Content uploaded by Einar Meister
Author content
All content in this area was uploaded by Einar Meister on Mar 05, 2014
Content may be subject to copyright.
STEFAN WERNER (Joensuu), EINAR MEISTER (Tallinn)
MICRODURATION IN FINNISH AND ESTONIAN VOWELS REVISITED:
METHODOLOGICAL MUSINGS
Abstract. The influence of vowel duration on the perception of different vowel
qualities in Finnish and Estonian has been the topic of several of our recent stud-
ies; for the present paper, we reconsider some of our methodological choices, compar-
ing various different solutions. In the area of test design, timed group tests are
evaluated as an alternative to our original self-paced individual test setup and test
reliability is explored through repeated tests on the same subject. In the area of test
evaluation, reaction time is added as a second dependent variable and a more sophis-
ticated statistical evaluation is applied. All these methodological variations confirm
the trends already visible in the results of our earlier studies.
Keywords: Estonian, Finnish, vowels, intrinsic duration.
1. Introduction
The studies on microprosody in several languages have established systematic differ-
ences in the intrinsic features of vowels — open vowels tend to have lower F0,
higher intensity and longer duration than close vowels (e.g. Peterson, Lehiste 1960;
Solé 2007; Di Cristo 1978; Wahlen, Levitt 1995; Meister, Werner 2006). Our recent
studies address the intrinsic vowel duration in quantity languages like Estonian and
Finnish, mainly focusing on the role of intrinsic duration on the perception of different
phonological categories, i.e. vowel contrasts in close-open dimension and short vs.
long durational oppositions. We have shown experimentally that in boundary condi-
tions when spectral as primary features do not provide sufficient information for
category discrimination in close-open vowel pairs, the intrinsic duration of vowels
acts as a secondary feature facilitating the perceptual decision (Meister, Werner 2009).
In a subsequent study we have found further evidence for the impact of intrinsic
vowel duration by examining the categorical short vs. long distinction — the vowel
quality (hence intrinsic duration of a vowel) plays a significant role in the discrim-
ination of Estonian short vs. long phonological category (Meister, Werner, Meister
2011). The latter result is rather surprising since in quantity language like Estonian
duration has to be intentionally controlled by a speaker to signal quantity contrasts
and this ”higher order” control can ”override” the intrinsic features.
The aim of our current paper is to verify our previous findings on short vs.
long category discrimination by different groups of subjects involving Estonian and
Finnish subjects, and to address a number of methodological issues like different
test setups, intra-subject variations in repeated experiments, different methods
applied in the statistical analysis of the results.
180
LINGUISTICA URALICA XLVIII 2012 3 doi:10.3176/lu.2012.3.03
2. Methods and data
2.1. Stimulus corpus
For the perception experiments a stimulus corpus involving short vs. long cate-
gory oppositions in close vowel /i/ and open vowel /a/ in CV(:)CV carrier words
was designed. The stimuli were created from the nonsense words /kaka/, /kiki/,
/papa/, /pipi/, /tata/, and /titi/ pronounced in isolation by a native Estonian
male speaker. In all words the duration of the stressed vowel (V1) was manipu-
lated from 100 ms to 190 ms in 10 ms steps which consequently resulted in six
stimulus sets from CVCV to CV:CV — /kaka/ vs. /ka:ka/, /papa/ vs. /pa:pa/,
/tata/ vs. /ta:ta/, /kiki/ vs. /ki:ki/, /pipi/ vs. /pi:pi/, /titi/ vs. /ti:ti/. The dura-
tions of the other segments were kept constant (C1(burst) = 25 ms for /k/, 15 ms
for /p/ and /t/; C2 = 75 ms; V2 = 240 ms); the F0 was set to a constant value of
100 Hz in both vowels. The number of different stimuli in all sets was 10. The
manipulation of stimuli was done with Praat (Boersma, Weenink 2011).
In Estonian, the stimulus sets constitute a continuum from a word in quantity
one (Q1) to a word in quantity two (Q2) achieved by changing the duration of the
first-syllable vowel.
2.2. Test variations
Factors whose potential influence we wanted to assess were:
• test setup: self-paced (individual) vs. timed (group) test
• test-retest intra-subject variation
• test evaluation: reaction times as additional support
• test evaluation: statistical modelling
To investigate the possible effect of imposing a time limit on the subjects we
designed two slightly different group versions of our quantity perception tests, one
of which was administered to Estonian subjects, the other to Finnish subjects. The
Estonian version contained the full set of different stimuli from the individual tests
– two vowel qualities and three consonant articulation places — with each stimulus
played three times and an inter-stimulus interval of five seconds, whereas the Finnish
version only used three of the six stimuli (/kaka/, /kiki/, /papa/) but played every
stimulus five times using an inter-stimulus interval of three seconds. The Estonian
group involved 40 subjects whereas 30 subjects where native speakers of Estonian
(EST-L1) and 10 non-native subjects with Russian-language background (EST-L2);
the Finnish group involved 17 native speakers (FIN-L1).
