Content uploaded by Ander Beristain
Author content
All content in this area was uploaded by Ander Beristain on Jun 17, 2020
Content may be subject to copyright.
Lenition of word-final plosives in Basque
José Ignacio Hualde, Ander Beristain, Ane Icardo Isasa, Jennifer Zhang
University of Illinois at Urbana-Champaign
{jihualde | anderb2 | icardoi2 | jz13}@illinois.edu
i
ABSTRACT
Basque has a phonological contrast between voiceless
/ptk/ and voiced /bdg/ in onset position. Word-finally
there is only /t/ and /k/. These word-final consonants,
although rare stem-finally, have great textual
frequency, since they are found in several frequent
inflectional suffixes. We examine the realization of
final /t k/ before a vowel across word boundaries,
comparing them with word-medial and word-initial
intervocalic consonants. Based on a corpus of natural
speech, we test the hypothesis that prevocalic word-
final plosives are weaker than other intervocalic
plosives, since they do not contrast with voiced
phonemes in this position. We test two methodologies
for quantifying plosive lenition based on differences
in intensity that have been proposed in prior work.
Both methodologies returned very similar, but not
identical, results. Intensity results are consistent with
the hypothesis, especially for velars. /t k/ are also
more voiced word-finally than elsewhere.
Keywords: Basque, lenition, contextual
neutralization, plosives
1. INTRODUCTION
We test the effects of phonological contrast on
lenition by comparing the realization of intervocalic
voiceless stops in Basque across morphological
contexts.
Basque has a phonological contrast between
voiceless and voiced plosive phonemes in onset
position, both word-initially and word-medially (e.g.
word-initial: puru ‘pure’, buru ‘head’; post-
consonantal: arto ‘corn’, ardo ‘wine’; word-medial
intervocalic: ekin ‘undertake’, egin ‘do’). Word-
finally, however, only /k/ and /t/ are allowed. These
two word-final consonants have very low lexical
incidence, but high textual frequency. There are no
major-category words that end with a plosive, but
several inflectional suffixes end in /t/ or /k/,
including, for /t/, the benefactive suffix (e.g. Peru-
rentzat ‘for Peru’) and the first person singular
transitive subject agreement marker (daki-t ‘I
know’). Suffixes ending in /k/ include the ergative,
which marks the subject of a transitive verb (Peru-k
‘Peru, erg.’), the absolutive and ergative plural
(mendi-ak ‘the mountains’), the partitive (mendi-rik
‘mountain, part.’), the ablative (mendi-tik ‘from the
mountain’) and the second person singular masculine
familiar transitive subject agreement marker (daki-k
‘thou, male, knowest’). In addition, there are several
function words ending in /t/, including the
numeral/indefinite article bat ‘one, a’.
Voiceless intervocalic stops have been reported to
voice and weaken to approximants, in at least some
Basque varieties [10, 11, 23].
In this paper, we examine the realization of word-
final plosives before a vowel in a corpus of natural
speech, comparing them with intervocalic plosives in
other positions. Since only /t/ and /k/ are possible
plosives word-finally, we limit our comparison to
these two consonants.
Our hypothesis is that word final plosives before a
vowel (VC#V) will show greater voicing lenition than
plosives in either word-internal intervocalic position
(VCV) or word-initially after a vowel (V#CV), given
the fact that, word-finally, /t k/ do not contrast with /b
g/. We also predict velar /k/ to undergo greater
lenition than dental /t/, based on the fact that word-
internal intervocalic velar /g/ tends be deleted in
many Basque varieties [9, 10]. Intervocalic voiceless
velars have also been found to show more lenition
than other consonants in Iberian Spanish [13].
Besides examining the effects of morphological
context and phonological contrast on intervocalic
lenition in Basque, we also have methodological
goals in this paper. We compare two methods that
have been used in recent work to quantify lenition
based on the intensity curve. We would like to
determine whether both methodologies produce
similar results when applied to the same data.
2. METHODS
Our data come from recorded interviews on everyday
topics with 6 native Basque speakers from Azpeitia,
Gipuzkoa (4 female, 2 male). Among all towns in the
Basque Country with over 5000 people, Azpeitia has
the highest proportion of Basque speakers (82%).
Basque is by far the most commonly used language
within the town. The interviewer was also a native
speaker of Azpeitia Basque. All participants signed a
consent form before being recorded. The
conversations took place in Azpeitia, in a quiet place
familiar to the participants, and were recorded with a
MicroTrack 24/96 digital recorder using a SONY F-
720 external microphone.
