A new look at Pattani Malay Initial Geminates: a statistical and machine learning approach

Abstract

In this paper, we present a statistical and machine learning approach to the acoustic discrimination of a cross-linguistically unusual phonological contrast, initial geminates vs. singletons in Pattani Malay. We show that the only statistically significant difference between geminates and singletons is the duration of the consonant itself. No differences in F0 and intensity were observed on the following vowel, contra earlier reports. We further investigated the robustness of this contrast using linear discriminant analysis. Results show that discrimination is above chance, but poor (~62%). The large overlap between the two categories may be partly due to the naturalistic nature of our speech samples. However, we also found that the contrast is neutralized in some minimal pairs. This merger is surprising since initial geminates are often the sole realization of lexical and morphosyntactic contrasts. We suggest that the singleton/initial geminate contrast is now best characterized as a marginal contrast. We hypothesize that this marginally contrastive status may be the result of an ongoing sound change, perhaps connected with the more modest role that initial geminates play in Pattani Malay morphophonological alternations.

Full text

A new look at Pattani Malay Initial Geminates:
a statistical and machine learning approach
Francesco Burroni
Sireemas Maspong
Department of Linguistics and
Department of Linguistics and
Cognitive Science Program
Southeast Asia Program
Cornell University
Cornell University
203 Morrill Hall,
203 Morrill Hall,
Ithaca, NY, USA 14850
Ithaca, NY, USA 14850
fb279@cornell.edu
sm2627@cornell.edu
Pittayawat Pittayaporn
Pimthip Kochaiyaphum
Department of Linguistics and
Department of Linguistics and
Southeast Asian Linguistics Research Unit
Southeast Asian Linguistics Research Unit
Faculty of Arts, Chulalongkorn University
Faculty of Arts, Chulalongkorn University
Phayathai Road, Pathumwan,
Phayathai Road, Pathumwan,
Bangkok, Thailand 10330
Bangkok, Thailand 10330
pittayawat.p@chula.ac.th
pimthip.ko@student.ac.th
Abstract
In this paper, we present a statistical and
machine learning approach to the acoustic
discrimination of a cross-linguistically unusual
phonological contrast, initial geminates vs.
singletons in Pattani Malay. We show that the
only statistically significant difference between
geminates and singletons is the duration of the
consonant itself. No differences in F0 and
intensity were observed on the following vowel,
contra earlier reports. We further investigated
the robustness of this contrast using linear
discriminant analysis. Results show that
discrimination is above chance, but poor
(~62%). The large overlap between the two
categories may be partly due to the naturalistic
nature of our speech samples. However, we also
found that the contrast is neutralized in some
minimal pairs. This merger is surprising since
initial geminates are often the sole realization of
lexical and morphosyntactic contrasts. We
suggest that the singleton/initial geminate
contrast is now best characterized as a marginal
contrast. We hypothesize that this marginally
contrastive status may be the result of an on-
going sound change, perhaps connected with the
more modest role that initial geminates play in
Pattani Malay morphophonological alternations.
1 Introduction
Pattani Malay (PM), an Austronesian language
spoken in Southern Thailand (Uthai 2011), exhibits
a cross-linguistically unusual phonological ‘length’
contrast for all word-initial consonants, e.g., [matɔ]
‘eye’ vs [mːatɔ] ‘jewelry’. The long forms of initial
consonants, usually termed initial geminates (IGs),
have been reported to differ from singletons along
multiple acoustic dimensions. With regards to
duration, PM IGs have been reported to be, on
average, three times longer than their singleton
counterparts (Abramson 1987). Durational
differences are hardly a surprising finding since
closure duration is usually considered the most
reliable acoustic correlate of phonological length
cross-linguistically (Ladefoged and Maddieson
1996). If previous work is representative, however,
the IG/singleton duration ratio of 3:1 in PM would
be on the extreme side of the spectrum (Ladefoged
and Maddieson 1996).
