Content uploaded by Christopher Carignan
Author content
All content in this area was uploaded by Christopher Carignan on Sep 27, 2019
Content may be subject to copyright.
THE PHONETIC BASIS OF PHONOLOGICAL VOWEL NASALITY:
EVIDENCE FROM REAL-TIME MRI VELUM MOVEMENT IN GERMAN
Carignan, C.1∗, Hoole, P.1, Kunay, E.1, Joseph, A.2, Voit, D.2, Frahm, J.2, Harrington, J.1
1Inst. of Phonetics and Speech Processing - LMU Munich, 2Max Planck Inst. for Biophysical Chem.
∗c.carignan@phonetik.uni-muenchen.de
ABSTRACT
It has been suggested that the development of con-
trastive vowel nasality in VN sequences may de-
pend partly on the nature of the following conso-
nant. In particular, there may be a preference for
VN sequences preceding voiceless oral consonants
to be phonologized due to aerodynamic constraints
on velum height, resulting in temporal overlap of
the vowel with a durationally constant velum ges-
ture. We investigate the phonetic basis of this claim
via direct imaging of velum kinematics in real-time
MRI videos (50 fps) from 35 German speakers. The
results show that, while the velum gesture does in-
deed begin and end earlier in /Vnt/ than in /Vnd/
sequences, the duration of the gesture itself is also
shorter in this context. This suggests that increased
temporal co-articulation in /Vnt/ sequences is not
necessarily due to durational maintenance of the
velum gesture, but to a temporally truncated velum
gesture that is shifted in time.
Keywords: Vowel nasalization, velum kinematics,
co-articulation, sound change, rtMRI, German.
1. INTRODUCTION
Co-articulatory vowel nasalization (i.e., [ ˜
V] in [ ˜
VN]
sequences) has been shown to exhibit systematic
temporal variation depending on a variety of pho-
netic contexts. In particular, a trading relation be-
tween the temporal extent of nasalization in the
vowel and the duration of the nasal consonant has
been observed to depend on the voicing of an oral
consonant that follows the VN sequence. For exam-
ple, nasal airflow studies of English have shown [5]
that there is greater co-articulatory vowel nasality (in
some cases, fully nasalized) combined with shorter
(in some cases, fully deleted) nasal consonants in
/Vnt/ vs. /Vnd/ sequences. [2] has proposed that this
trading relation may be a result of the interaction be-
tween aerodynamic constraints and a tendency for
maintenance of the duration of the velum gesture.
Voiceless obstruents require a sufficient build up of
intra-oral air pressure in order to produce the high-
airflow release that is necessary for the perception
of voicelessness [10]. The velum must close in or-
der to produce this pressure build up, an articulatory
requirement that is aerodynamically incompatable
with the requirement that the velum be open for the
production of a nasal consonant. If speakers main-
tain a roughly stable duration of the velum gesture, a
resolution to this aerodynamic incompatibility is for
the velum opening gesture to begin earlier (in the V)
and end earlier (in the N), resulting in a fully closed
velum during the following voiceless obstruent.
This particular phonetic pattern has important im-
plications for the diachronic development of con-
trastive vowel nasality. Typological evidence sug-
gests that VN sequences preceding voiceless oral
consonants may be predisposed to the development
of contrastive vowel nasality, in comparison with
those preceding voiced oral consonants [2]. Stud-
ies of sound change mostly in Romance languages
have shown that nasal consonants are preferentially
deleted and vowels preferentially nasalized before
voiceless obstruents [7, 11]. This typological asym-
metry may come about because listeners parse the
co-articulatory effect of nasalization with the vowel
rather than with the source (the nasal consonant) that
gives rise to it [2, 3].
Thus, accurate knowledge of the temporal pat-
terns involved in co-articulatory nasalization is
paramount to our understanding of how phonemic
vowel nasality can emerge diachronically. The re-
search presented here examines the prediction that
a greater degree of co-articulatory vowel nasaliza-
tion of, e.g., /Vnt/ compared to /Vnd/, is due to ear-
lier onset of a constant-sized nasal gesture. Here,
we test this hypothesis for phonetic vowel nasal-
ity in German using real-time magnetic resonance
imaging (rtMRI), which allows direct observation
of velum movement, rather than indirect measure-
ment of nasalization via its observed effect on air-
flow or the acoustic record (e.g., A1-P0, A1-P1 [4]).
Since there is no a priori reason to assume an on-
going process of phonologization of vowel nasality
in German, this study may help uncover basic pho-
netic mechanisms in contextual nasalization that can
413
explain the typological asymmetry discussed above.
2. METHODS
Real-time MRI, speakers, and stimuli
rtMRI data were collected at the Biomedizinische
NMR, Max Planck Insitute for Biophysical Chem-
istry in Göttingen, Germany, and reconstructed with
a temporal resolution of 20 ms (i.e., 50 fps) and an
in-plane spatial resolution of 1.4 mm [9, 12], along
with synchronized, noise-suppressed audio. Data for
35 native speakers of German are presented here.