642
The recordings were analyzed in Praat [2]. From
these recordings we extracted 2482 tokens of /t/ and
/k/ in the following contexts: (a) word-final before a
vowel (VC#V), (b) word-medial intervocalic (VCV),
(c) word-initial following a vowel (V#CV) and (d)
word-final before another consonant (VC#C). For this
study, however, we are excluding the word-final
preconsonantal tokens, as we are focusing on the
realization of intervocalic consonants. Only tokens
produced without a pause between words have been
included.
After removing pre-consonantal tokens and some
other tokens for the additional reasons explained
below, a resulting total of 2252 intervocalic tokens
were analyzed for this paper, with the distribution of
contexts shown in Table 1.
Table 1. Number of analyzed tokens by context
Final
Initial
Medial
k
180
75
821
t
83
126
967
To quantify lenition, we took measurements of
intensity between the consonant and the following
vowel, as a correlate of the degree of constriction. The
less constricted the consonant, the smaller the
difference in intensity is expected to be.
Intensity was measured in two different ways.
First, we calculated the difference between the
intensity maximum and minimum in the CV sequence
(IntDiff), using a similar methodology as in other
work on intervocalic lenition [3, 4, 5, 6, 13, 15, 17,
18,19, 22]. To take this measurement, an interval tier
was created in Praat, manually placing boundaries
around the target consonant, taking care that the
selected intervals contained the minimum intensity in
the consonant and the maximum in the following
vowel. While delineating the boundaries of lenited
consonant can be difficult, exact segmentation is not
required for this measurement. The only requirement
is for the interval to include the intensity minimum
within the consonant and the intensity maximum
during the following vowel. Some, but not all, of the
studies mentioned above have used high and/or low-
pass filters in order to remove energy from voicing at
low frequencies and possibly from background noise
at high frequencies. For instance, in [22] energy was
measured applying a Hann band-pass filter between
250 Hz and 10 kHz. Here, we do not apply a filter and
compare the results with the those obtained by
applying the methodology described in [7], where
changes in energy are taken from a band between 400
Hz and 1200 Hz.
Thus, we used the R script by [7], which focuses
on changes in intensity from the beginning of the
consonant. To run this script, we created a separate
point tier, placing a single boundary near the
beginning of each target consonant. Edges of closing
and opening gestures are automatically determined
from changes in intensity velocity, and segment
boundaries are placed in relation to these points. The
script returns Delta-i (Δi) values, which reflect the
magnitude of change in intensity within the defined
segment. Tokens where there was no visible dip in
intensity or other evidence for a consonantal gesture
in the spectrogram were discarded. We take these as
instances of consonant deletion. To make the two
analyses fully compatible, we removed those tokens
from the computation of IntDiff as well.
We consider to what extent both methodologies
produce similar results. As noted, neither method
requires accurate manual placement of segment
boundaries. We will use the term Delta-i to refer to
the results obtained with the script by [7] and IntDiff
to refer to the measurement that calculates the
difference between intensity maximum and minimum
within the CV sequence, without applying a filter.
A second potential dimension in the lenition of
voiceless stops, besides a decrease in the degree of
constriction, is voicing. To analyze voicing
separately, we used the intervals created for the
IntDiff measurement, classifying each token as either
fully voiced or not. The voice report in Praat returns
the percentage of frames in an interval that are
produced as voiced, but since our intervals include
not only the target consonant but also part of the
following vowel, we are treating voicing as a binary
feature. We consider a consonant to be voiced only if
there is voicing throughout the entire interval
(unvoiced frames in the selected interval = 0 in
Praat’s voice report). This avoids the need to place
exact boundaries at the beginning and end of the
consonant. A justification for our binary analysis is
that only fully voiced tokens can be said to show
potential neutralization between underlyingly voiced
and voiceless phonemes [13].
For the statistical analysis of the intensity results,
we ran linear mixed effects regressions on IntDiff and
Delta-i in R [20] and RStudio [21] with the function
lmer in the package lme4 [1]. P-values were obtained
with the Satterthwaite approximation in lmerTest
[14]. Voicing was analyzed with the function glmer
(family= binomial) in lme4.
Details regarding the structure of the regressions
are given in the results section.
3. RESULTS
3.1. IntDiff
The boxplots in Figure 1 show the results of our
IntDiff measurement by context (made with the
package ggplot2 [25]). From visual inspection of the
plots, word-final tokens are generally more lenited
643
(smaller IntDiff) than other tokens. For /k/, word-
initial tokens appear to be strongest.
