1 Introduction
This study investigates the syllabification of medial geminates in Moroccan Arabic using patterns of temporal stability. While previous phonological research suggests that both medial clusters and geminates in Moroccan Arabic are heterosyllabic (Dell & Elmedlaoui, 2002; Noamane, 2021), this claim has not been tested phonetically for medial geminates. Temporal stability patterns have been established as a reliable diagnostic for the syllabification of consonant clusters. If these patterns can be generalized to geminates, they should yield results consistent with the independently established syllabification. This study therefore has two aims: (1) to test the competing heterosyllabic and tautosyllabic representations for Moroccan Arabic medial geminates, and (2) to evaluate whether temporal stability metrics produce results that align with independent phonological evidence.
Temporal stability studies examine syllable structure by investigating the alignment of consonant gestures relative to a following vowel (Browman & Goldstein, 1988; Goldstein et al., 2007; Shaw et al., 2009). Two patterns of stability have been identified in the literature: c-centre-to-anchor and right-edge-to-anchor. As shown in Figure 1, in the c-centre-to-anchor pattern, the midpoint of the consonantal gestures (the c-centre) is aligned at a stable distance from the end of the following vowel (the anchor), regardless of the number of consonants. This consistent timing of clusters of varying size toward a following vowel reflects the kind of tight gestural coordination characteristic of complex onsets. In the right-edge-to-anchor pattern, however, the midpoint of the last consonant in a sequence is aligned at a stable distance from the end of the following vowel, regardless of the number of consonants. Since only the last consonant shows stable alignment, this means that earlier consonants are not co-timed with the vowel, which suggests that the consonant sequence cannot form a complex onset.
Figure 1: Schematic representations of c-centre-to-anchor (left) and right-edge-to-anchor (right) stability patterns, adapted from Shaw et al. (2009).
![]()
Previous research on temporal stability patterns has predominantly focused on initial clusters. Languages like English (Browman & Goldstein, 1988; Marin & Pouplier, 2010), Romanian (Marin, 2011, 2013), and Georgian (Goldstein et al., 2007) have been observed to exhibit a c-centre-to-anchor pattern, indicating that they allow complex onsets word-initially. On the other hand, languages such as Moroccan Arabic (Shaw et al., 2009), Jazani Arabic (Durvasula et al., 2021), Tashlhiyt Berber (Goldstein et al., 2007; Hermes et al., 2011), and Hebrew (Tilsen et al., 2012) show a right-edge-to-anchor pattern, and, therefore, have simplex onsets word-initially. Most of these studies use articulatory methods, such as Electromagnetic Articulography (EMA), to examine temporal stability patterns. Durvasula et al. (2021) proposed using acoustic measurements to study the syllabification of initial clusters in American English and Jazani Arabic. The advantage of using acoustic data is the fact that such data is more accessible and easier to collect than articulatory data, thereby facilitating studies with a larger number of participants. Durvasula et al. (2021)’s methodology successfully identified consistent acoustic landmarks (e.g., burst releases, vowel onset, and formant transitions) to demarcate consonantal intervals relative to the adjacent vowel. They measured the temporal distance between these landmarks across different sizes of consonant clusters, which resulted in an approximation of the articulatory timing captured by prior articulatory methods. Their findings for American English show that it exhibits a c-centre-to-anchor stability pattern, which is consistent with previous articulatory studies on English (Browman & Goldstein, 1988; Marin & Pouplier, 2010). Medial consonant clusters, however, have not been extensively studied using this methodology (but see Lialiou et al. 2021 for English).
The syllabification of geminates has been a topic of theoretical importance in phonology. Two main approaches have been proposed to represent medial geminates: one approach analyzes them as heterosyllabic, in which the geminate span a syllable boundary (VCi.CiV; McCarthy, 1979a; Leben, 1980; Clements & Keyser, 1983; Levin, 1985; Hayes, 1989; Selkirk, 1990; Tranel, 1991; Ringen & Vago, 2011; Morén, 2013), while a second approach treats them as tautosyllabic (V.CiCiV; Martinet, 1975; Mohanan, 1989; Topintzi, 2008, 2022; Wolf, 2008; Bailiuk, 2018). Importantly, neither approach is assumed to be universal: it has been shown that some languages are better analyzed as having heterosyllabic geminates (e.g., Italian (Zmarich & Fivela, 2005; Di Benedetto et al., 2021) and Tashlhiyt Berber (Ridouane, 2010; Lahrouchi, 2017)), while others favor a tautosyllabic analysis (e.g., Marshallese (Topintzi, 2008), Cypriot Greek (Armosti, 2009; Topintzi, 2022), and Ukrainian (Bailiuk, 2018)). In addition, the position of a geminate also matters: word-initial and word-final geminates often show different behavior from word-medial ones (Ham, 2001; Topintzi, 2008). This study focuses on medial geminates, for which both heterosyllabic and tautosyllabic parses have been proposed.