Interestingly, duration is not the only cue that
distinguishes IGs from singletons in PM. IGs have
been reported to produce acoustic effects on the
following vowel as well. In particular, previous
research has reported that vowels following IGs

display longer duration, higher fundamental
frequency (F0), and higher intensity (Abramson
1987; Abramson 1998; Phuengnoi 2010). These F0
and intensity cues alone have been shown to be
reliable enough for native speakers to correctly
identify IGs vs singleton onsets; even in
environments where durational cues are ambiguous,
such as in absolute utterance-initial position where
closure duration cannot be distinguished from
preceding silence (Abramson 2003). Similar
acoustic features in production and perceptual
results have been reported for another closely
related variety, Kelantan Malay (Hamzah et al.
2019; Hamzah et al. 2020).
The concomitant manifestation of IGs in the
form of local durational differences and of effects
on intensity and F0 of the following vowel has led
scholars to hypothesize that PM speakers may be in
the process of reanalyzing consonantal length as a
prosodic contrast based on stress/pitch accent, or
that the language may even be on its way to
tonogenesis (Abramson 2004).
The possibility that IGs may be the target of
ongoing sound change warrants by itself a fresh
look at the realization of this unusual phonological
contrast. However, we should be cautious in
considering previous work the last word on PM IGs.
For one thing, previous studies were based on a
limited number of speakers (4 for Abramson, 7 for
Phuengnoi). Moreover, the difference between IGs
and singletons was studied only in words produced
in isolation or in words that appeared in a carrier
sentence. Finally, in previous studies, speakers were
explicitly instructed about the production of the
contrast in question. All these factors combined may
have led to an exaggeration of the differences
between IGs and singletons.
Given such limitations in previous studies, we
investigate again the acoustic correlates of IGs in
PM by comparing words with and without IGs, but
we do so in more ecologically valid speech, which
was elicited outside the lab using natural sounding
sentences. To characterize the differences between
IGs and singletons we make use of both statistical
and machine learning techniques.
Statistical analyses showed that IGs are longer
than their singleton counterparts, but the difference
is much smaller than reported by previous studies.
We also found no difference in F0 and intensity on
the vowel following IGs vs singletons, contra the
reports of previous studies.
Additionally, to quantify the robustness of the
IG/singleton contrast and to find out which
dimensions best discriminate the two categories, we
performed classification using linear discriminant
analysis (LDA) with a variety of models that
employ different combinations of acoustic features.
We found that the model performances are above
chance, but still poor, peaking at only about 62%
accuracy for the best feature combinations.
We speculate that the limited statistical
differences and low accuracy of the LDA may be
partly due to the naturalistic nature of the speech
materials we collected and to ongoing neutralization
of the contrast in some minimal pairs. We conclude
by discussing several hypotheses concerning the
mechanisms that may be at the heart of the observed
neutralization.
2 Acoustic Analyses
2.1 Methodology
14 native speakers of PM (6M; 8F) were asked to
pronounce 13 disyllabic minimal pairs differing
only for their word-initial onsets, which were either
geminate or singleton, as shown in Table 1. Stimuli
were presented orally with natural-sounding Thai
sentences containing the target words. Participants
were asked to translate the sentence into PM. Each
sentence was repeated six times.
singleton
(CVCV)
gloss
geminate
(CːVCV)
gloss
pagi
‘morning’
pːagi
‘early
morning’
paka
‘to
use/wear’
pːaka
‘usable’
tanɔh
‘land’
tːanɔh
‘outside’
dapo
‘kitchen’
dːapo
‘at the
kitchen’
katoʔ
‘hammer’
kːatoʔ
‘frog’
kabo
‘Java
kapok’
kːabo
‘beetle’
gaɟɨ
‘wage’
gːaɟɨ
‘saw’
ɟalɛ
‘path’
ɟːalɛ
‘to walk’
ɟuɣi
‘to steal’
ɟːuɣi
‘thief’
misa
‘mustache’
mːisa
‘to grow a
moustache’
labɔ
‘profit’
lːabɔ
‘spider’

singleton
(CVCV)
gloss
geminate
(CːVCV)
gloss
bulɛ
‘moon’
bːulɛ
‘month’
buŋɔ
‘flower’
bːuŋɔ
‘to bloom’
Table 1. Stimuli
Audio was collected at 44.1 kHz in Praat
(Boersma and Weenink 2020). All recordings were
made in quiet rooms at the Prince of Songkla
University Pattani Campus.