Figure 1: IntDiff for /t/ and /k/ by context: fv =
word-final prevocalic, i =word initial after a vowel,
m = word-medial intervocalic
We ran a linear mixed effects regression on IntDiff
with Context (three levels: i= V#CV, m=VCV and fv
=VC#V), Consonant (two levels: /t/, /k/) and their
interaction as fixed factors, and Speaker and Word as
random factors.
The output of the regression shows a significant
difference between word-final prevocalic and both
word-initial postvocalic (β = 13.7, t= 10.4, p < 0.001)
and word-medial intervocalic tokens (β = 5, t= 6.1, p
< 0.001), as well between /t/ and /k/, with the velar
being more lenited (β = 3.6, t= 2.3, p = 0.02).
Post-hoc comparisons (with the package
emmeans [16]) return significant differences among
all three contexts for both consonants, with the initial
position being strongest (largest IntDiff) and the final
position being weakest, see Table 2.
Table 2: Post-hoc comparisons, IntDiff
/k/
Estimate
t.ratio
p.value
fv-i
-13.7
-10.4
<0.0001
fv-m
-5
-6.1
<0.0001
i-m
8.7
7.7
<0.0001
/t/
fv-i
-10
-4.9
<0.0001
fv-m
-4.1
-3
=0.0076
i-m
5.9
3.7
=0.0007
Regarding our research question, we find that the
target intervocalic consonants are more lenited when
they are word-final than when they are word-initial or
word-medial. The differences between contexts are
greater for /k/ and there is a clear hierarchy of strength
for this consonant among the three positions
examined.
3.2. Delta-i
The boxplots in Figure 2 show differences in intensity
calculated with the methodology of [7]. The results
for /k/ are very similar to those in Figure 1, with a
clear effect of position on degree of constriction:
initial > medial > final. For /t/, on the other hand, the
differences between contexts are less clear in Figure
2 (for ease of comparison, the Delta-i output has been
multiplied by -1).
Figure 2: Delta-i for /t/ and /k/ by context: fv = word-
final prevocalic, i = word initial after a vowel, m =
word-medial intervocalic
Delta-i values were modeled as the dependent value
in an lmer with the same structure as the one for the
IntDiff results. Significant effects were found for
both fixed factors. Regarding context, word-final is
significantly weaker than word-initial (β = 10, t= -
8.3, p < 0.001) and word-medial (β = 3.8, t= 4.9, p <
0.001). There is also a significant difference between
/t/ and /k/, with the velar being weaker (β = 7, t= 4.9,
p =0.001). An interaction was also found between
Consonant and the comparison between the final and
initial positions (β = -4.8, t= -2.1, p =0.032).
In post-hoc comparisons with emmeans, all three
place comparisons are significant for /k/ (p <
0.0001), but for /t/, only final vs initial and initial vs
medial contexts approach significance at the p <0.01
level (see Table 3).
Table 3: Posthoc comparisons, Delta-i
/k/
Estimate
t.ratio
p.value
fc-i
-10
8.3
<0.0001
fv-m
-3.8
-4.9
<0.0001
i-m
6.3
6.1
<0.0001
/t/
Estimate
t.ratio
p.value
fv-i
-5.2
-2.2
=0.016
fv-m
-0.9
-0.7
=0.8 n.s.
i-m
4.3
2.9
=0.0103
644
3.3. Voicing
Regarding voicing, word-final prevocalic consonants
also show a greater propensity to be realized as fully
voiced than other tokens in intervocalic position. On
the other hand, word-initial postvocalic consonants
are almost never fully voiced (see Table 4). Both
medially and finally, /k/ voices more often than /t/.
Table 4: Percent of fully voiced tokens of
intervocalic /t/ and /k/ by morphological context
Context
/t/
/k/
Final VC#V
13/83 (15.7%)
58/180 (32.2%)
Initial V#CV
7/127 (5.5%)
3/75 (4%)
Medial VCV
85/976 (8.7%)
146/821(17.8%)
A binomial mixed-effects logistic regression was
fit to the voicing data with the same fixed effects
structure as for the intensity data and with Speaker as
a random factor. (A model including Word as an
additional random factor failed to converge). Our
statistical model returned a significant effect of
consonant, where /k/ is significantly more likely to
be realized as fully voiced than /t/ (β = -1.0279, z= -
2.588, p < 0.01). Intervocalic /t/ and /k/ are voiced
with significantly higher frequency when word-final
than when word-initial (β = -2.6626, z= -4.161, p <
0.001) and word-medial (β = -1.0594, z= -4.665, p <
0.001). No significant interaction between consonant
and context was found.