The present study proposes that temporal stability measures offer a phonetically grounded way of distinguishing between the two syllabification parses for medial geminates. The two parses make different predictions about how geminates should pattern with respect to temporal alignment: if a geminate is a single unit, its mid-point should show temporal stability in the same way a c-centre does in clusters. This would suggest that the geminate has a tautosyllabic representation (V.CiCiV). On the other hand, if the geminate consists of two timing units, only the second unit should align stably, mirroring the onset consonant in a heterosyllabic cluster (VCi.CiV). It should be noted, however, that this pattern would only partially identify the syllabification: while it would rule out a tautosyllabic parse, it would not determine whether the first timing unit functions as a coda or forms a degenerate syllable (see Section 4 for discussion). To date, however, this prediction has not been explicitly tested. The present study addresses this gap by adopting a temporal stability methodology that treats the two halves of a geminate as separate timing units, which allows for testing whether geminates align with clusters (as predicted by the heterosyllabic parse) or behave as single units (as predicted by a tautosyllabic parse).
Moroccan Arabic provides an ideal testing ground for this question. In Moroccan Arabic, the syllabification of medial clusters and geminates as heterosyllabic is widely assumed in the phonological literature (Al Ghadi, 1990, 1994; Jebbour, 1996; Boudlal, 2001; Dell & Elmedlaoui, 2002; Noamane, 2019, 2020, 2021; Bensoukas, 2025; among others). This assumption is supported by multiple independent lines of evidence including phonological, phonotactic and prosodic constraints. One major source of evidence comes from versification and poetic meter; in Dell & Elmedlaoui’s ( 2002) versification analysis of Moroccan Arabic poetry, medial clusters and geminates are consistently parsed as heterosyllabic to satisfy metrical constraints. By analyzing text-to-tune alignments in traditional sung poetry (melhun), they show that the Moroccan Arabic meter requires alternating patterns of Light and Heavy syllables. To satisfy these rhythmic requirements, medial geminates must be split: the first half acts as a weight-bearing coda that makes the preceding syllable heavy, while the second half serves as the onset for the following syllable (e.g., /zewweq/ is scanned as zew.weq).
Another source of evidence comes from schwa epenthesis; Noamane (2021) emphasized that medial geminates and clusters share a heterosyllabic parse using the example of schwa epenthesis in CCCC words (/CCCC/ → [CəCCəC]), which forces a heterosyllabic parse for both geminates and clusters, as a tautosyllabic parse would create a schwa in an open syllable (Cə.CCəC) or an onsetless syllable (CəCC.əC), which are both prohibited in Moroccan Arabic (Al Ghadi, 1990; Boudlal, 2001; Bensoukas & Boudlal, 2012; among others). Finally, the assumption that medial clusters and geminates have a heterosyllabic parse is also drawn from the syllabification of initial and final consonant sequences. For instance, the foot binarity constraint require that prosodic words be minimally bimoraic, leading to heterosyllabic parsing of initial and final clusters and geminates in triconsonantal words like CCəC or CəCC, where the initial consonant in CCəC and the final consonant in CəCC form degenerate syllables (Al Ghadi, 1994; Jebbour, 1996; Boudlal, 2001). Based on these constraints, as well as additional evidence from the phonotactics and phonology of Moroccan Arabic, it has been assumed that medial geminates and clusters pattern similarly, sharing a heterosyllabic parse. This provides a baseline syllabification against which the results of temporal stability measures can be evaluated.