Segmental boundaries were obtained in Praat
TextGrids by forced alignment using the Montreal
Forced Aligner (McAuliffe et al. 2017). The
TextGrids were inspected and manually corrected
when necessary. The corrected TextGrids
containing segmental boundaries and the audio
signals of each word were read back in MATLAB®
for analysis.
Eight acoustic measurements were collected:
(1) Duration of initial segments (ms)
(2) Duration of initial syllables (ms)
(3) F0 mean of initial syllables (semitone)
(4) Intensity peak of initial syllables (dB)
(5) F0 mean over initial 10% of vowel
following target consonants (semitone)
(6) Intensity mean over initial 10% of vowel
following target consonants (dB)
(7) Difference between semitone transformed
mean F0 of initial and final syllable
(8) Ratio of mean RMS amplitude of initial to
final syllable
F0 was calculated using a MATLAB®
implementation of Talkin’s robust algorithm for
pitch tracking (Talkin 1995) contained in the
Voicebox toolbox for MATLAB®. (Brookes 1997).
F0 was further processed within all trials and
separately by participant by removing all data points
with standard deviation scores greater than 2 from
the mean; datapoints deviating ±10 Hz from
neighboring samples were also excluded. When the
F0 vector of a word contained less than 5 datapoints
per each syllable, the contour was no longer
processed, as interpolation over the entire word
would not be reliable. In the other cases, F0 was
subsequently interpolated using spline interpolation
and smoothed using a median filter. F0 was
transformed by converting from Hz to semitones
according to the equation
!"
#$%!""!×!𝑙𝑜𝑔!&
&
'(
)'(
' in
Zhang (2018).
Sound Pressure Level (SPL) normalized
intensity was calculated by transforming the root
mean squared intensity of the signal to dB and
normalizing to human auditory threshold using the
formula
20 ×𝑙𝑜𝑔!&
*
+&
. In this formula P represents
the power of the signal and P0 represents the
normalizing term for the auditory threshold of a
1000 Hz sine wave, equal to
2 × 10,-
(Huang et al.
2001).
Statistical analyses were conducted by fitting
linear mixed effect regressions. We compared a
model where the fixed effect was the
presence/absence of IGs to an intercept-only model.
Random effects were subject, word, and position of
the word in the phrase (medial or final). Random
intercepts were present in the model for each
random effect. Random slopes were added when
they resulted in a better fit as determined via a
loglikelihood ratio test. Loglikelihood ratio tests
were, thus, used to assess statistical significance and
to determine the random effect structure.
2.2 Results
Consonant Duration: Comparing the initial
segment in the IG and no IG condition, we found
that IGs are significantly longer than singletons
(χ2(1) = 4.03, p = .04) with an effect size estimated
at 17 ms, as illustrated in Figure 1.
Figure 1. Comparison of initial segment duration
(ms)
Syllable Duration: The presence of IGs does not
significantly affect the duration of the initial
syllable (χ2(1) = 1.34, p= .24), as illustrated in
Figure 2.

Figure 2. Comparison of syllable duration of initial
syllables (ms)
F0: The presence of IGs does not significantly
affect the mean F0 of the initial syllable (χ2(1) =
0.16, p= .69), as illustrated in Figure 3.
Figure 3. Comparison of mean F0 of initial
syllables (semitones)
Intensity: IGs do not significantly affect the
maximum SPL normalized intensity of the initial
syllable (χ2(1) = 0.49, p= .48), as illustrated in
Figure 4.
Figure 4. Comparison of maximum SPL
normalized intensity of initial syllables (dB)
To further investigate whether the effects of IGs
on the following vowel may be limited to the region
immediately following the release, we also
examined mean F0 and intensity over the first 10%
of the vowel, following previous work on Kelantan
Malay (Hamzah et al. 2020).
We found no significant differences between
mean F0 over the initial 10% of the vowel following
IGs vs. singletons (χ2(1) = 0.06, p= .79). F0 contours
over the vowel are presented in Figure 5.
Figure 5. Comparison of time normalized F0
trajectory of initial vowel in semitone. Shaded
areas represent ±2 Standard Errors
We also found no significant difference between
mean SPL normalized intensity over the initial 10%
of a vowel following IGs vs. singletons (χ2(1) =
0.95, p= .33). The intensity contours of the
following vowel are presented in Figure 6.