Although we chose to focus on numbers of fully
voiced tokens in each context for the reasons
mentioned above, an analysis based on ‘fractions of
locally unvoiced frames’ in each token returns fully
comparable results, with more voicing for /k/ than for
/t/ and the same hierarchy among contexts.
3.4. Comparison of the results of methods of
quantifying lenition
The two intensity measures that we have employed
show a relatively strong correlation, r = 0.7087. Our
IntDiff and our Delta-i measurements produced more
similar results for /k/, as is clear from comparing the
left-hand panels of Figures 1 and 2 (for /k/ only, r
= -0.7454). Regressions using these two
measurements as the dependent variable returned a
significant effect of context, with the initial context
being the strongest and the final context being the
most lenited. We thus conclude that there is a strong
effect of word position on the lenition of /k/ that is
captured by both measurements.
For /t/, on the other hand, the measurement that is
used matters. Whereas the regression on IntDiff
returned a significant difference between final tokens
of /t/ and tokens in other positions (as well as initial
vs final), the regression on Delta-i found much
smaller differences.
The explanation for the difference in the results
of applying the two methodologies may be that Delta-
i excludes energy at low frequency ranges, where the
voice bar is found, and IntDiff does not (since
intervocalic /t/ is most often realized as fully voiced
when word-finally and least frequently when word-
initial). To answer this question more directly, a voice
report could be obtained for the intervals created to
calculate Delta-i by adding this functionality to the
script.
As mentioned for /k/, on the other hand, the
correlation between IntDiff and Delta-i is somewhat
higher and the statistical results are very similar. We
may suspect that for /k/, voicing and the reduction of
the oral gesture tend to go hand in hand more
frequently than for /t/. This difference between the
lenition of /k/ and /t/ is consistent with what we know
about the allophony of their voiced counterparts in
Basque. Whereas intervocalic /g/ tends to delete in
Basque dialects, /d/ is very frequently realized as a
flap, neutralizing with phonemic /ɾ/ (e.g. bide ~ bire
‘path’) [9], including in the variety of Azpeitia [8].
Dentals and velars thus show somewhat different
paths of reduction.
The voicing of a phonological stop is obviously
independent from the magnitude of the oral
articulatory gesture. From a phonological point of
view, on the other hand, it may make sense to
consider these two articulatory dimensions together
as part of a single process of lenition, with differences
among places of articulation perhaps to be explained
by articulatory factors.
4. CONCLUSIONS AND DISCUSSION
All consonant tokens that we have analysed are found
in the same phonetic context: between two vowels.
There are, however, morphological differences
among them related to word boundaries, which have
an effect on the realization of these consonants. The
VC#V context conditions more lenited consonants
than other contexts. In this context, consonants are
less constricted (especially for /k/) and are more
frequently fully voiced. Since there is no
phonological contrast between voiced and voiceless
consonants word-finally in Basque, these results can
be interpreted as showing an effect of phonological
contrast on phonetic realization. On the other hand,
this hypothesis does not explain the finding that
word-medial tokens are also more lenited than word-
initial intervocalic ones. The explanation may be
found in differences in the pattern of gestural
coordination of intervocalic consonants depending on
position in the word (see [12, 24]). The fact that final
stops are usually in suffixes may also be relevant.
645
7. REFERENCES
[1] Bates, D., Maechler, M., Bolker, B., Walker, S. 2015.
Fitting linear mixed-effects models using lme4.
Journal of Statistical Software 67(1), 1-48.
[2] Boersma, P., Weenink, D. 2018. Praat: doing phonetics
by computer, version 6.035, www.praat.org [computer
program].
[3] Carrasco, P., Hualde, J. I., Simonet, M. 2012.
Dialectal differences in Spanish voiced obstruent
allophony: Costa Rican vs. Iberian Spanish.
Phonetica, 69, 149–179.
[4] Colantoni, L., Marinescu, I. 2010. In: Ortega-Llebaria,
M. (ed.), The scope of stop weakening in Argentine
Spanish. Selected Proceedings of the 4th Conference on
Laboratory Approaches to Spanish Phonology.
Somerville:MACascadilla Proceedings Project, 100–
114.
[5] Cole, J., Hualde, J. I., Iskarous, K. 1999. In: Fujimura,
O., Joseph, B. D. and Palek, B. (eds.), Effects of
prosodic and segmental context on/g/-lenition in
Spanish. Proceedings of the 4th international
linguistics and phonetics conference. Prague: The
Karolinium Press 2: 575–589.
