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

net/publication/261376774

Articulatory vowel spaces of male and female speakers

Conference Paper · May 2014

CITATIONS READS

8 2,224

2 authors:

Melanie Weirich Adrian P. Simpson

Friedrich Schiller University Jena Friedrich Schiller University Jena
69 PUBLICATIONS 522 CITATIONS 103 PUBLICATIONS 1,483 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Melanie Weirich on 07 April 2014.

The user has requested enhancement of the downloaded file.

 Articulatory vowel spaces of male and female speakers
Melanie Weirich, Adrian P. Simpson
Friedrich-Schiller-University, Jena, Germany
[email protected], [email protected]

between accent/stress, degree of coarticulation and target

Abstract undershoot in terms of vowel reduction seems to exists. How-
ever, if females reach their articulatory targets earlier and more
Studies have shown that although females on average have a
larger acoustic vowel space than males, they exhibit a smaller often than men, as is hypothesized above, then they should be
articulatory vowel space. From this it is hypothesized that sex- less influenced by accent induced undershoot. To test this,
specific differences in undershoot might exist. Articulatory articulatory vowel spaces are compared between male and
vowel space sizes and Euclidean distances between vowel female speakers and the potential interaction of accent and
positions are analyzed in nine German speakers (5m, 4f) by sex-specific differences is analyzed. In addition, if females
means of electromagnetic articulography. Analyses include reach their articulatory targets earlier, they should also be less
different sentence accent positions and two different sequences affected by coarticulatory induced undershoot. Therefore,
varying in their expected coarticulatory induced degree of articulatory vowel spaces of male and female speakers are
undershoot. Results show a relationship between undershoot compared a) in temporally privileged positions where the
and speaker sex with males being more affected by accent and effect of coarticulatory induced undershoot is expected to be
coarticulatory induced undershoot than females. minimal and b) in the sequence /gV/.

Keywords: vowel space, sex-specific differences, articulation 2. Method

1. Introduction 2.1. Speakers and speech material

For several languages it has been found that female speakers Articulatory recordings of 5 male and 4 female German speak-
have a larger acoustic vowel space than male speakers (e.g. ers were made at Potsdam University with the NDI-Wave
Diehl et al. (1996) for American English; Whiteside (2001) for system. The speakers were between 23 and 43 years old and
British English; Weirich & Simpson (2014) for German). revealed no known speech or hearing impairments. All of
Different reasons have been held responsible for this including them came from the Eastern Central German dialect area but
behavioral and physiological factors. Despite having a larger showed very little dialectal influence based on the auditory
acoustic vowel space, females have been found to have a impression of the authors. Altogether, six coils were attached
smaller articulatory space (Simpson 2001, 2002). Furthermore, to the tongue, the lips, and the lower jaw. Four coils, one
Simpson (1998) found sex-specific differences in correlations above the upper incisors, one at the bridge of the nose and two
between formant values and duration. While male speakers behind the participants’ ears served as reference coils which
showed the expected significant relationship in read speech could be used to compensate for head movements. The bite
between duration and e.g. F1 of /aː oː ɔ ʊ/, female speakers plane was established using three sensors attached to a set
often did not exhibit this relationship. From this and the mis- square which a speaker held between his/her teeth.
match between articulatory and acoustic vowel spaces the The speech material presented here is twofold and part of a
question arises whether there might be sex-specific differences larger corpus comprising 20 different target words in varying
in terms of undershoot. In other words, do females reach their accent conditions.
articulatory targets more often than males? The first set of data used in the current study includes the three
The phenomenon of undershoot has been found cross- point vowels /aː uː iː/ contained in the vowel sequences in the
linguistically in many studies investigating the role of accent abbreviation IAA, AUU and BII. The articulatory positions of
and lexical stress (Öhmann 1967, Fowler 1981, Rietveld & the vowels were expected to be extreme and only minimally
Koopmans-van Beinum 1987, de Jong 1995, Palethorpe et al. effected by coarticulatory influences due to their temporally
1999, Harrington et al. 2000, Cho 2004). Vowel undershoot privileged occurrence in the abbreviations. Each target word
patterns have been explained both in terms of paradigmatic was repeated 10-12 times and embedded in a carrier sentence,
enhancement and by varying degrees of coarticulation. With e.g. Sie fuhren letzte Woche zur IAA ganz schnell (‘They went
respect to paradigmatic enhancement, more peripheral vowels to the IAA very fast last week’).
are supposed to be found in stressed or accented positions, The second set of data comprises the sequence /gV/ with V
whereas in unstressed or unaccented positions vowel centrali- being /iː ɪ eː ɛ aː a oː ɔ uː ʊ/ in the name GVbi embedded in the
zation is expected. Following the coarticulatory explanation, carrier sentence Ich sah GVbi an (‘I looked at GVbi’). For the
undershoot in unstressed or unaccented positions is a product /gV/-material, three different accent conditions were recorded.
of a higher degree of coarticulation related to contextual reduc- First, the participants were asked to read the sentences pre-
tion. Vowel reduction due to increased coarticulation is ex- sented to them from a screen (neutral condition, n). Second,
plained by target undershoot reflecting the shorter durations in speakers produced the name in response to questions from the
unstressed/unaccented positions. Mooshammer & Geng (2008) experimenter. For the accented condition (a) the experimenter
examined acoustic and articulatory vowel reduction patterns of asked 1) Sahst du /gVbi/ oder /gVbi/ an? (‘Did you look at
tense and lax vowels in German in stressed and unstressed /gVbi/ or /gVbi/?’), eliciting in the reply an accentuation of the
syllables. They found a higher degree of coarticulation in name. For the unaccented condition (u) the experimenter asked
unstressed vowels than in stressed vowels. Thus, a relationship 2) Siehst du oder sahst du /gVbi/ an? (‘Do you or did you look
at /gVbi/?’), eliciting in the reply an accentuation of the verb. mm² vs. 93 mm²). Moreover, the dimensions of the space
Each vowel was again repeated 10-12 times in each accent differ between the sexes: while the male speakers reveal a
condition. The sentences were recorded in a randomized order larger vertical than horizontal expansion (around 1.3 times
with additional filler sentences in between. larger), the relationship between vertical and horizontal expan-
sion for the females is ca. 1 (or even less). A significant differ-
2.2. Articulatory labeling and analyses ence between males and females was found for the Euclidean
distance (ED) between /iː/ and /aː/ with a larger distance for
All articulatory labeling was done with the help of mview
males (Welch two-sample t-test, t = -2.7, df = 5.9, p < .05).
(software written by M. Tiede). Here, we will concentrate on
males females
the horizontal and vertical position of the backmost lingual
sensor (tongue dorsum). First, the articulatory position at the

