845,t=3.866,p=
.008], although not in the Korean group’s English
[r(6) = 0.480,t=1.342,p=.228]. Thus, in gen-
eral, bilinguals with more compact vowels while
3. Speech Production and Speech Physiology ID: 1035
1184
Figure 3: Vowel spaces by language (L1 Korean,
L2 English) and condition in participant K004.
Figure 4: Vowel spaces by language (L1 Korean,
L2 English) and condition in participant K006.
sober also produced more compact vowels while in-
toxicated, whether speaking in L1 or L2.
Finally, there were significant differences among
vowels in terms of area, but not in terms of intoxi-
cation effects. Compared to the average vowel area
over all vowels, the vowels /i I u U/ (in German and
English), /y A:/ (in German), /1/ (in Korean), and
/A/ (in English) were significantly more diffusely
distributed, whereas /e E o/ (in German, Korean,
and English), /a/ (in German and Korean), /O/ (in
German and English), /2/ (in Korean and English),
/ø œ/ (in German), and /Ä/ (in English) were sig-
nificantly less diffusely distributed (i.e., more com-
pact). Crucially, none of the Condition ×Vowel
coefficients were significant, meaning that different
vowels did not vary significantly in terms of how
their vowel area was affected by intoxication.
4. GENERAL DISCUSSION
Taken together, the results of this study do not sup-
port the hypothesis of an L2-specific facilitative ef-
fect of intoxication on speech production. Contrary
to expectation, intoxication decreased vowel com-
pactness in L2 (English), just as it did in L1 for both
German and Korean. Intoxication-induced changes
in compactness also left a wide range in compact-
ness at the individual level (Table 1). Individual
variation in compactness, moreover, tended to be
correlated between sober and intoxicated speech, in
both L1 and L2. These findings provide evidence
that individual differences in compactness persist
across speaking conditions and throughout a bilin-
gual’s repertoire, consistent with the view that the
articulatory precision reflected in compactness is a
stable dimension of individual differences.
The contrast between our results and those of [10]
invites the question of why [10] observed facilita-
tive effects of intoxication on L2 speech but we did
not. There are two possible, and mutually compat-
ible, explanations. First, our L2 speakers may have
been more proficient than those in [10]. Second, our
study analyzed a specific, objective aspect of seg-
mental production, whereas [10] focused on a sub-
jective measure of overall pronunciation. What we
take away from these differences is that the effects
of intoxication on speech may be multifaceted and
variable over development. For example, intoxica-
tion may worsen acoustic consistency at the segmen-
tal level, but it may not necessarily affect proximity
to native norms, and it may well improve supraseg-
mentals that contribute to perceived pronunciation
quality, none of which were tested in this study.
While this study found persistent individual dif-
ferences in the acoustic consistency of vowels, the
findings are limited in a few ways that leave open
several directions for future research. For one, as
mentioned above, we examined only compactness,
so we cannot say for sure whether more-compact
speakers were also more “on target” vis-a-vis norms
than less-compact speakers. Thus, further work in-
vestigating central tendencies of L2 phonetic cate-
gories in conjunction with their distributions would
provide a fuller picture of individual differences in
L2 speech. In addition, for reasons of space, we
were only able to discuss the range of individual
differences in compactness, glossing over potential
variation in changes in compactness by vowel. In fu-
ture research, it would therefore be insightful to ex-
amine, for example, whether speakers showing aver-
age compactness pattern differently from the most-
compact and/or least-compact speakers in terms of
intoxication effects on specific vowels. Finally, our
results showed suggestive effects of language (Ger-
man vs. Korean) and language status (L1 vs. L2) on
compactness, which could be usefully elaborated in
future studies crossing languages with different-size
vowel inventories as L1 and L2.
3. Speech Production and Speech Physiology ID: 1035
1185
5. REFERENCES
[1] Trojan, F., Kryspin-Exner, K. 1968. The decay of
articulation under the influence of alcohol and par-
aldehyde. Folia Phoniatrica et Logopaedica 20,
217–238.
[2] Beam, S. L., Gant, R. W., Mecham, M. J. 1978.
Communication deviations in alcoholics; A pilot
study. Journal of Studies on Alcohol and Drugs
39, 548–551.
[3] Chin, S. B., Pisoni, D. B. 1997. Alcohol and
Speech. San Diego: Academic Press.
[4] Pisoni, D. B., Martin, C. S. 1989. Effects of alcohol
on the acoustic-phonetic properties of speech: Per-
ceptual and acoustic analyses. Alcoholism: Clinical
and Experimental Research 13(4), 577–587.
[5] Johnson, K., Pisoni, D. B., Bernacki, R. H. 1990.
Do voice recordings reveal whether a person is in-
toxicated? A case study. Phonetica 47, 215–237.
[6] Hollien, H., DeJong, G., Martin, C. A., et al. 2001.
Effects of ethanol intoxication on speech supraseg-
mentals. JASA 110, 3198–3206.
[7] Wieling, M., Blankevoort, C., Hukker, V., et al.