Figure 6. Comparison of time normalized SPL
normalized intensity trajectory in dB. Shaded areas
represent ±2 Standard Errors
Finally, also following previous work
(Abramson 1998; Hamzah et al. 2020), we
examined whether differences between IGs and
singletons may be manifested more globally in the

F0 difference and RMS amplitude ratios of the two
syllables. We found no differences for both F0
(χ2(1) = 0.007, p= .93) and RMS amplitude (χ2(1) =
0.07, p= .79), as illustrated in Figure 7.
Figure 7. F0 difference and RMS amplitude ratio
2.3 Summary
We found that the durations of IGs and singletons
are significantly different, but, unlike in previous
studies, the duration of IGs is not three times longer
than singletons. The durational differences are
estimated at about 17 ms. Furthermore, there is a
significant overlap between the two distributions.
Contrary to previous descriptions, the presence or
absence of IGs does not have a significant effect on
syllable duration, mean F0, or peak intensity of the
following vowel; no effect is observed even if only
10% of the vowel is examined. We also observed no
significant differences in the F0 difference and
amplitude ratios of the two syllables.
In sum, we found only very small durational
differences between IGs and singletons and the
other acoustic measurements do not display
significant differences.
3 Linear Discriminant Analysis
To further address the question of whether the
singleton/IG contrast in PM is comparable in terms
of its magnitude to the singleton/geminate contrast
of other languages, we performed classification of
IGs vs. singletons using linear discriminant analysis
(LDA). In a nutshell, LDA is a classification
technique (and also a dimensionality reduction
technique) that uses linear combinations of features
to maximize the separation between two or more
categories. LDA is of interest here because it has
been successfully applied to the study of various
phonetic contrasts, including geminate vs non-
geminate contrasts in both word-medial, in Japanese
(Idemaru and Guion-Anderson 2010) and Lebanese
Arabic (Khattab and Al-Tamimi 2014), and word-
initial position, in Salentino (Burroni and Maspong
to appear). We tried to extend this methodology to
characterize the word-initial geminate contrast of
PM.
3.1 Methodology
We fitted LDA models using cross-validation to
evaluate the accuracy of our models. We randomly
assigned 80% of the data to a training set and the
remaining 20% to a test set. 10,000 such LDA
models were fitted for each combination of
predictors. The mean accuracy and standard
deviations reported here were taken over these
10,000 iterations.
To determine which acoustic dimensions were
more apt to discriminate the singleton/IG contrast,
we considered that duration of the first segment
(CDur) and ratio of the duration of the first segment
to the entire word (CDur / WordDur) are the only
two statistically significant differences present in
our data. We then tested whether adding
information concerning the duration (σi Dur), mean
F0 (σi MeanF0), and maximum intensity (σi MaxInt)
of the target syllable would improve LDA
classification. All features were z-scored by
participants before performing LDA, as this
procedure is known to improve LDA classification.
3.2 Results
We found that the model performance is above
chance (that is, above 50%), but still quite poor, as
summarized in Table 2, peaking at only about 62%
accuracy for the best linear combination of features:
the duration of the first segment (CDur) alone or in
combination with the duration ratio of the first
segment to the entire word (CDur / WordDur).

Model Structure
Mean
Accuracy
Standard
Deviation
CDur +
CDur/WordDur +
σiDur + σiMaxInt +
σiMeanF0
58.84%
2.18%
CDur +
CDur/WordDur +
σiDur + σiMeanF0
58.20%
2.07%
CDur +
CDur/WordDur +
σiDur + σiMaxInt
58.88%
2.20%
CDur +
CDur/WordDur +
σiDur
58.19%
2.10%
CDur +
CDur/WordDur
61.40%
2.06%
CDur/WordDur
59.84%
2.14%
CDur
62.36%
2.11%
Table 2. Accuracy of LDA models for different
combinations of features
Optimizing the hyperparameters of the model
does not greatly improve performance in the
identification of IGs as is clear from the confusion
matrix of the optimized model presented in Figure
8.