The corpus consists of ≈300 German lexical items,
balanced for coda composition over a wide range of
phonetic contexts (e.g., vowel quality, stops vs. ob-
struents, etc.). A subset of this corpus is presented
here, consisting of minimal (or near-minimal) pairs
containing the tautosyllabic structure /Vnt/ or /Vnd/:
Table 1: Minimal and near-minimal pairs used.
Spelling Gloss IPA transcription
Bande ‘gang’ /band@/
bannte ‘averted’ /bant@/
Bunde ‘bunches’ /bUnd@/
bunte ‘colorful’ /bUnt@/
finde ‘find’ /fInd@/
Finte ‘trick’ /fInt@/
Panda ‘panda’ /panda/
Panther ‘panther’ /pant5/
Sande ‘sand(s)’ /zand@/
sandte ‘sent’ /zant@/
sende ‘send’ /zEnd@/
Senta ‘(woman’s name)’ /zEnta/
Sonde ‘probe’ /zOnd@/
sonnte ‘sunned’ /zOnt@/
winde ‘coil/wreathe’ /vInd@/
Winter ‘winter’ /vInt5/
During the MRI scanning sessions, the words ap-
peared on a computer screen, as reflected on a mir-
ror placed inside the scanner. The words appeared
in a variety of carrier phrases constructed to vary the
stress placement of the word in three primary condi-
tions: accentuated, de-accentuated, and neutral.
Velum movement signal
For each speaker’s data set, image registration was
carried out with reference to the superior portion
of the head, in order to correct for minor move-
ments of the head throughout the scanning session.
A velum opening/closing (henceforth “velum move-
ment”) signal was created from the registered im-
ages according to the following method. First, a re-
gion of interest (RoI) was manually selected around
the spatial range of velum opening/closing for each
speaker. The voxels (i.e., 3-D volume elements ob-
tained from the MRI scan) in the RoI were then ex-
tracted for the images pertaining to words containing
VN sequences. The voxel intensities were used as
dimensions in principal components analysis (PCA)
models, and the scores from the first PC (PC1) were
logged for each image frame, resulting in a time-
varying signal. Since there is only one primary de-
gree of freedom in the movement of the velum (i.e.,
opening/closing) in VN sequences, PC1 will relate
to this dimension of movement in every case. An
example of the PC1 loadings/coefficients for one of
the speakers is shown in Figure 1. The positive load-
ings (bright voxels) and negative loadings (dark vox-
els) are associated with the velum in its closed and
opened states, respectively, revealing that the feature
captured by PC1 is indeed velum opening/closing.
The time-varying signal derived from this method
can be interpreted as the magnitude of velum open-
ing: smaller values represent a more closed velum,
while larger values represent a more open velum.
Figure 1: An example of the region-of-interest
(RoI) based principal components analysis for
generating a time-varying velum opening/closing
signal. PC1 loadings are denoted by light (posi-
tive) and dark (negative) voxels within the RoI.
Measurements used
Using this velum movement signal, several measure-
ments were derived from key time points occurring
in the VN segment of each token. Firstly, the onset
and offset of the velum gesture was determined by
20% thresholds of the peak positive velocity (corre-
414
sponding to the gesture onset) and the peak negative
velocity (corresponding to the gesture offset) of the
velum movement signal, in the same manner as for
kinematic signals generated by electromagnetic ar-
ticulometry. The duration of the velum gesture is
therefore defined as the temporal distance between
these two time points. Secondly, the vowel offset—
representing the point of transition between the V
and the N in the VN sequence—was identified man-
ually in the acoustic signal. This time point was
used to create articulatory/acoustic hybrid measure-
ments for the timing of the onset and the offset of the
gesture for each token, which results in more stable
measurements (i.e., less variance) than using the raw
gesture onset/offset measurements themselves. The
time points for the gesture onset and offset are thus
defined with reference to the acoustic vowel offset
(e.g., offset = gesture offset - vowel offset), although
the interpretations for the values remain the same:
smaller values represent earlier time points, while
larger values represent later time points.
In addition to these three temporal measurements,
a spatial measurement was also created to charac-
terize the magnitude of velum opening. This mea-
surement is defined simply as the value of the velum
movement signal at the time point of the trajec-
tory peak (i.e., the maximum degree of nasalization).
This measurement was also modified in order to cre-
ate a more stable measure and to more accurately
capture difference in the relative magnitude within
each token: the value at the onset of the gesture (i.e.,
a baseline for each token) was subtracted from the
value at the point of maximum constriction. The in-
terpretations for these baseline-compensated values
remain the same: smaller values represent a smaller
degree of velum opening, while larger values repre-
sent a larger degree of velum opening.