[6] Eager, C. 2017. Contrast preservation and constraints
on individual phonetic variation. Doctoral dissertation.
Urbana-Champaign, IL: University of Illinois at
Urbana-Champaign
[7] Ennever, T., Meakins, F., Round, E. 2017. A replicable
acoustic measure of lenition and the nature of
variability in Gurindji stops. Laboratory Phonology:
Journal of the Association for Laboratory Phonology,
8(1), 20. DOI: http://doi.org/10.5334/labphon.18
[8] Hualde, J. I., Beristain, A. 2019. “Bi alofonia gertakari
Azpeitiko euskaran”. In: Festchrift [to appear].
[9] Hualde, J.I., Elordieta, G., Elordieta, A. 1994. The
Basque dialect of Lekeitio (Supplements of ASJU).
Bilbao/Donostia: Univ.del País Vasco – Euskal Herriko
Unib./ Diputación de Gipuzkoa – Gipuzkoako Foru
Aldundia.
[10] Hualde, J.I., Lujanbio, O., Zubiri, J.J. 2010. Goizueta
Basque. Journal of the International Phonetic
Association, 40, 113–127.
[11] Hualde, J. I., Nadeu, M. 2011. Lenition and phonemic
overlap in Rome Italian. Phonetica 68(4), 215–242,
DOI: https://doi.org/10.1159/000334303
[12] Hualde, J. I., Prieto, P. (2014). “Lenition of sibilants
in Catalan and Spanish”. Phonetica 71, 109–127, DOI:
https://doi.org/10.1159/00036819
[13] Hualde, J. I., Simonet, M., Nadeu, M. 2011.
Consonant lenition and phonological recategorisation.
Laboratory phonology 2(2): 301–329, DOI:
https://doi.org/10.1515/labphon.2011.011
[14] Kuznetsova A., Brockhoff, P.B., Christensen R.H.B.
2017. lmerTest Package: Tests in Linear Mixed
Effects Models. Journal of Statistical Software,
82(13): 1-26. doi: 10.18637/jss.v082.i13
[15] Lavoie, L. M. 2001. Consonant strength: Phonological
patterns and phonetic manifestations. New York:
Routledge.
[16] Lenth, R. 2018. emmeans: Estimated Marginal Means,
aka Least-Squares Means. R package
version 1.2.3.
https://CRAN.R-project.org/package=emmeans
[17] Nadeu, M., Hualde, J.I. 2015. Biomechanically
conditioned variation at the origin of diachronic
intervocalic voicing. Language and Speech 58(3): 351–
370
[18] Lewis, A. 2001. Weakening of Intervocalic /p, t, k/ in
Two Spanish Dialects: Towards the Quantification of
Lenition Processes. Doctoral dissertation. Urbana-
Champaign, IL: University of Illinois at Urbana-
Champaign
[19] Ortega-Llebaria, M. 2004. Interplay between phonetic
and inventory constraints in the degree of spirantization
of voiced stops: Comparing intervocalic /b/ and
intervocalic /g/ in Spanish and English. In: Face, T. L.
(ed.), Laboratory approaches to Spanish phonology.
Berlin: Mouton de Gruyter, 237–253.
[20] R Core Team. 2017. R: A language and environment
for statistical computing. R Foundation for Statistical
Computing, Vienna, Austria. https://www.R-
project.org/.
[21] RStudio Team. 2016. RStudio: Integrated
Development for R. RStudio, Inc., Boston, MA URL
http://www.rstudio.com/.
[22] Recasens, D. 2017. The effect of contextual
consonants on voiced stop lenition: Evidence from
Catalan. Language and Speech 59(1): 139-161.
[23] Salaburu, P. 1984. Hizkuntzaren soinu-egitura.
Hizkuntz teoria eta Baztango euskalkia. Leioa, Bizkaia,
Spain: Univ.del País Vasco-Euskal Herriko Unib.
[24] Strycharczuk, P., Kohlberger, M. (2016).
“Resyllabification Reconsidered: On the Durational
Properties of Word-Final /s/ in Spanish”. Laboratory
Phonology: Journal of the Association for Laboratory
Phonology, 7(1), 3.
DOI: http://doi.org/10.5334/labphon.5
[25] Wickham, H. 2016. ggplot2: Elegant Graphics for
Data Analysis. Springer-Verlag New York.
646