Figure 8. Confusion matrix showing the number of
IGs (class 1, top) and singletons (class 2, bottom)
classified correctly (gray diagonal) and incorrectly
(orange diagonal).
If we inspect the predicted boundary between the
two classes, as shown in Figure 9, the reason for the
low performance of the model becomes clear: IGs
and singletons are not linearly separable in the
investigated acoustic dimensions, thus, they cannot
be captured by an LDA classifier.
Figure 9. Output of LDA showing large overlap
between categories
The low LDA accuracy for geminates contrasts
sharply with high accuracy reported for other
languages. For instance, for medial geminates in
Japanese, accuracy is at ~85-95% (Idemaru and
Guion-Anderson 2010) and, for IGs in Salentino,
accuracy is at ~80% (Burroni and Maspong to
appear).
3.3 Summary
In sum, the discrimination above chance shows that
there is indeed a contrast between words with and
those without IGs that can be picked up by a simple
model, such as an LDA classifier. This is in line
with previous phonetic and phonological research
on PM and justifies looking for contrasts between
words with and without IGs. On the other hand, the
low classification accuracy suggests that the
contrast is subtle.
We now discuss what factors may be responsible
for the observed overlap between IGs and
singletons.
4 Discussion
We have three non-mutually exclusive hypotheses
to explain why the contrast between IGs and
singletons looks much less robust than previously
reported.
The first possibility that comes to our mind is
that the differences between the result of our study

and previous work is due to different methods of
data collection. Previous work (Abramson 1987;
Abramson 1998) examined IGs only in isolation and
in a carrier sentence. Our data, on the other hand,
presented IGs and their singleton counterparts in
naturalistic sentences. Accordingly, the difference
could be due to less carefully articulated speech.
A second possibility is that the contrast may be
neutralized for some speakers. The size of our
dataset does not allow for a full quantitative
assessment of this claim; however, our impression
is that almost all speakers produce IGs that are
longer than singletons on average, as illustrated in
Figure 10.
Figure 10. Mean duration of singletons (left) and
IGs (right) by speaker (ms)
A third possibility is that the contrast only exists
for a subset of minimal pairs. This means that, for
many lexical items, the contrast between singletons
and IGs is not realized.
Indeed, our data suggests that closure duration of
the initial consonants is distinct only for a subset of
minimal pairs, as illustrated in Figure 11.
Figure 11. Mean duration of singletons (left) vs
IGs (right) by word (ms)
Given this observation, we ask what
generalizations may explain the observed
neutralizations, as well as the non-neutralizations.
In the framework of Evolutionary Phonology
(EP), IGs have been hypothesized to be
diachronically unstable (Blevins 2004).
Furthermore, EP holds that the stability of phonetic
cues to IGs may be related to their wider role in the
grammar. IGs survive only in languages where they
represent the only cue to lexical contrasts and
produce “sentential minimal pairs”. In other words,
IGs survive only when they compete lexically with
singletons and cannot be disambiguated by context
(Blevins and Wedel 2009; Burroni and Maspong to
appear).
Interestingly, PM seems a counterexample to this
generalization, as IGs are being lost in this
language, even though they are the unique
realization of morphosyntactic contrasts. For
instance, under an EP approach, the observed
merger of [dapo] ‘kitchen’ and [dːapo] ‘at the
kitchen’ is expected, since these forms appear in
different positions and can be disambiguated by
context. Similarly, the non-merger of [katoʔ]
‘hammer’ and [kːatoʔ] ‘frog’ is expected since these

forms appear in the same context and the IG or lack
thereof is the only cue distinguishing them.
However, other mergers, such as [kabo] ‘Java kapok
(type of plant)’ and [kːabo] ‘beetle (type of bug)’,
are not expected, because context does not allow for
disambiguation, thus, the neutralizing IG would be
one that is a unique cue to the contrast, just like the
non-merging one in [katoʔ]/[kːatoʔ]. However, the
merger of [kabo]/[kːabo] may suggest some role for
word frequency effects. Phillips (2006) explained
that retrieving low-frequency word is a challenge
for the learner. These difficulties, in turn, may lead
to alterations of the phonetic forms of low frequency
words on the model of unmarked patterns, that IGs
may be altered to singletons. At any rate, for another
counterexample to the EP claim that cues to IGs are
dependent on lexical competition, we refer the
reader to Burroni and Maspong (to appear). Since
lexical competition alone does not explain the
paradox of IGs merging with singletons in PM,
other factors need to be considered.