Statistical validation
Linear mixed-effects (LME) models were created
in R using the lmer function in the lme4 package
[1]. Estimates for F-statistics and corresponding
p-values were generated using the lmerTest pack-
age [8]. The three temporal measurements and one
spatial measurement were speaker-normalized via z-
score transformation before inclusion in the models.
For each model, fixed effects included the VOICING
of the coda oral consonant (/Vnd/, /Vnt) and STR ES S
(accentuated, de-accentuated, neutral), and full ran-
dom effects were included for SP EA KER and WOR D.
For the purposes of this study, “word” is defined as
the phonetic segments up to and including the vowel,
but excluding the coda, since the coda context is in-
herently part of the fixed effect VOI CI NG.
3. RESULTS
An example of the velum movement trajectories for
the minimal pair Panda-Panther is shown in Figure
2, averaged over all 35 speakers. In this figure, the
trajectories have been time-aligned with the (acous-
tic) vowel offset, which is denoted by the middle set
of symbols (circles and squares). The left set of sym-
bols denote the vowel onset (as determined by the
acoustics), and the right set of symbols denote the
offset of the coda consonants /t/ or /d/ (as determined
by the acoustics). Voicing of the oral coda conso-
nant is denoted by line and symbol (/Vnd/ = solid
line +circles, /Vnt/ = dotted line +squares). Stress
is denoted by color (accentuated = blue, neutral =
red). From this figure, it appears that the velum ges-
ture for /Vnt/ is shorter in duration and begins and
ends earlier compared to /Vnd/. Additionally, for
both words, the accentuated stress condition results
in a larger gestural magnitude (i.e., the blue lines are
higher than the red lines).
Figure 2: Ensemble averages (over all 35 speak-
ers) of velum movement signals for the mini-
mal pair Panda-Panther. The gesture trajecto-
ries are time-aligned with respect to the acoustic
vowel offset (Time = 0). Smaller values indicate a
smaller degree of velum opening and larger values
indicate a larger degree of velum opening.
-0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2
Time (s); lineup: Vowel offset
-1
-0.5
0
0.5
1
1.5
2
mean normalized velum opening
/panda/ vs. /pantɐ/
panda_A
panda_N
pantɐ_A
pantɐ_N
By way of comparison, Figure 3 displays velum
trajectories for the minimal pair sende-Senta. Al-
though the same patterns can be observed for the
gestural magnitude and duration, no differences in
the timing of the onset of the gesture can be seen. In
other words, while the onset of the velum gesture oc-
curred earlier in /Vnt/ vs. /Vnd/ for Panda-Panther,
the same difference cannot be seen for sende-Senta,
although the same reduction in the duration of the
gesture is evidenced. This suggests that, rather than
a temporal shifting of the velum gesture, the gesture
is instead temporally truncated in /Vnt/ vs. /Vnd/.
415
Figure 3: Ensemble averages (over all 35 speak-
ers) of velum movement signals for the minimal
pair sende-Senta. The gesture trajectories are
time-aligned with respect to the acoustic vowel
offset (Time = 0). Smaller values indicate a
smaller degree of velum opening and larger val-
ues indicate a larger degree of velum opening.
-0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2
Time (s); lineup: Vowel offset
-1
-0.5
0
0.5
1
1.5
2
mean normalized velum opening
/zɛndə/ vs. /zɛnta/
zɛndə_A
zɛndə_N
zɛnta_A
zɛnta_N
In order to observe the effects for all of the words
combined, Table 2 displays the results for the LME
models. With regard to voicing, there are signifi-
cant effects for all four measurements: gesture dura-
tion, onset, offset, and magnitude. In other words, in
/Vnt/ sequences compared to /Vnd/ sequences: the
velum gesture is shorter, begins earlier, ends earlier,
and has a smaller magnitude (i.e., less velum open-
ing). With regard to stress, no significant effect is
observed for the gesture onset, and a marginally sig-
nificant effect is observed for the gesture duration;
however, this marginal effect is most likely a con-
sequence of the large effect that stress has on the
timing of the gesture offset: the velum gesture ends
sooner in de-accentuated and neutral conditions than
in accentuated condition. Moreover, stress has a sig-
nificant effect on velum magnitude: there is a greater
degree of nasalization in accentuated condition.
4. CONCLUSION
The results from this study reveal that the velum ges-
ture begins and ends earlier in /Vnt/ vs. /Vnd/ se-
quences in German, as predicted by [2, 3]. How-
ever, it is not the case that the duration of the gesture
was maintained in these data: the velum gesture was
shorter in /Vnt/ vs. /Vnd/ sequences. Although the
onset of the velum gesture began earlier in /Vnt/ se-
quences, the temporal shift is not as great as for the
gesture offset: on average, there is a 13 ms tempo-
ral difference between /Vnt/ and /Vnd/ at the gesture
onset, but a 28.4 ms difference at the offset. More-
over, /Vnt/ sequences were found to be less nasal-
Table 2: Results for LME models constructed to
test the effect of coda voicing (/Vnt/, /Vnd/) and
stress (accentuated, deaccentuated, neutral) on the
overall duration, onset timing, offset timing, and
peak magnitude of the velum gesture.
