Vowel Space, Speech Rate and Language Space

Abstract

Acceleration of speech rate is often said to be correlated with a reduction of the vowel space. However, a monocausal explanation of the vowel space reduction by speech rate is surely too simplistic. With our regionally balanced database of a German text read in two reading tempi we present geolinguistic maps of • the different sizes of the vowel space and • different changes of the sizes of the vowel space when comparing normal and accelerated speech rates. These maps for the normal reading tempo show regional patterns of vowel space sizes for the long and short vowel system. Accelerating speech rate affects the vowel space size in regionally specific patterns. In addition, increasing reading tempo shows a surprising general effect: initially large vocal spaces are reduced while initially small vocal spaces are enlarged. Interestingly, vowel space size and change due to accelerating reading tempo does only limitedly reflect traditional dialect regions.

Full text

VOWEL SPACE, SPEECH RATE AND LANGUAGE SPACE
Beat Siebenhaar and Matthias Hahn
Universität Leipzig, Germany
siebenhaar@uni-leipzig.de, matthias.hahn@uni-leipzig.de
ABSTRACT
Acceleration of speech rate is often said to be corre-
lated with a reduction of the vowel space. However,
a monocausal explanation of the vowel space reduc-
tion by speech rate is surely too simplistic. With our
regionally balanced database of a German text read in
two reading tempi we present geolinguistic maps of
• the different sizes of the vowel space and
• different changes of the sizes of the vowel space
when comparing normal and accelerated speech
rates.
These maps for the normal reading tempo show
regional patterns of vowel space sizes for the long and
short vowel system. Accelerating speech rate affects
the vowel space size in regionally specific patterns. In
addition, increasing reading tempo shows a surprising
general effect: initially large vocal spaces are reduced
while initially small vocal spaces are enlarged.
Interestingly, vowel space size and change due to
accelerating reading tempo does only limitedly reflect
traditional dialect regions.
Keywords: German, speech rate, tempo, vowel
space, geophonetics
1. INTRODUCTION
Variationist geolinguistic and sociophonetic research
focuses on specific sounds or the relation of sounds
within a linguistic system. Realisations of specific
sounds are correlated to areal, social or interactional
factors. The vowel space is architecturally conceptu-
alized in terms of relative questions like: Is /eː/ more
fronted than /iː/? Where is the position of an /aː/?
Usually, the size of the vowel space itself is not under
investigation. However, comparison of different
studies makes it obvious that the vowel space area is
quite variable. Fig. 1 shows different representations
of the vowel space for the long monophthongs of
Standard German. These studies – except for [13]
with data from a word list – represent data from rea-
ding tasks and have been analysed differently, so that
the results are only comparable to a limited extent.
Recordings of [13] were made in Berlin; however,
the authors claim that their students had no dialectal
influence. [15] analyses the Kiel Corpus [5] with
recordings from the northwestern part of Germany.
[11] examines data from Vienna, representing an
Austrian standard, and [14] looks at data from Leipzig
in East Central Germany.
Figure 1: Comparison of the vowel space of long
vowels in standard German intended recordings.
This comparison makes it evident that there is
variation in the geometry of the vowel space as well
as in the size of the vowel space. It also suggests that
there is regional variation in Standard-intended Ger-
man speech – that is, Standard German as performed
regionally [4, 7, 8]. Moreover, the analyses [5, 11, 12,
14, 15] document stylistic differences and [18] shows
that speech rate affects articulation accuracy in many
ways. In addition to segment elision and an increase
in coarticulation phenomena, vowel undershoot is a
typical feature attributed to higher speech rate. The
hypothesis in [9] postulates that a vowel target may
not be reached by the articulators under the temporal
constraints of an accelerated speech rate, resulting in
formant undershoot and a reduction in the vowel
space size; however, this theory is by no means uni-
versally accepted and findings are inconsistent [12].
With a geographically balanced dataset of German
read speech in two speech rates [1, 2, 3] we will focus
on the following questions:
• Is there regional variation in vowel space size?
• Is there regional variation in how long and short
vowels pattern in terms of vowel space size?
• How does acceleration of speech rate affect these
patterns?
2. DATA AND METHODS
2.1. Data
The data are part of the “Deutsch heute” corpus from
the Institute for German Language (IDS) [1, 7]. For
this study the Aesop fable The North Wind and the
Sun was used as reading material. This was recorded