For both native groups short vs. long category discrimination is natural since
both Estonian and Finnish exploit the duration cue contrastively; also L2 subjects
are able to discriminate Estonian short and long contrasts despite non-categorical
role of duration in Russian (Meister, Meister 2011).
In order to check for test-retest variation, one native Finnish and two native
Estonian subjects underwent the same test several times; for the Finnish subject,
reaction times were now also recorded. Instead of a linear regression analysis of
response frequency in terms of duration (as in Meister, Werner 2009), we fitted
more complete binomial logistic regression models, with and without random effects.
3. Results
3.1. Short vs. long boundaries
Overall group test results are in line with our previous studies: vowel openness
correlates positively with stimulus duration in all subjects’ groups (Figure 1). In
EST-L1 group the boundary mean in the case of high vowel lies at 146.8 ms and
Microduration in Finnish and Estonian Vowels...
181
in the case of low vowel 151.7 ms, in EST-L2 the categorical boundary values are
slightly lower — 142 ms and 147.3 ms for high and low vowel, correspondingly.
The boundary values for the Finnish group lie even at shorter vowel durations —
at 135.9 ms in the case of high vowel and at 144.5 ms in the case of low vowel.
The difference between the two means is significant in two native groups, in EST-L1
group at the 0.001 level (Welch two-sided t-test, t = 3.9; df = 174; p < 0.001) and in
FIN-L1 group at the 0.01 level (t = 2.7; df = 35; p < 0.01); in the EST-L2 group the
difference in category mean values between low and high vowel turned out to be
insignificant (t = 1.4; df = 35.9; p = 0.17).
The variation of responses broken down by stimulus duration is shown in Figure
2 (the difference in frequency range is due to the higher number of observations
per duration and subject in the Estonian test). The area of indecision around the
category boundary seems to spread out slightly more in the Estonian data. But this
can be due to, at least partly, the greater variation of segmental contexts in the
Estonian stimulus material and the larger number of Estonian test subjects (here
EST-L1 and EST-L2 groups are pooled together).
3.2. Test setup
Test setup does not seem to have a systematic influence on the perception test
results in our one-subject Finnish case study. The subject’s category boundary ranged
Stefan Werner, Einar Meister
182
Figure 1. Boxplots of the distributions of Estonian and Finnish speakers’ category
boundaries between ”long” and ”short” in low and high vowels.
Figure 2.Boxplots of the proportions of Estonian and Finnish speakers’ ”long” responses
across stimulus durations. The whiskers extend to 1.5 times the interquartile range,
indicating a 95% confidence interval for the difference in medians.
EST-L1 EST-L2 FIN-L1
Vowel duration (ms)
Estonian Finnish
Vowel duration (ms) Vowel duration (ms)
Frequency of ”long” responses
/a/ /i/ /a/ /i/ /a/ /i/
from 148.4 ms to 133.7 ms in the four identical individual tests which were self-
paced, and was 142.3 ms in the time-controlled group test.
3.3. Test-retest variation
As illustrated in the mosaic plots of Figures 3 and 4 intra-subject test-retest variation
turns out to be moderately high for our Finnish case who went through the test six
time (see previous section), but minimal for the two Estonian two-test cases: cate-
Microduration in Finnish and Estonian Vowels...
183
Figure 3.Test-retest comparison of response distributions for two Estonian subjects.
Figure 4.Test-retest comparison of response distributions for Finnish subject.
Estonian case AK Estonian case MK
Finnish case EP
Vowel duration (ms)
Vowel duration (ms)
Vowel duration (ms)
Response distributions
Response distributions Response distributions
gory boundaries are at 151.7 and 150.0 for subject AK and at 144.7 in both tests
for subject MK. All in all, the category boundaries between long and short are not
affected in a way that would challenge our overall results for both language and
intra-subject variation between tests does not exceed within-test intra-subject vari-
ation for repeated stimuli. Even in the Finnish case, median durations for long vs.
short only fluctuate between adjacent conditions: 170 vs. 160 ms and 120 vs. 110
ms for long and short responses, respectively.
3.4. Reaction time
In two of the four self-paced tests of Finnish subject EP reaction times were meas-
ured. As can be seen from Figure 5, there seems to be a slight trend for reaction
time to increase towards the category boundary which lies at 135.5 ms for these
two tests. There is a weak but significant negative correlation (r = –0.16, p < 0.001)
between the squared distance of stimuli’s’ duration from the category boundary
and reaction time. If a similar trend could be observed in other subjects as well it
would lend additional support to our estimation of the category boundaries.
Figure 5.Boxplots of Finnish subject EP’s reaction times (from two tests). The whiskers
extend to 1.5 times the interquartile range, indicating a 95% confidence interval for
the difference in medians (the zero value at 160 ms must be due to an inadvertent
keypress).
3.5. Statistical models
We fitted binomial logistic regression models using R's glm() and glmer() functions.