5 10

5 10
acoustic midpoint of the double vowel sequences were meas-

TBy_trans0

TBy_trans0
ured (cf. Figure 1, left graph).This provided a notional extreme

0
articulatory vowel space for each speaker (the IAU-polygon).
To compare the articulatory positions across speakers, the data

-10

-10
was translated with the midpoint set to the origin (0/0). The
vowel space size of the IAU-polygon and the Euclidean Dis- -15 -5 5 15 -15 -5 5 15
tances (EDs) between the vowels were calculated. Also, the TBx_trans0 TBx_trans0
EDs from the midpoint of the polygon to each vowel were
measured for each speaker. Second, the start and end of the Figure 2: Vowel spaces for male (left) and female (right)
opening gesture /gV/ was labeled for each speaker, stimulus speakers centered around 0/0
and accent condition oriented on the tangential velocity of the
tongue dorsum sensor (cf. Figure 1, right graph). The end When we look at the articulatory positions of the different
points of the gestures (which correspond to the vowels) were vowels, we can see that the sexes differ particularly in /aː/ and
used to parameterize the polygon area spanned by the vowels /iː/, while the positions for /uː/ overlap (Figure 3).
contained in the /gV/-sequences and the EDs from the mid-
point. Third, each speaker’s /gV/-polygon was normalized by
expressing it as a percentage of the temporally privileged IAU-
5
TB_y (mm)
vowel space. The percentages were calculated for the polygon
0
size and the EDs between vowels and midpoint.
For statistical analyses, Welch two sample t-tests and linear
-5

mixed models (LMMs) were conducted. LMMs as implement-

-10

ed in the lme4 package (Bates et al. 2011) were run in R (ver-

sion 2.14.1, R Development Core Team 2008) with the articu- -10 -5 0 5 10
latory measurements as dependent variable, the fixed factors
speaker sex, vowel and accent condition (for the /gV/-data), TB_x (mm)

and the random factors speaker and repetition. Likelihood ratio Figure 3: Vowel spaces of the /IAU/-polygon for all male
tests were used for model comparisons with different factors (blue) and female (red) speakers
included to find the model with the best fit to the data.
This is mirrored in the statistical results. For both vertical and
horizontal tongue position as dependent variable, LMMs
revealed the best fit to the data with the interaction term
Sex*Vowel included. For the horizontal tongue position,
significant differences between males and females were found
for /aː/, for the vertical tongue position males and females
differ significantly in /aː/ and /iː/ (see Table 1). Thus, the
female vowel space is smaller in terms of a higher and more
fronted position for /aː/ and a lower position for /iː/.