Medial clusters and geminates in Moroccan Arabic have also been studied phonetically. Recent articulatory work shows that medial clusters in Moroccan Arabic exhibit a right-edge-to-anchor stability pattern, indicating a heterosyllabic temporal organisation (Gafos et al., 2020). As for geminates, EMA data on medial geminate plosives show that they are produced as two overlapped consonantal gestures (Zeroual et al., 2015), since their closing and opening movements share similar temporal and kinematic properties with those of clusters. Although this latter study does not explicitly address syllabification, it suggests that the internal temporal structure of geminates in Moroccan Arabic may be compatible with a heterosyllabic parse. Taken together, these observations make Moroccan Arabic an ideal case to test whether geminates and clusters pattern similarly with respect to temporal alignment (as predicted by Zeroual et al., (2015)), and whether temporal stability measures align with the established phonological evidence for the heterosyllabicity of medial geminates, supporting this methodology as a useful diagnostic for probing the syllabification of geminates.
Using acoustic measurements similar to the ones proposed by Durvasula et al., (2021), this study examines the temporal stability patterns of medial geminates and clusters in Moroccan Arabic. It improves on Gafos et al.’s (2020) EMA study of medial clusters in Moroccan Arabic by including a larger number of speakers and using only real word stimuli. I also investigate whether the sonority profile of medial clusters affects temporal stability patterns and whether vowel duration before geminates and clusters provides converging evidence for syllabification. The results show that both geminates and clusters exhibit a right-edge-to-anchor stability pattern, supporting a heterosyllabic representation.
The remainder of this paper is organized as follows. Section 2 describes the experimental methodology, including details about the participants, materials, procedure, and measurements. Section 3 presents the results of the temporal stability analysis. Finally, Section 4 discusses the findings in relation to previous research and proposes directions for future research.
2 Methods
To determine whether a medial geminate is a heterosyllabic unit that span a syllable boundary (VCi.CiV) or tautosyllabic single unit (V.CiCiV), I conducted a production experiment examining the temporal stability patterns of both medial geminates and clusters in Moroccan Arabic. Participants produced target items embedded in a carrier phrase, and acoustic measurements of the c-centre-to-anchor and right-edge-to-anchor intervals were identified and extracted from the recordings. Additionally, the duration of the vowel preceding the geminate/cluster was also extracted.
2.1 Participants
Ten native speakers of Moroccan Arabic participated in this study (four male, six female). All participants were between 19 and 39 years of age and were all from Rabat, the capital city of Morocco. Most participants had a basic to intermediate knowledge of French, while a few had basic knowledge of English. None of the participants spoke Berber. The participants were recruited through word-of-mouth, and were provided with monetary compensation to acknowledge their contribution. The study was approved by the Institutional Review Board, and formal consent was obtained from each participant prior to their involvement in the study.
The sample size of ten speakers is consistent with prior work on temporal stability. Articulatory studies have often used between 3 and 7 speakers (e.g., 5 speakers in Zeroual et al., 2015), (4 speakers in Shaw et al., 2009; Gafos et al., 2020), (7 speakers in Marin & Pouplier, 2010; Lialiou et al., 2021), (3 speakers per language in Hermes et al., 2017), reflecting the resource-intensive nature of articulatory data collection. Durvasula et al. (2021), in the only prior acoustic implementation of temporal stability measures, tested 10 speakers of American English (Experiment 1) and 7 speakers of Jazani Arabic (Experiment 2). Therefore, the present sample size of 10 speakers exceeds typical articulatory studies and matches the acoustic study of Durvasula et al. (2021). Importantly, because acoustic data is less resource-intensive to collect than articulatory data, this methodology can be efficiently scaled to larger samples in future work.
2.2 Materials
The stimuli consisted of twenty target items, which were divided into two subsets: five singleton–geminate pairs (ten items) with a medial VCiV/VCiCiV contrast (Table 1) and five pairs with a medial VCjV/VCiCjV contrast (Table 2). Unlike Gafos et al. (2020), who used some nonce words as target items, all twenty items in this study were real words; this encourages natural pronunciation and avoids the potential timing disruptions that often occur with nonce words such as hyperarticulation or increased variability due to speakers’ unfamiliarity with the items. While real words may be influenced by lexical factors such as word frequency, the statistical modeling approach used in this study controlled for these potential confounds.