It has been reported that PM speakers no longer
make use of IGs for the purpose of morphological
derivation due to contact with Thai (Uthai 1993),
accordingly, it is possible that the contrastive
phonological status of IGs is being eroded in
connection with their reduced ‘functional’ role in
the grammar. If IGs and singletons will be merging
at evolutionary timescales, the loss of PM IG
contrasts would be a striking example of sound
change via lexical diffusion connected with a
reduced functional load, an information theoretic
measurement that has been argued to correlate with
geminate to singleton ratio (Tang and Harris 2014)
and resistance to merger (Wedel et al. 2013).
Further research is necessary to test the merits of
these hypotheses on the basis of a larger PM dataset.
Corpus frequencies also need to be obtained in order
to calculate information theoretic measurements,
such as functional load (Surendran and Niyogi
2006).
At any rate, since the contrast between IGs and
singletons is only observed for some minimal pairs,
it may be best interpreted as a quasi-phonemic or
marginal contrast (Hall 2013; Renwick and Ladd
2016). If this interpretation is correct, our acoustic
results would align with recent work demonstrating
that marginal phonological contrasts may display
large overlaps when data is collected outside the lab,
in more naturalistic contexts (Cohn and Renwick
2019).
5 Conclusion
In this paper, we have shown that the only
significant difference between PM IGs and
singletons in naturalistic speech is the duration of
the consonants themselves. We have further shown
that an LDA model is able to discriminate between
syllables with and without IGs slightly above
chance level (~62%). This is much below usual
LDA performances for geminates in other
languages.
The striking difference between our findings and
earlier reports regarding the robustness of cues to
IGs in PM calls for an explanation. One possibility
is that previous experimental work may have
exacerbated the difference between IGs and
singletons. After all, highly controlled lab speech is
very different from less carefully articulated
naturalistic speech. IGs in PM could then be an
example showing that a more nuanced
characterization of phonological contrasts requires
an integrated analysis of both laboratory and more
naturalistic phonetic data, as advocated by Cohn and
Renwick (2019).
However, we have also shown that, although
speakers on average produce longer IGs than
singletons, they produce the contrast only for a
subset of minimal pairs. We have speculated that an
appropriate characterization of the subsets that
undergo and resist merger will require further
collection of information theoretic measurements,
such as functional load. One thing is relatively
clearer: IGs are moving towards a more marginally
contrastive role in the grammar of PM, a fact that
may be reflected in their phonetic realization.
Acknowledgements
We would like to thank Santhawat Thanyawong for
his help with the data collection. We would also like
to thank three anonymous reviewers and Sam Tilsen
for their feedback. We are grateful of the Southeast
Asian Linguistics Research Unit, Faculty of Arts,
Chulalongkorn University for the financial support.
References
Abramson, Arthur S. 1987. Word-initial consonant
length in Pattani Malay. International Congress of
Phonetic Science (ICPhS) 6, 68–70.

Abramson, Arthur S. 1998. The complex acoustic output
of a single articulatory gesture: Pattani Malay word-
initial consonant length. Annual meeting of the
Southeast Asian Linguistics Society (SEALS) 4, 1–20.
Abramson, Arthur S. 2003. Acoustic cues to word-initial
stop length in Pattani Malay. International Congress
of Phonetic Science (ICPhS) 15, 387–390.
Abramson, Arthur S. 2004. Toward prosodic contrast:
Suai and Pattani Malay. In International Symposium
on Tonal Aspects of Languages: Emphasis on Tone
Languages, 1–4.
Blevins, Juliette. 2004. Evolutionary Phonology: the
emergence of sound patterns. Cambridge; New York:
Cambridge University Press.
Blevins, Juliette & Andrew B. Wedel. 2009. Inhibited
sound change: An evolutionary approach to lexical
competition. Diachronica 26(2), 143–183.
Boersma, Paul & David Weenink. 2020. Praat: doing
phonetics by computer. [Computer software].