Table 1: Summary statistics for fixed factors Sex and Vowel,

dependent variable is vertical tongue position (default level is
vowel /iː/ for females, observation: 263, rep.: 12, speakers: 9)
Figure 1: Articulatory labeling in mview for the abbrevations Estimate Std.Error t-value pMCMC
(left graph: midpoint of double vowel /iːiː/ marked by the red Intercept 4.1 0.2 19.1 <.001
vertical line) and the /gV/-sequences (right graph: start and Sex [f vs. m] 1.7 0.2 5.8 <.001
end of the sequence marked by the red lines) Vowel [/iː/ vs. /aː/] -10.4 0.3 -34.5 <.001
Vowel [/iː/ vs. /uː/] -2.0 0.3 -6.5 <.001
Sex[m]*Vowel[/aː/] -3.7 0.4 -9.1 <.001
3. Results Sex[m]*Vowel[/uː/] -1.1 0.4 -2.8 <.01
3.1. The impact of speaker sex on the articulatory
space of the IAU-polygon
Figure 2 shows the mean articulatory vowel spaces measured 3.2. The impact of speaker sex and accent condition
in the double vowel sequences separated by sex. The data was on the articulatory space of the /gV/-sequence
translated and centered on the origin. As expected, the females Figure 4 (left graph) shows the polygon area spanned by the
on average reveal a smaller vowel space than the males (66 tense vowels from the /gV/-sequence separated by sex and
 accent condition. First of all, the figure reveals a higher inter- 3.3. The impact of speaker sex and accent condition
speaker variability for the males than for the females for all on the relationship between IAU-polygon and /gV/-
conditions. A possible reason for this might be that males are sequence
less restricted than females in terms of physiological bounda-
ries and thus show this articulatory freedom in more variabil- To take the differences in speaker-specific physiology into
ity. The figure also shows that while males reveal on average account (in terms of its effect on the articulatory space that can
larger polygon sizes for the neutral and the accented condition, be used in a temporally privileged condition), the /gV/-
the picture is reversed for the unaccented condition. The same sequence was analyzed with respect to the speaker specific
was found for the lax vowels, but here high inter-speaker (extreme) IAU - vowel spaces. Figure 5 visualizes this rela-
variability was found for males and females. The right graph tionship for four speakers. While the filled red points show the
of Figure 4 shows the EDs measured from the midpoint of the IAU - vowel space, the black and grey markers show the
polygons to the tense vowels /iː eː aː oː uː/. Again, the data is gestures /gV/ (start /g/ = black crosses, end /V/= grey dots).
separated by speaker sex and accent condition. It is apparent The genders vary in terms of the use of their articulatory vow-
that the difference between the accented and unaccented con- el space: both male speakers show a bigger vowel space meas-
dition is much higher for the males than for the females. ured in the temporally privileged IAU (red) – especially in
terms of a lower tongue position for /aː/ – than in the /gV/
sequence (black and grey symbols). The females do not seem
ED MP-tense vowels (mm)

f m
Polygon tense (mm²)

15 to differ between the two conditions.

n a u
M1 M2
100 10
80

5
60

TB_y (mm)

TB_y (mm)
5
40

-5

-5
f m f m f m 0

-15

-15
n a u n a u

Figure 4: Polygon sizes spanned by the tense vowels of the 35 45 55 65 35 45 55 65

/gV/-sequence (left graph) and ED from the midpoint to the
TB_x (mm) TB_x (mm)
tense vowels of the /gV/ sequence (right graph) separated by
sex and accent condition F1 F2

15
15

In order to increase the number of data points that can be used,

TB_y (mm)
TB_y (mm)

but also to enable a vowel specific analysis to be undertaken,

0 5
0 5

the EDs from the midpoint of the vowel space to each vowel
were used as dependent variable for the statistical analysis
-10