Table 1: Target items with medial singletons and geminates.
| VCiV |
VCiCiV |
| sˤarˤa |
‘Sara’ |
dˤarˤrˤa |
‘harmful’ |
| ʕuma |
‘swim’ |
ʕamma |
‘general’ |
| kala |
‘support’ |
lalla |
‘grandma’ |
| sˤatˤa |
‘girl’ |
ħatˤtˤa |
‘put’ |
| riʃa |
‘feather’ |
kaʃʃa |
‘blanket’ |
Table 2: Target items with VCjV and VCiCjV.
| VCjV |
VCiCjV |
| baqa |
‘still’ |
sabqa |
‘surpassed’ |
| sala |
‘finished’ |
makla |
‘food’ |
| nuba |
‘turn’ |
rakba |
‘riding’ |
| mama |
‘mom’ |
raʒma |
‘stoning’ |
| ʕana |
‘suffered’ |
baʕna |
‘sold us’ |
All target words were presented as sets differing in the number of consonants word-medially. Within each pair, the final consonant was held constant. There is, however, a variety of consonants and consonant clusters across pairs. The medial clusters included both plateau sonority sequences (stop-stop: [bq] in sabqa, [kb] in rakba) and rising sonority sequences (obstruent-sonorant: [kl] in makla, [ʒm] in raʒma, [ʕn] in baʕna). No falling sonority clusters were included in the stimulus set, as the study was not originally designed to systematically test sonority effects. This decision was informed by prior studies on Arabic varieties including Moroccan Arabic (Gafos et al., 2020) and Jazani Arabic (Durvasula et al., 2021). Gafos et al. (2020), for instance, found a uniform pattern of interval stability across the different sonority profiles in word medial clusters, which indicates that sonority does not play a crucial role in these timing patterns. However, in response to reviewer feedback, post-hoc analyses were conducted following the approach of Durvasula et al. (2021) to examine whether the available sonority profiles (plateau vs. rising) affect temporal stability patterns.1
To isolate the effect of the medial consonants, the following vowel was fixed as [a] across all target items. In addition to the target items, forty filler items were added, yielding 60 total sentences. These fillers were carefully selected to differ from the target items in various ways: they varied in length, had different syllabic templates (e.g., CCVC, CVCVC) compared to the CVCV and CVCCV templates of the target words, and contained a more diverse set of consonants and vowels.
2.3 Procedure
Each sentence was repeated five times, which resulted in 300 tokens per participant (60 sentences × 5 repetitions). These repetitions were distributed across five blocks, meaning participants produced one utterance of each word per block. The stimuli were embedded within the carrier phrase [ʒibi __ hnaja] ‘bring __ here’. The 300 sentences were arranged in a pseudo-randomized order and presented on PowerPoint slides generated using a Python script.
Before the main recording session, participants completed a practice phase where they read three sample sentences, each repeated once, to familiarize themselves with the task. During the main phase, participants were presented with the 300 sentences divided into five blocks of 60, with short breaks between each block. The recordings were made in a controlled, quiet environment using WaveEditor Android app (Sound-Base Audio, 2024) on a Google Pixel 8 Pro. Participants were asked to read the sentences fluently. In cases of hesitation or errors, they were asked to repeat the sentence, which allowed for the inclusion of the complete dataset for analysis. Durvasula et al. (2021) used a different approach where they did not request repetitions during recordings; instead, they excluded tokens with hesitations later from their analysis. Each session lasted approximately 20 minutes.
2.4 Measurements
The recordings were manually cleaned using Audacity (Audacity Team, 2024) to remove errors and filler items. Automatic annotations were generated using WebMAUS (Schiel, 1999; Kisler et al., 2017), a tool designed for aligning recordings with their corresponding transcripts. These annotations were then reviewed and manually corrected in Praat (Boersma & Weenink, 2022) to address any misalignments. Identifying the right-edge-to-anchor interval for geminates presents a methodological challenge, since geminates are usually realized as a single, continuous consonantal gesture and therefore do not have an internal boundary that corresponds to a second consonantal unit. To allow temporal stability to be evaluated in a way that is comparable to clusters, I divided each geminate into two equal halves in terms of duration. While previous studies have shown that geminates may behave like two identical consonants (e.g., Italian geminates showing double bursts, Di Benedetto et al. (2021); or Moroccan Arabic geminate stops having two overlapped gestures, Zeroual et al. (2015)), these studies, however, treat the geminate as one single interval in their measurements. I take a different approach by treating the two temporal halves of a geminate as independent units for the purpose of measurement, which enables a direct comparison with the temporal alignment of clusters. This choice is methodologically motivated, and, to my knowledge, this study is the first one to analyze geminates by explicitly splitting them into two temporal units.