2019. The influence of alcohol on L1 vs. L2
pronunciation. In: Calhoun, S., Escudero, P.,
Tabain, M., Warren, P. (eds), Proceedings of the
19th ICPhS. Canberra: ASSTA Inc., 3622–3626.
[8] Tang, K., Chang, C. B., Green, S., et al. 2022. In-
toxication and pitch control in tonal and non-tonal
language speakers. JASA Express Lett. 2, 065202.
[9] Guiora, A. Z., Beit-Hallahmi, B., Brannon,
R. C. L., et al. 1972. The effects of experimen-
tally induced changes in ego states on pronuncia-
tion ability in a second language: An exploratory
study. Comprehensive Psychiatry 13, 421–428.
[10] Renner, F., Kersbergen, I., Field, M., et al. 2018.
Dutch courage? Effects of acute alcohol consump-
tion on self-ratings and observer ratings of foreign
language skills. Journal of Psychopharmacology
32, 116–122.
[11] Dörnyei, Z., Skehan, P. 2003. Individual differ-
ences in second language learning. In: Doughty,
C., Long, M. (eds), The Handbook of Second Lan-
guage Acquisition. Malden: Blackwell, 589–630.
[12] Chandrasekaran, B., Sampath, P. D., Wong,
P. C. M. 2010. Individual variability in cue-
weighting and lexical tone learning. JASA 128,
456–465.
[13] Perrachione, T. K., Lee, J., Ha, L. Y. Y., et al. 2011.
Learning a novel phonological contrast depends
on interactions between individual differences and
training paradigm design. JASA 130, 461–472.
[14] Bowles, A. R., Chang, C. B., Karuzis, V. P. 2016.
Pitch ability as an aptitude for tone learning. Lan-
guage Learning 66, 774–808.
[15] Chang, C. B., Kwon, S. 2020. The contributions of
crosslinguistic influence and individual differences
to nonnative speech perception. Languages 5, 49.
[16] Baker, W., Trofimovich, P. 2006. Perceptual paths
to accurate production of L2 vowels: The role of
individual differences. IRAL 44, 231–250.
[17] Raizada, R. D. S., Tsao, F.-M., Liu, H.-M., et al.
2010. Quantifying the adequacy of neural repre-
sentations for a cross-language phonetic discrim-
ination task: Prediction of individual differences.
Cerebral Cortex 20, 1–12.
[18] Idemaru, K., Holt, L. L., Seltman, H. 2012. Indi-
vidual differences in cue weights are stable across
time: The case of Japanese stop lengths. JASA 132,
3950–3964.
[19] Darcy, I., Park, H., Yang, C.-L. 2015. Individual
differences in L2 acquisition of English phonology:
The relation between cognitive abilities and phono-
logical processing. Learning and Individual Differ-
ences 40, 63–72.
[20] Schertz, J., Cho, T., Lotto, A. J., et al. 2016. Indi-
vidual differences in perceptual adaptability of for-
eign sound categories. Attention, Perception, and
Psychophysics 78, 355–367.
[21] Kartushina, N., Frauenfelder, U. H. 2014. On the
effects of L2 perception and of individual differ-
ences in L1 production on L2 pronunciation. Fron-
tiers in Psychology 5, 1246.
[22] Kartushina, N., Hervais-Adelman, A., Frauen-
felder, U. H., et al. 2015. The effect of phonetic
production training with visual feedback on the per-
ception and production of foreign speech sounds.
JASA 138, 817–832.
[23] British Council, IDP: IELTS Australia, Cam-
bridge English. 2023. IELTS in CEFR scale.
https://www.ielts.org/about-ielts/ielts-in-cefr-scale.
[24] Lemhöfer, K., Broersma, M. 2012. Introducing
LexTALE: A quick and valid Lexical Test for Ad-
vanced Learners of English. Behavior Research
Methods 44, 325–343.
[25] Miller, W. R., Muñoz, R. F. 1982. How to Control
Your Drinking. Englewood Cliffs: Prentice Hall.
[26] McAuliffe, M., Socolof, M., et al. 2017. Montreal
Forced Aligner: Trainable text-speech alignment
using Kaldi. In: INTERSPEECH-2017, 498–502.
[27] Boersma, P., Weenink, D. 2022. Praat: Doing pho-
netics by computer. Version 6.2.23, praat.org.
[28] Nearey, T. M. 1977. Phonetic Feature Systems for
Vowels. PhD thesis, University of Alberta.
[29] Kendall, T., Thomas, E. R. 2018. vowels: Vowel
manipulation, normalization, and plotting. R pack-
age version 1.2-2. Available from https://cran.r-
project.org/web/packages/vowels/.
[30] R Development Core Team, 2022. R: A language
and environment for statistical computing. Version
4.2.1, www.r-project.org.
[31] Kuznetsova, A., Brockhoff, P. B., Christensen,
R. H. B. 2017. lmerTest package: Tests in linear
mixed effects models. J. Stat. Softw. 82, 1–26.
1Random slopes for Condition by Participant were ex-
plored but not included because they caused the model to
be singular. Results remained the same regardless.
2The full model output showing all fixed-effect coeffi-
cients is available at osf.io/s9vkw/.
3. Speech Production and Speech Physiology ID: 1035
1186