(and not the overall polygon size). Model comparisons showed -10
that the LMM with the interaction terms Sex*Accent and 30 40 50 60 30 40 50 60
Sex*Vowel as fixed factors revealed the best fit to the data
TB_x (mm) TB_x (mm)
(random factors included were speaker and repetition). Table 2
Figure 5: Tongue dorsum positions during the IAU (extreme)
gives the summary statistics for the fixed effects. To summa-
rize, there is a significant difference between males and fe- vowel space (red) and /gV/-sequence of two male (M1, M2)
and two female (F1, F2) speakers
males in the EDs only for the vowel /aː/. Second, males show
a significant difference between the accented and unaccented
Analogous to Figure 4 above, Figure 6 shows the polygon
condition, while females do not. Since there was no three-way
interaction of Accent, Vowel and Sex, this sex-specific differ- spanned by the tense vowels of the /gV/-sequence separated by
sex and accent condition (left graph). However, this time the
ence in the effect of accent seems to hold for all vowels.
values are expressed in percent of the IAU-polygon (calculat-
ed for each speaker and accent condition separately). Interest-
Table 2: Summary statistics for factors Sex, Vowel and Accent,
dependent variable is absolute ED between midpoint and all ingly, the males reveal smaller values than the females in
every accent condition, but especially in the unaccented one.
tense vowels (in mm) (default level: vowel /aː/ in accented
condition for females, observations:1104, rep:12, speakers:9) The right graph shows the EDs from the midpoints to the tense
vowels /iː aː uː/ expressed as percent of the EDs measured in
Estimate Std.Error t-value pMCMC
the IAU-polygon. Here again, the males differ between the
(Intercept) 5.1 0.6 9.1 <.001
Accent [a vs. n] 0.5 0.2 2.7 <.001 accented and unaccented condition, while the females do not.
Accent [a vs. u] -0.1 0.2 -0.3 0.8
ED MP-tense vowels (%)

f m
Sex [f vs. m] 2.3 0.7 3.1 <.01
Polygon tense (%)