Two temporal intervals were extracted from the resulting annotations: (i) right-edge-to-anchor – from the midpoint of the rightmost consonant (or right half of a geminate) to the end of the following vowel, and (ii) c-centre-to-anchor – from the mean of the midpoints of all medial consonants (or the whole geminate) to the end of the following vowel. Defining the anchor varies across the temporal stability literature, and different anchors can result in qualitatively different stability patterns (Shaw et al., 2009; Sotiropoulou et al., 2020). Articulatory studies have used both consonantal anchors—such as the target or maximum constriction of a postvocalic consonant (Shaw et al., 2009; Marin & Pouplier, 2010)—and vocalic anchors—such as vowel target attainment (Goldstein et al., 2007) or vowel spatial extremum (Sotiropoulou et al., 2020; Lialiou et al., 2021). To assess robustness, some articulatory studies have used multiple anchors simultaneously (Sotiropoulou et al., 2020; Lialiou et al., 2021). Shaw et al. (2009) directly compared consonantal and vocalic anchors and found that vocalic anchors resulted in higher variability across all interval types, even though the underlying stability patterns were generally preserved.
Articulatory landmarks such as gestural target attainment or spatial extremum do not correspond directly to acoustic segment boundaries. For acoustic implementations, the relevant landmarks are the onset and offset of segments. Vowel offset is generally more identifiable than vowel onset, since the transition from a vowel into a following consonant typically produces clearer spectral changes than the gradient formant transitions at vowel onset, particularly following sonorant consonants (cf. Goldstein et al., 2007, who noted that vowel onsets were difficult to measure reliably even in articulatory data). Following Durvasula et al. (2021), I therefore use vowel offset as the anchor.
Figure 2 and Figure 3 show sample annotations of four items with a medial cluster and geminate with these temporal landmarks indicated.
Figure 2: Sample annotations showing c-centre and right-edge landmarks for two tokens with a medial cluster: [rakba] ‘riding’ (left) and [raʒma] ‘stoning’ (right).
![]()
Figure 3: Sample annotations showing c-centre and right-edge landmarks for two tokens with a medial geminate: [kaʃʃa] ‘blanket’ (left) and [ʕamma] ‘general’ (right).
![]()
The two intervals — c-centre-to-anchor and right-edge-to-anchor — were compared across conditions in order to examine temporal stability. The c-centre-to-anchor interval was considered stable if its duration showed no significant variation across singleton and geminate contexts. Such a pattern would be consistent with a tautosyllabic analysis of geminates, under which the geminate is treated as a single long unit whose midpoint behaves like the c-centre of a complex onset. On the other hand, the right-edge-to-anchor interval was considered stable if its duration is consistent across medial singletons and their corresponding geminates. This pattern would be compatible with a heterosyllabic analysis, in which only the second half of the geminate aligns as the onset of the following syllable. The same logic applies to the comparison between singleton and cluster conditions.
To quantify the stability of each interval, the mean duration, standard deviation (SD), and relative standard deviation (RSD) were calculated for each pair of items for each participant. The RSD was calculated using the formula in (1). Following Shaw et al. (2009, 2011) and Durvasula et al. (2021), I use RSD as a measure of stability, since it addresses the inherent bias in using SD when comparing intervals of different lengths. This is relevant here since the c-centre-to-anchor interval is inherently longer than the right-edge-to-anchor interval. RSD normalizes the variance, which results in a more robust measure of stability. Lower RSD values indicate more temporal stability.
(1)
RSD=(SDMean)×100
\documentclass[10pt]{article}
\usepackage{wasysym}
\usepackage[substack]{amsmath}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage[mathscr]{eucal}
\usepackage{mathrsfs}
\usepackage{pmc}
\usepackage[Euler]{upgreek}
\pagestyle{empty}
\oddsidemargin -1.0in
\begin{document}
\[
{\rm RSD}=\left(\frac{\rm SD}{\rm Mean}\right)\times 100
\]
\end{document}
![]()
Stability was assessed by comparing RSD values across VCiV–VCiCiV pairs and VCjV–VCiCjV pairs.