twice per speaker, once in a “normal” reading tempo,
and then in a “fast” reading tempo. The subjects were
two male and two female high school students per
place. They were aged 17–20, local to the area under
investigation, and not professional speakers. The
recording took place during lesson time in school in a
quiet room. The study area covers the whole conti-
guous German-speaking area of Germany, Austria,
Switzerland, Liechtenstein, East Belgium and South
Tyrol, from which 161 evenly distributed locations
were selected. Altogether, there are 644 recordings in
two reading tempi.
2.2. Preprocessing
For forced alignment we used the Munich Automatic
Segmentation tool (MAUS) [6]. Then the first two
formants were measured in the central 60% of the
monophthongs. A script calculated formant values at
30 positions, then the median of these formant values
was taken for each of the 260,000 individual vowels.
In order to equalise gender and individual differences,
the data were z-normalised [10], that means that the
data were transformed in a way that the centre of all
vowels got a 0 and the distribution reached a standard
deviation of 1. These z-normalised data were the basis
for the measurement of the vowel space, represented
in z-normalised squared units.
2.3. Measuring the vowel space
The vowel space of the German standard comprises
six vowels. However, the North Wind and the Sun
features only one occurrence of /ɛː/ and of /uː/ for
each recording, so the data density for these two
vowels is quite weak. Therefore, we only measure the
vowel space for long vowels in the area between /iː/,
/eː/, /aː/, and /oː/, marked in Fig. 2 in grey.
Figure 2: Vowel space of the Standard German
long vowel system (solid line) and the area used for
this paper (shaded area).
The exclusion of /ɛː/ is not of great concern for this
analysis, since there is a tendency for /ɛː/ and /eː/ to
merge as /eː/ in the north of the German-speaking area
and in Austria [7]. If /ɛː/ were taken into account, the
vowel spaces would no longer be directly compa-
rable. Omitting /uː/ is more problematic due to its
position in the high back corner of the vowel space.
However, the available data are not sufficient for the
analysis of /uː/.
For the vowel space of the short vowels, we use
the corresponding area between /ɪ/, /ɛ/, /a/, and /ɔ/.
To calculate each vowel space, the first two for-
mants of every individual monophthong were mea-
sured. Thirty measurements were taken within the
central 60% of each vowel, and the median of these
was calculated. The data were then Lobanov- or z-
normalised [16] – transformed so that the centre of all
vowels received a value of 0 and the distribution had
a standard deviation of 1. This allows a comparison
of speakers with various vocal physiologies [17].
These normalised formant measurements formed the
basis for the calculation of vocal space size. The
vowel space area for every speaker was measured for
long and short vowels separately, each in the normal
and fast reading tempi.
The vowel space values were mapped with Arc-
GIS using a local smoothing to level out outliers. This
means that the value mapped for every polygon
represents the median value of this polygon and all its
neighbouring values. Thus each area point is
represented by 20 to 32 speakers, and because of the
robustness against outliers the median is used.
3. RESULTS
3.1. Vowel spaces of males and females
Sociolinguistic research often finds differences be-
tween male and female speakers. Therefore, we first
compare them to test whether we can treat the datasets
of males and females as one larger and therefore ge-
nerally more reliable dataset, or whether they are
distinct. Over the whole dataset, we find correlations
of male and female data of r=0.45 for the long vowel
system and r=0.48 for the short vowel system, so the
chance remains that the datasets of males and females
can be treated as one. Fig. 3 maps the vowel spaces of
long vowels in normal reading tempo. The brown
areas mark larger vowel spaces, the blue areas mark
smaller vowel spaces.
Figure 3: Comparison of the vowel space of long
vowels in normal reading tempo of male (left) and
female (right) speakers. Larger vowel spaces in
brown, smaller vowel spaces in blue.