Brookes, Mike. 1997. Voicebox: Speech processing
toolbox for MATLAB. [Computer Software].
Burroni, Francesco & Sireemas Maspong. to appear. Re-
examining Initial Geminates: Typology, Evolutionary
Phonology, and Phonetics. In Historical Linguistics
2019.
Cohn, Abigail C. & Margaret E. L. Renwick. 2019.
Doing Phonology in the age of big data. In Cornell
Working Papers in Phonetics and Phonology 2019,
1–36.
Hall, Kathleen Currie. 2013. A typology of intermediate
phonological relationships. The Linguistic Review
30(2), 215 – 275.
Hamzah, Mohd Hilmi, J Fletcher & J Hajek. 2019. The
secondary roles of amplitude and F0 in the perception
of word-initial geminates in Kelantan Malay.
International Congress of Phonetic Science (ICPhS)
19, 2735-2757.
Hamzah, Mohd Hilmi, John Hajek & Janet Fletcher.
2020. Non-durational acoustic correlates of word-
initial consonant gemination in Kelantan Malay: The
potential roles of amplitude and f0. Journal of the
International Phonetic Association. Cambridge
University Press 50(1), 23–60.
Huang, Xuedong, Alex Acero, Hsiao-Wuen Hon & Raj
Foreword By-Reddy. 2001. Spoken language
processing: A guide to theory, algorithm, and system
development. Prentice hall PTR.
Idemaru, Kaori & Susan Guion-Anderson. 2010.
Relational timing in the production and perception of
Japanese singleton and geminate stops. Phonetica
67(1–2), 25–46.
Khattab, Ghada & Jalal Al-Tamimi. 2014. Geminate
timing in Lebanese Arabic: The relationship between
phonetic timing and phonological structure.
Laboratory Phonology 5(2), 231–269.
Kraehenmann, Astrid. 2011. Initial geminates. The
Blackwell companion to phonology. Wiley Online
Library.
Ladefoged, Peter & Ian Maddieson. 1996. The sounds of
the world’s languages. Oxford; Cambridge, MA:
Blackwell Oxford.
McAuliffe, Michael, Michaela Socolof, Sarah Mihuc,
Michael Wagner & Morgan Sonderegger. 2017.
Montreal Forced Aligner. [Computer software].
Phillips, Betty S. 2006. Word frequency and lexical
diffusion. New York: Palgrave Macmillan.
Phuengnoi, Nattaphon. 2010. An acoustic study of
stressed and unstressed syllables in Pattani Malay
and Urak Lawoi’. Bangkok: Chulalongkorn
University thesis.
Renwick, Margaret E. L. & D. Robert Ladd. 2016.
Phonetic distinctiveness vs. lexical contrastiveness in
non-robust phonemic contrasts. Laboratory
Phonology 7(1). 1–29.
Surendran, Dinoj & Partha Niyogi. 2006. Quantifying the
functional load of phonemic oppositions, distinctive
features, and suprasegmentals. In Ole Nedergaard
Thomsen (ed.), Competing Models of Linguistics
Change: Evolution and Beyond. In commemoration of
Eugenio Coseriu (1921-2002), 43–58. Amsterdam &
Philadelphia: Benjamins.
Talkin, David. 1995. A robust algorithm for pitch
tracking (RAPT). In Speech Coding and Synthesis.
495-518. New York, NY: Elsevier Science.
Tang, Kevin & John Harris. 2014. A functional-load
account of geminate contrastiveness: a meta-study. In
Linguistics Association of Great Britain 2014.
Uthai, Ruslan. 1993. A comparison of word formation in
Standard Malay and Pattani Malay. Bangkok:
Chulalongkorn University thesis.
Uthai, Ruslan. 2011. Keistimewaan dialek Malayu
Patani. Bangi: Penerbit Universiti Kebangsaan
Malaysia.
Wedel, Andrew, Abby Kaplan & Scott Jackson. 2013.
High functional load inhibits phonological contrast
loss: A corpus study. Cognition 128(2). 179–186.
Zhang, Jingwei. 2018. A Comparison of Tone
Normalization Methods for Language Variation
Research. Pacific Asia Conference on Language,
Information and Computation (PACLIC) 32, 823-
831.