Two mixed-effects linear regressions were conducted to test whether the right-edge-to-anchor interval was significantly more stable than the c-centre-to-anchor interval and whether the stability patterns differed between geminates and clusters. The regressions were performed in R (R Development Core Team, 2014) using the lme4 package (Bates et al., 2015). General data analysis, and plotting were performed using Python scripts. For extracting measurments for TextGrid files, I used the praatio library (Mahrt, 2016).2
For the first regression model, which examines the stability difference between the two intervals, the dependent variable was RSD. The independent variable was Interval (c-centre-to-anchor vs. right-edge-to-anchor, with c-centre-to-anchor as the baseline). Participant and set were included as random intercepts, along with a by-participant random slope for interval type. For the second regression model, which compares the temporal stability patterns of medial geminates and clusters, the dependent variable was the RSD values of the right-edge-to-anchor interval. The independent variable was Sequence Type (geminate or cluster, with cluster as the reference level). Participant and Pair were included as random intercepts, and a by-participant random slope for Sequence Type was also included. In addition to these two main regression models, I also fitted two regression models to test the effect of sonority profile on temporal stability patterns and the shortening of the vowel preceding geminates and clusters. Significance was determined by comparing the p-values of the fixed effects to an alpha level, with results showing significance when p < 0.05.
3 Results
A total of 6 out of 1000 token productions (0.6%) were annotated as ‘bad’ and discarded, as they exhibited acoustic artifacts, possibly caused by the speaker being too close to the microphone, which made identifying acoustic landmarks impossible. In what follows, I present the results of the temporal-stability patterns for medial geminates and clusters. In both cases, it will be shown that the right-edge-to-anchor interval is more stable than the c-centre-to-anchor interval, supporting a heterosyllabic representation. I also show that the sonority profile of clusters does not affect temporal stability patterns. Finally, I report the results of an analysis of preceding vowel duration, showing that the duration of the preceding vowel is not significantly affected by the presence of a geminate or cluster.
3.1 Medial clusters
Figure 4 plots overall raw durations of the two intervals across the two sequence types. The c-centre-to-anchor interval is systematically longer in VCiCjV compared to VCjV, whereas the right-edge-to-anchor interval shows no increase in raw durations between the two conditions.
Figure 4: Overall raw durations for the c-centre and right-edge intervals (VCjV vs. VCiCjV). Center lines indicate medians, box edges show 25th and 75th percentiles, and whiskers extend to 1.5 × interquartile range with outliers plotted individually.
![]()
Inter-speaker variability was assessed by examining participant-wise patterns. The overall pattern observed in Figure 4 holds participant-by-participant as shown in Figure 5; the right-edge-to-anchor interval remains stable across medial sequences regardless of sequence size, while the c-centre-to-anchor interval duration does not show the same stability.
Figure 5: Participant-wise raw durations for the two intervals (VCjV vs. VCiCjV). Box plots as in Figure 4.
![]()
To further evaluate the temporal stability of the two intervals, the overall RSD values were calculated for each interval. As shown in Figure 6, the lower RSD values for the right-edge-to-anchor interval indicate more stability compared to the c-centre-to-anchor interval.
Figure 6: RSD distributions for c-centre and right-edge intervals (VCjV vs. VCiCjV). Violin plots show the probability density of RSD values across participants, where wider sections indicate more frequently observed values. Box plots are overlaid showing medians (center lines), quartiles (box edges), and whiskers extending to 1.5 × interquartile range.
![]()
To assess whether the right-edge stability pattern was consistent across participants, participant-wise RSD values were examined. As shown in Figure 7, all ten participants exhibited lower RSD values for the right-edge-to-anchor interval compared to the c-centre-to-anchor interval, confirming that the pattern is robust across the sample rather than driven by a subset of speakers.
Figure 7: Participant-wise RSD distributions for c-centre and right-edge intervals (VCjV vs. VCiCjV). Violin plots as in Figure 6.
![]()
Table 3 presents the mean, SD, and RSD for each interval across the five word pairs. The values with the lowest RSD in each set are highlighted in green. The Right-edge-to-anchor interval exhibits the lowest RSD across all word pairs.
I fit a mixed-effects linear regression model to determine whether the right-edge-to-anchor interval exhibited significantly greater temporal stability (lower RSD values) than the c-centre-to-anchor interval. The dependent variable was RSD, and the independent variable was Interval (c-centre-to-anchor vs. right-edge-to-anchor, with c-centre-to-anchor as the baseline). Participant and set were included as random intercepts, along with a by-participant random slope for interval type.