DV Effect F-stat. Pr(>|F|)
Dur. Voicing 76.15 p<0.001 ***
Stress 3.72 p<0.050 *
Ons. Voicing 85.06 p<0.001 ***
Stress 3.09 p=0.086
Off. Voicing 221.84 p<0.001 ***
Stress 42.59 p<0.001 ***
Mag. Voicing 42.81 p<0.001 ***
Stress 30.87 p<0.001 ***
ized than /Vnd/ sequences. These results suggest
that increased temporal co-articulation in /Vnt/ se-
quences in German is not due to durational mainte-
nance of the velum gesture. Rather, the velum ges-
ture is temporally truncated and, in some cases (e.g.,
Panda-Panther but not sende-Senta), this truncated
gesture is also shifted in time, resulting in increased
co-articulatory vowel nasalization.
In these cases, the phonetic bias is the same as
predicted by [2, 3]: /Vnt/ sequences involve greater
nasal co-articulation, which can give rise to con-
trastive vowel nasality if listeners parse the effect of
nasalization with the vowel. Instead of this phonetic
bias emerging from the maintenance of a constant-
sized velum gesture, it may be the case that contex-
tual vowel nasalization involves a stage of tempo-
ral truncation and reduced magnitude of the gesture
when preceding voiceless obstruents, as a precursor
to the temporal shift of the gesture onto the vowel.
Finally, the results for accentuation are inter-
esting because previous investigations of prosodic
strengthening with respect to velum activity have
given conflicting results [6]. However, since even
the extensive material presented here is actually only
part of that available on prosodic contrasts in our
rtMRI corpus we will consider this issue in more de-
tail in a separate publication.
ACKNOWLEDGMENTS
This research was funded by ERC Advanced Grant
295573 Human interaction and the evolution of
spoken accent (J. Harrington) and DFG grant HA
3512/15-1 Nasal coarticulation and sound change:
a real-time MRI study (J. Harrington & J. Frahm).
416
5. REFERENCES
[1] Bates, D., Mächler, M., Bolker, B., Walker, S.
2015. Fitting Linear Mixed-Effects Models Using
lme4. Journal of Statistical Software 67(1), 1–48.
[2] Beddor, P. S. 2009. A coarticulatory path to sound
change. Language 85(4), 785–821.
[3] Beddor, P. S. 2012. Perception grammars and
sound change. In: Solé, M.-J., Recasens, D.,
(eds), The Initiation of Sound Change. Perception,
Production, and Social Factors. Amsterdam: John
Benjamins 37–55.
[4] Chen, M. Y. 1997. Acoustic correlates of English
and French nasalized vowels. Journal of the Acous-
tical Society of America 102, 2360–2370.
[5] Cohn, A. C. 1990. Phonetic and Phonological
Rules of Nasalization. PhD thesis University of
California, Los Angeles. Published as UCLA Work-
ing Papers in Linguistics 76.
[6] Fougeron, C. 2001. Articulatory properties of ini-
tial segments in several prosodic constituents in
French. Journal of Phonetics 29, 109–135.
[7] Hajek, J. 1997. Universals of sound change in
nasalization. Oxford: Blackwell.
[8] Kuznetsova, A., Brockhoff, P. B., Chris-
tensen, R. H. B. 2016. lmerTest: Tests in
Linear Mixed Effects Models. Computer soft-
ware program available from https://cran.r-
project.org/package=lmerTest.
[9] Niebergall, A., Zhang, S., Kunay, E., Keydana, G.,
Job, M., Uecker, M., Frahm, J. 2012. Realtime
MRI of speaking at a resolution of 33 ms: Under-
sampled radial FLASH with nonlinear inverse re-
construction. Magnetic Resonance in Medicine 69,
477–485.
[10] Ohala, J. J., Ohala, M. 1993. The phonetics of
nasal phonology: Theorems and data. In: Huffman,
M. K., Krakow, R. A., (eds), Nasals, Nasalization,
and the Velum volume 5 of Phonetics and Phonol-
ogy. San Diego: Academic Press 225–249.
[11] Sampson, R. 1999. Nasal vowel evolution in Ro-
mance. Oxford: Oxford University Press.
[12] Uecker, M., Zhang, S., Voit, D., Karaus, A., Mer-
boldt, K., Frahm, J. 2010. Real-time mri at a reso-
lution of 20 ms. NMR Biomed. 23, 986–994.
417