Vowel [/aː/ vs. /eː/] 2.3 0.3 9.0 <.001 n a u 200

Vowel [/aː/ vs. /iː/] 2.4 0.2 9.7 <.001 120
150
Vowel [/aː/ vs. /oː/] 1.9 0.2 8.1 <.001 100
80
Vowel [/aː/ vs. /uː/] 1.2 0.2 4.9 <.001 100
60
Accent [n]*Sex [m] -0.8 0.3 -3.1 <.01 40 50
Accent [u]*Sex [m] -1.4 0.3 -5.1 <.001 f m f m f m
0
Sex[m]*Vowel[/eː/] -1.4 0.3 -4.0 <.001 n a u n a u
Sex[m]*Vowel[/iː/] -1.6 0.4 -4.4 <.001
Sex[m]*Vowel[/oː/] -2.8 0.3 -8.6 <.001 Figure 6: Normalized polygon sizes (in % of the IAU-space)
Sex[m]*Vowel[/uː/] -1.9 0.3 -5.7 <.001 spanned by tense vowels of /gV/-sequence (left) and EDs
between midpoint and /iː aː uː/ of the /gV/-sequence (in % of
the IAU-EDs) separated by sex and accent condition (right)
 For the statistical analysis LMMs were run with the normal- female speakers. We suggest that this might be caused by a
ized EDs as dependent variable and likelihood ratio tests were stronger impact of coarticulation in males than in females.
used to find the model with the best fit to the data. This time Since females have on average a smaller articulatory space due
the model with only the interaction term Accent*Sex as fixed to physiological restrictions, they reach their targets earlier and
factor turned out to be best (random factors were again speak- thus are less affected by coarticulatory influences and also by
er and repetition). Table 3 summarizes the results of the fixed accent induced undershoot.
factors. In contrast to the absolute values, there is no sex-
specific difference anymore in the ED for /aː/, or for the other 5. Acknowledgements
vowels. However, as in the model on the absolute data, the
factor Accent reveals its significance in terms of sex-specific This work was supported by a German Research Council
differences regarding its influence: while male speakers differ Grant (SI 743/6-1/2) awarded to the second author. We would
between the accented and unaccented condition, females do like to thank all participating subjects, our student assistant
not. Again, the three way interaction of Sex, Vowel and Ac- Miriam Oschkinat and Jana Brunner, Christian Geng and
cent was not significant and including it did not reveal a better Adamantios Gafos at the Dept. of Linguistics, University of
fit to the data which indicates that the interaction of sex and Potsdam where the articulatory recordings were made.
accent is valid for all vowels.
6. References
Table 3: Summary statistics for factors Sex and Accent condi- Cho, T. (2004). Prosodically conditioned strengthening and vowel-to-
tion, dependent variable is normalized ED between midpoint vowel coarticulation in English. Journal of Phonetics 32(2),
and tense vowels (in %) (default level is vowel /aː/ in accented 141–176.
condition for females, observations: 664, rep.: 12; speakers 9)
de Jong, K. (1995). The supraglottal articulation of prominence in
Estimate Std.Error t-value pMCMC English: Linguistic stress as localized hyperarticulation. J.
(Intercept) 83.1 5.4 15.4 <.001 Acoust. Soc. Am. 97, 491–504.
Accent [a vs. n] 7.1 3.1 2.3 <.05
Accent [a vs. u] 0.5 3.2 0.2 0.9 Diehl, R. L.; Lindblom, B.; Hoemeke, K. A. & Fahey, R. P. (1996) On
explaining certain male-female differences in the phonetic real-
Sex [f vs. m] 5.1 7.3 0.7 0.5 ization of vowel categories. Journal of Phonetics 24, 187-208
Accent [n]*Sex [m] -12.4 4.4 -2.9 <.01
Accent [u]*Sex [m] -18.1 4.4 -4.1 <.001 Fowler, C. (1981). Production and perception of coarticulation among
stressed and unstressed vowels. Journal of Speech and Hearing
Research 46, 127–139.
4. Discussion and conclusion
Harrington, J., Fletcher, J., & Beckman, M. (2000). Manner and place
To summarize, our results replicate the findings of Moosham- conflicts in the articulation of accent in Australian English. In
mer & Geng (2008) regarding more undershoot in unaccented M. Broe & J. Pierrehumbert (Eds.), Papers in Laboratory Pho-
than in accented condition in German vowels. At the same nology V: Acquisition and the lexicon, 40–51. Cambridge:
time, we have found evidence for sex-specific differences in Cambridge University Press.
the degree of undershoot already hinted at by Simpson (1998). Mooshammer, C. & Geng, C. (2008). Acoustic and articulatory mani-
Males reveal larger articulatory vowel spaces, as found in a festations of vowel reduction in German, Journal of the Interna-
temporally privileged context in terms of a lower and more tional Phonetic Association 38/2, 117-136.
retracted tongue position for /aː/ and a higher tongue position
Öhman, S. (1967). Numerical model of coarticulation. J. Acoust. Soc.
for /iː/. The ED between these two vowels is significantly Am. 41,310–320.
larger in males and thus, the vowels are articulatory further
apart than in females. However, these sex-specific differences Palethorpe, S., Beckman, M., Fletcher, J. & Harrington, J. (1999).
were found to be less in the coarticulatorily more vulnerable The contribution of schwa vowels to the prosodic accent con-
trast in Australian English. The 14th International Congress of
/gV/-sequence (in the neutral and accented condition) or even Phonetic Sciences, San Francisco, 695–698.
reversed (in the unaccented condition). Thus, sex-specific
differences strongly depend on the accent condition. In detail, R Development Core Team (2008). R: A language and environment
this means that while females did not differ in articulatory for statistical computing, R Foundation for Statistical Compu-
ting, Vienna, Austria, http://www.R-project.org. (Last viewed
dimensions between accented and unaccented tense vowels,
07/19/2012).
males revealed smaller excursions in the unaccented condition.
Hence, the sexes differed in vowel reduction patterns that are Rietveld, A. C. M., & Koopmans-van Beinum, F. J. (1987). Vowel
due to target undershoot in an unaccented condition. In addi- reduction and stress. Speech Communication 6, 217–229.
tion, in terms of normalized articulatory dimensions expressed Simpson, A. P. (1998) Phonetische Datenbanken des Deutschen in der
in percent (with the help of a speaker-specific extreme articu- empirischen Sprachforschung und der phonologischen Theorie-
latory vowel space measured in a temporally privileged con- bildung. Arbeitsberichte des Instituts für Phonetik und digitale
text) sex-specific differences decrease or disappear (as for the Sprachverarbeitung der Universität Kiel (AIPUK) 33).
ED from the midpoint to /aː/) or even go in the other direction Simpson, A. P. (2001). Dynamic consequences of differences in male
with higher values for females than males (as for the polygon and female vocal tract dimensions, J. Acoust. Soc. Am. 109,
sizes). 2153–2164.
Although the neutral accent condition was elicited as well, our
Simpson, A. P. (2002). Gender-specific articulatory-acoustic relations
analysis here has focused mainly on comparing the accented in vowel sequences, J. Phonetics 30, 417–435.
and the unaccented condition. But as we can see from the
figures and the tables for the EDs, there is a tendency for the Weirich, M. & Simpson, A. P. (2014). Acoustic vowel space size and
female values to be highest in the neutral condition, for males perceived speech tempo, Journal of Phonetics 43. 1-10.
in the accented condition. Whiteside, S. P. (2001) Sex-specific fundamental and formant fre-
To conclude, we interpret our findings in terms of a higher quency patterns in a cross-sectional study. Journal of the
amount of undershoot in unaccented conditions in male than in Acoustical Society of America 110, S. 464-478.

View publication stats