As shown in Table 4, the model shows a significant negative estimate for the Interval (Right Edge) predictor (β = –6.86, p = 0.009). Given that the c-centre-to-anchor interval served as the baseline for comparison, this negative estimate confirms that the right-edge-to-anchor interval has significantly lower RSD values, indicating greater temporal stability.
Table 3: Mean, SD and RSD for clusters. Lowest RSD in each set highlighted in green.
|
C-Centre |
Right Edge |
| Set |
Mean |
SD |
RSD |
Mean |
SD |
RSD |
| baqa ∼ sabqa |
138.3 |
20.7 |
15.1 |
117.9 |
15.7 |
13.3 |
| sala ∼ makla |
117.7 |
22.1 |
19.1 |
98.7 |
13.0 |
13.1 |
| nuba ∼ rakba |
129.6 |
23.2 |
18.2 |
109.5 |
11.2 |
10.3 |
| mama ∼ raʒma |
123.6 |
28.3 |
22.6 |
104.4 |
12.3 |
11.5 |
| ʕana ∼ baʕna |
121.2 |
27.5 |
22.9 |
103.1 |
15.7 |
15.3 |
| Average |
126.1 |
24.4 |
19.6 |
106.7 |
13.6 |
12.7 |
Table 4: Mixed-effects model for clusters.
| Fixed effect |
β |
SE |
t |
p |
| Intercept (C-Centre) |
19.56 |
1.53 |
12.72 |
<0.001 |
| Interval (Right Edge) |
–6.86 |
1.54 |
–4.43 |
0.009 |
Figure 8: RSD distributions for c-centre and right-edge intervals by sonority profile. Left: plateau sonority clusters (stop-stop). Right: rising sonority clusters (obstruent-sonorant). Violin plots as in Figure 6.
![]()
To assess whether sonority profile affects temporal stability patterns, an additional analysis was conducted following the approach of Durvasula et al. (2021), who tested sonority effects in Jazani Arabic word-initial clusters. The clusters were coded as plateau (stop-stop: [bq], [kb]) or rising sonority (obstruent-sonorant: [kl], [ʒm], [ʕn]). A mixed-effects model was fit with RSD as the dependent variable and fixed effects for Interval (C-Center vs. Right Edge), Sonority Profile (plateau vs. rising), and their interaction. The model included random intercepts for Participant.
Figure 8 shows the RSD distributions for plateau and rising sonority clusters. As shown in Table 5, there was a significant main effect of Sonority Profile (β = 4.90, p < 0.001), with rising sonority clusters having higher overall RSD than plateau clusters. More importantly, the Interval × Sonority Profile interaction was significant (β = –3.39, p = 0.035). This interaction suggests that the right-edge stability advantage may be larger for rising sonority clusters than for plateau clusters. For plateau clusters, the right-edge interval was more stable than the c-center interval by 4.8% (16.6% vs. 11.8% RSD). For rising sonority clusters, the right-edge advantage was 8.2% (21.5% vs. 13.3% RSD). Therefore, while both cluster types show right-edge stability, the pattern appears to be more pronounced for rising sonority clusters. These findings, however, should be interpreted cautiously given that the stimulus set was not designed to systematically test sonority effects.
Table 5: Mixed-effects model for clusters with sonority profile. The intercept represents the condition with C-Center interval and plateau sonority.
| Fixed effect |
β |
SE |
t |
p |
| Intercept (C-Center, Plateau) |
16.62 |
0.98 |
16.89 |
<0.001 |
| Interval (Right Edge) |
–4.83 |
1.26 |
–3.83 |
0.0003 |
| Sonority Profile (Rising) |
4.90 |
1.12 |
4.39 |
<0.001 |
| Interval × Sonority Profile |
–3.39 |
1.58 |
–2.15 |
0.035 |
3.2 Medial geminates
With respect to medial geminates, Figure 9 shows that having medial geminates lengthens the raw duration of c-centre-to-anchor interval compared to singletons, while the duration of right-edge-to-anchor interval remains stable across both conditions. The participant-by-participant plots shown in Figure 10 reflect the same trends.
Figure 9: Overall raw durations for the two intervals (singleton vs. geminate). Box plots as in Figure 4.
![]()
Figure 10: Participant-wise raw durations for singleton vs. geminate pairs. Box plots as in Figure 4.