The comparison for the long vowels in normal rea-
ding tempo shows very similar regional distributions
across genders, with only a few larger differences in
the central western part and around Berlin. The cor-
respondences between male and female speakers for
the fast reading tempo and for the short vowel system
are comparable, but each shows different areal
distributions. Finally, it can be said that the regional
distributions of vowel space size for men and women
are quite similar, though only two men and two
women were recorded per place. Therefore, in the
following analyses we treat data from males and
females in a common dataset.
3.2. Ratio of vowel space sizes of the short vowel to
long vowel systems at different reading tempi
Before presenting the vowel spaces for the long and
short vowel systems, we will have a look at the
relationship between the two systems. Fig. 4 shows
the size of the vowel space of the short vowel system
as a proportion of the size of the vowel space of the
long vowel system. Red areas mark big differences in
the size of the vowel spaces of the long and the short
vowel system, while blue areas mark smaller differ-
ences in the size of the two vowel systems. The two
maps in Fig. 4, representing the two reading tempi,
are very consistent. This means that changes in
reading tempo affect long and short vowels in a
similar matter.
Figure 4: Ratio of the vowel space sizes of the short
vowel system to the long vowel system at normal
(left) and fast (right) reading tempi. Small differe-
nces in red, large differences in blue.
3.3. Comparing normal and fast reading tempi
When we compare the vowel space sizes for long
vowels in normal (left) and fast (right) reading tempo
(Fig. 5), we find quite different maps. This means that
change in tempo does not affect the vowel space size
of all regions in the same manner.
Figure 5: Vowel space size of the long vowel sys-
tem at normal (left) and fast (right) reading tempi.
Large vowel spaces in brown, small vowel spaces
in blue.
While the eastern part of Austria and the Ale-
mannic south have quite small vowel space areas in
both tempi, there are striking changes in the transition
zone from Bavarian to Alemannic and in the north-
eastern area of Mecklenburg-Vorpommern.
The results for the short vowel system (Fig. 6)
seem to be as disparate as for the long vowel system.
Here, the south is a bit more consistent than for the
long vowel system. For both normal and fast reading
tempi, the transition zone from Bavarian to Ale-
mannic and Franconian is separated from the neigh-
bouring areas. Moreover, in the north lower German
area and in the Ostfalen area, we find contrary distri-
butions for the two reading tempi.
Figure 6: Vowel space size of the short vowel sys-
tem at normal (left) and fast (right) reading tempi.
Large vowel spaces in brown, small spaces in blue.
Figs. 5 and 6 show that increasing reading tempo
does not universally result in the expected reduction
of the vowel space. In some regions there is even an
enlargement of the vowel space. However, com-
parison of the corresponding maps in Figs. 5 and 6 is
rather confusing. Therefore, the ratio of the vowel
space size of the fast reading tempo to the normal
reading tempo was calculated for the long vowel
system and the short vowel system. The result is
given in Fig. 7, showing in red the reduction of the
vowel space with an increase in reading tempo, and

in blue the expansion of the vowel space with an
increase in reading tempo.
Figure 7: Vowel space size change with increasing
tempo of the long vowel system (left) and the short
vowel system (right). Reduction of vowel spaces in
red, enlargements of vowel spaces in blue.
When we compare the changes for the long and
short vowel systems, the two pictures show a very
high correspondence. The areas where the vowel
space is enlarged when people speak faster are almost
identical. So, concerning the adjustments of vowel
space size, strategies for speaking faster are identical
for the long and the short systems.
When comparing the percent change of the vowel
space size with total vowel space size at normal rea-
ding tempo (the left-side maps in Figs. 6 and 7), we
get a quite surprising result. Fig. 8 shows a negative
correlation of vowel space size in normal tempo with
amount of reduction when reading faster. The regres-
sion models are highly significant (Long vowel sys-
tem (left): F(1,160) = 271.9169; p < 0.001; Short vo-
wel system (right): F(1,160) = 201.2398; p < 0.001).
So, we find that regions with a relatively large vowel
space in normal reading tempo show reduction of the
vowel space when reading faster. In regions where
people use a relatively small vowel space in normal
reading tempo, they enlarge it when reading faster.
Figure 8: Linear regressions, change of vowel
space size (%) in fast reading tempo by vowel space
size in normal reading tempo by place. Long vowel
system (left) and short vowel system (right).
As the vowel space size shows regional patterning,
we also find a related geolinguistic distribution of the
different strategies when speeding up the speaking
tempo (Fig. 7). The maps reflect how the strategies of
vowel space size change by increasing reading tempo.
The distribution of the data in the Standard-intended
reading task does not fit the traditional dialect areas,
but it shows a new regional distribution.
4. DISCUSSION AND CONCLUSION
Our results indicate that there are regional patterns of
vowel space size in Standard-intended German.
Because the patterns are consistent across both male
and female voices, the variation seems to not be
idiolectal but rather has a strong areal imprint.
Accelerating reading tempo does not only affect the
temporal domain but also vowel space size. Our
analyses likewise show regionally specific reduction
patterns for the vowel space.
The relationship between the vowel space sizes of
long and short vowels remain fairly stable between
reading tempi. Increasing the reading tempo has two
different effects: large vowel space sizes are reduced,
small vowel space sizes are enlarged.
The geographical distributions of vowel space size
in read speech and change due to reading tempo only
slightly reflect traditional dialect regions. Overall, the
geographical distribution is surprising as the vowel
space is based on the qualities of the vowels. How-
ever, the traditional maps show relatively large,
mostly phonologically motivated differences, whe-
reas there is a phonetic analysis here. While we do not
have the corresponding geolinguistic data on the vo-
wel space of the dialects to make a direct comparison,
we have to refer to other dialect classifications based
on the segmental phonetic level and on morphological
structures. Yet there are hardly any similarities to
these dialect classifications. The geographical
distribution of vowel space size may be interpreted as
having little correspondence to segmental phonetic
and morphological structures because the linguistic
system is independent of its phonetic realisation. This
difference would then have nothing to do with the
opposition of dialect and standard. One could also
argue that this distribution has more to do with the
regional distribution of speech rate or segment reduc-
tion than with dialect areas [2, 3]. But even these
relationships cannot be established unambiguously.
Thus, the design and change of the vocal space seems
to be a relatively independent parameter, probably
due to the fact that it is hardly perceived consciously.
Despite the apparent independence from traditio-
nal dialect classifications, reasonably stable geogra-
phical patterns emerge with regard to the design of
the vocal space. These show that theories of vocal
space [9, 12, 17, 18] must not only take into account
general aspects of influence such as speech rate, style,
and coarticulation, but that social and geographical
components must also be taken into account. This
means that the influence of speech rate on the shaping
of the vowel space is not a universal phenomenon, but
can vary even within one language.