The mixed model analyses adding subject, stimulus, and/or presentation order as
random effects did not produce results that significantly differed from the fixed-
effects-only model: the only relevant factor, in addition to stimulus duration, is the
consonantal context. Vowel openness, although affecting the categorical boundary
in a minimal model with duration as the only factor (146 ms vs. 151 ms and 136
ms vs. 144 ms for high vs. low in Estonian and Finnish speakers, respectively),
does not improve the model fit significantly. Subjects' sex and age reduced model
deviance even less. Table 1 shows as an example the deviance analysis of a three-
factor model for the Finnish group data.
Stefan Werner, Einar Meister
184
Vowel duration (ms)
Reaction time (s)
Table 1
Analysis of deviance table for a binomial logistic regression model (logit link
function) of duration perception in the Finnish group test with factors duration,
consonant place of articulation and vowel openness
Df Deviance Resid. Df Resid. Dev. Pr(Chi)
NULL 2549 3523.4
dur 1 1733.08 2548 1790.4 2e-16 ***
cons 1 108.41 2547 1682.0 2e-16 ***
vow 1 1.47 2546 1680.5 0.2247
4. Discussion
Our new tests with Estonian and Finnish subjects lend further support to our
previous findings on the connection between intrinsic vowel duration and percep-
tual vowel categorization in quantity languages. Our case study of reaction time
measurements in addition to perceptual ratings also shows the same trend.
On the basis of our one-subject case study it seems that the influence of vari-
ations in the test set-up can be neglected but more data will be needed to prove
this point. Finally, more sophisticated statistical analyses with mixed models instead
of fixed-factors-only models do not introduce new insights into our data.
All in all, the collection of results from new data and reconsiderations of method-
ological solutions presented here consolidates the concept of micro- and macro-
duration interplay developed already in our earlier studies.
Acknowledgement
This work has been partly supported by the target-financed theme No. 0140007s12
of the Estonian Ministry of Education and Research. We are grateful to Lya Meis-
ter for conducting the perception experiments with Estonian listeners and to all
subjects who participated in the study.
Addresses
Stefan Werner
Department of Linguistics, University of Eastern Finland
E-mail: stefan.werner@uef.fi
Einar Meister
Institute of Cybernetics, Tallinn University of Technology, Estonia
E-mail: einar@ioc.ee
REFERENCES
B o e r s m a, P., W e e n i n k, D. 2011, Praat: doing phonetics by computer
(Version 5.2.09) [Computer program]. http://www.praat.org/.
Di Cristo, A. 1978, De la microprosodie à l’intonosyntaxe. Thèse d’Etat, Aix-en-
Provence.
M e i s t e r, L., M e i s t e r, E. 2011, Perception of the Short vs. Long Phonological
Category in Estonian by Native and Non-Native Listeners. — Journal of
Phonetics 39, 212—224.
M e i s t e r, E., W e r n e r, S. 2006, Intrinsic Microprosodic Variations in Estonian
and Finnish. Acoustic Analysis. — Fonetiikan Päivät 2006, Helsinki (Publications
of the Department of Speech Sciences, University of Helsinki 53), 103—112.
—— 2009, Duration Affects Vowel Perception in Estonian and Finnish. — LU
XLV, 161—177.
Microduration in Finnish and Estonian Vowels...
185
M e i s t e r, E., W e r n e r, S., M e i s t e r, L. 2011, Short vs. Long Category
Perception Affected by Vowel Quality. — ICPhS XVII. The 17th Interna-
tional Congress of Phonetic Sciences, Hong Kong, 17—21 August 2011, Hong
Kong, 1362—1365.
P e t e r s o n, G. E., L e h i s t e, I. 1960, Duration of Syllable Nuclei in English.
— Journal of the Acoustical Society of America 32, 693—703.
S o l é, M. J. 2007, Controlled and Mechanical Properties in Speech. A Review
of the Literature. — Experimental Approaches to Phonology, Oxford, 302—
321.
W a h l e n, D. H., L e v i t t, A. G. 1995, The Universality of Intrinsic F0 of
Vowels. — Journal of Phonetics 23, 349—366.
СТЕФАН ВЕРНЕР (Йоэнсуу), ЭЙНАР МЕЙСТЕР (Таллинn)
O МИКРОДЛИТЕЛЬНОСТИ ФИНСКИХ И ЭСТОНСКИХ ГЛАСНЫХ.
МЕТОДОЛОГИЧЕСКИЕ РАЗМЫШЛЕНИЯ
Влияние длительности гласного на восприятие качества гласных в финском и
эстонском языках изучено в нескольких наших недавних исследованиях, в дан-
ной статье пересмотрivaœтся некоторые из наших методологических podhodов.
В äасти разработки слуховых экспериментов группoвой тест сравнивается с is-
hodным индивидуальным, a dostoverность результатов proverqется повтор-
ныmi экспериментami с одним субъектом. Для дополнительной оценки теста
измеряется время реакции и при анализе реzультатов ispolxzuœтся более
сложные статистические методы. Все эти методологические вариации под-
тверждают результаты наших предыдущих исследований.
Stefan Werner, Einar Meister
186