5. REFERENCES
[1] Brinckmann, C., et al. 2008. German Today: An
areally extensive corpus of spoken Standard German.
Proc. LREC Marrakech. Paris: ELRA., 3185–3191.
[2] Hahn, M., & Siebenhaar, B. 2016. Sprechtempo und
Reduktion im Deutschen (SpuRD). In: Jokisch, O.
(ed), Elektronische Sprachsignalverarbeitung 2016.
Dresden: TUDpress, 198–205.
[3] Hahn, M., & Siebenhaar, B. 2019. Spatial Variation of
Articulation Rate and Phonetic Reduction in Standard-
Intended German. In Calhoun, S., Escudero, P.,
Tabain, M., & Warren, P. (eds.), Proceedings of the
19th International Congress of Phonetic Sciences.
Melbourne, Australia.
[4] Iivonen, A. 1994. Zur gehobenen regionalen phone-
tischen Realisierung des Deutschen. In Viereck, W.
(ed), Verhandlungen des Internationalen
Dialektologenkongresses Bamberg 29. 7. – 4. 8. 1990
vol 3. Stuttgart, 311–330.
[5] IPDS. 1994–1997. The Kiel Corpus of Read Speech
(vol. 1–4). Kiel: Institut für Phonetik und digitale
Sprachverarbeitung.
[6] Kisler, T., Reichel, U. D., & Schiel F. 2017.
Multilingual processing of speech via web services.
Computer Speech & Language 45, 326–347.
[7] Kleiner, S. 2011ff. Atlas zur Aussprache des
deutschen Gebrauchsstandards (AADG).
http://prowiki.ids-mannheim.de/bin/view/AADG/.
[8] König, W. 1989. Atlas zur Aussprache des
Schriftdeutschen in der Bundesrepublik Deutschland.
Ismaning: Hueber.
[9] Lindblom, B. 1963. Spectrographic study of vowel
reduction. J. Acoustical Soc. of America 35, 1773–
1781.
[10] Lobanov, B. M. 1971. Classification of Russian
Vowels Spoken by Different Speakes J. Acoustical
Soc. of America 49: 606–608.
[11] Moosmüller, S. 2007. Vowels in Standard Austrian
German. An Acoustic-Phonetic and Phonological
Analysis. Wien.
http://www.kfs.oeaw.ac.at/publications/
habil_2007may28_tableofcontents_zus_final.pdf
[12] Schulz, E., Oh, M. O., Andreeva, B., & Möbius, B.
2016. Impact of prosodic structure and information
density on vowel space size. Proceedings of Speech
Prosody 2016. Boston: ISCA, 350–354.
[13] Sendlmeier, W. F. & Seebode, J. 2006. Formant-
karten des deutschen Vokalsystems.
https://www.kw.tu-berlin. de/fileadmin/a01311100/
Formantkarten_des_deutschen_Vokalsystems_01.pdf
[14] Siebenhaar, B. 2014. Instrumentalphonetische Analy-
sen zur Ausgestaltung des Sprechlagenspektrums in
Leipzig. Zeitschrift für Dialektologie und Linguistik
81, 151–190.
[15] Simpson, A. P. 1998. Phonetische Datenbanken des
Deutschen in der empirischen Sprachforschung und
der phonologischen Theoriebildung. Kiel: AIPUK.
[16] Thomas, E. R. & Kendall, T. 2007–2015. NORM:
The vowel normalization and plotting suite. [Online
Resource:
http://lingtools.uoregon.edu/norm/norm1.php]
[17] van der Harst, S. 2011. The Vowel Space Paradox. A
Sociophonetic Study on Dutch. Utrecht: LOT.
[18] Weiss, B. 2008. Sprechtempoabhängige Aussprache-
variationen. Berlin. http://edoc.hu-
berlin.de/dissertationen/weiss-benjamin-2008-05-
28/PDF/weiss.pdf