![]()
Figure 11 and Figure 12 show the overall and participant-wise RSD values, respectively, and Table 6 reports means, SDs and RSDs for each of the five word pairs. Similar to the results for clusters, the right-edge-to-anchor interval for geminates has lower RSDs, confirming its greater stability.
Figure 11: RSD distributions for c-centre and right-edge intervals (VCiV vs. VCiCiV). Violin plots as in Figure 6.
![]()
Figure 12: Participant-wise RSD distributions for c-centre and right-edge intervals (VCiV vs. VCiCiV). Violin plots as in Figure 6.
![]()
Table 6: Mean, SD and RSD for singleton-geminate pairs. Lowest RSD in each set highlighted in green.
|
C-Centre |
Right Edge |
| Pair |
Mean |
SD |
RSD |
Mean |
SD |
RSD |
| sˤarˤa ∼ dˤarˤrˤa |
105.4 |
24.6 |
23.9 |
95.1 |
15.4 |
16.7 |
| ʕuma ∼ ʕamma |
120.3 |
19.6 |
16.5 |
103.6 |
12.7 |
12.1 |
| kala ∼ lalla |
114.4 |
24.7 |
21.7 |
100.2 |
15.0 |
15.0 |
| sˤatˤa ∼ ħatˤtˤa |
126.4 |
19.7 |
15.6 |
108.2 |
12.2 |
11.3 |
| riʃa ∼ kaʃʃa |
142.8 |
16.9 |
11.9 |
122.2 |
14.2 |
11.8 |
| Average |
121.9 |
21.1 |
17.9 |
105.9 |
13.9 |
13.4 |
I fit a mixed-effects model identical to the one used in analyzing the results for clusters. As shown in Table 7, the results show a significant negative estimate for the Interval (Right Edge) predictor (β = –4.56, p = 0.029). This confirms that the right-edge-to-anchor interval has significantly lower RSD values, indicating greater temporal stability for the geminate pairs.
To further confirm these results, I conducted a third mixed-effects linear regression analysis where I compared the RSD values of the right-edge-to-anchor interval of medial geminates to those of medial clusters. The dependent variable was the RSD values of the right-edge-to-anchor interval, and the independent variable was Sequence Type, i.e., whether the medial sequence is a geminate or a cluster, with the latter being the reference level. Participant and Pair were included as random intercepts, and a by-participant random slope for Sequence Type was also included. As shown in Table 8, the model showed no significant effect of Sequence Type. In other words, there is no evidence for a difference in temporal stability between medial geminates and clusters (p = 0.473).
Table 7: Mixed-effects model for geminate pairs.
| Fixed effect |
β |
SE |
t |
p |
| Intercept (C-Centre) |
17.94 |
2.31 |
7.75 |
<0.001 |
| Interval (Right Edge) |
–4.56 |
1.46 |
–3.10 |
0.029 |
Table 8: Mixed-effects linear regression model results for geminates vs. clusters.
| Fixed effect |
β |
SE |
t |
p |
| Intercept (cluster) |
12.70 |
1.01 |
12.47 |
<0.001 |
| Sequence Type (geminate) |
0.67 |
0.93 |
0.72 |
0.473 |
These results indicate that medial geminates show temporal stability patterns that are comparable to those of medial clusters. Taken together, the results support a heterosyllabic representation for both medial geminates and clusters in Moroccan Arabic.
3.3 Preceding vowel duration analysis
Previous studies have used cues like shortening of the preceding vowel as an indicator that geminates and clusters have a heterosyllabic representation (Esposito & Di Benedetto, 1999; Ridouane, 2010; Khattab & Al-Tamimi, 2014; among others). I examined whether this pattern holds for Moroccan Arabic medial geminates and clusters. As shown in Figure 13 and Table 9, the duration of the preceding vowel was shorter before both clusters and geminates compared to their singleton counterparts. For clusters, the vowel in VCiCjV sequences (105.4 ms) was slightly shorter than in VCjV sequences (108.3 ms), a difference of –2.9 ms. For geminates, the vowel before the geminate (104.9 ms) was noticeably shorter than before the singleton (115.8 ms), a difference of –10.9 ms.
Figure 13: Preceding vowel duration for clusters (left) and geminates (right).
![]()
Table 9: Mean and standard deviation of preceding vowel duration in VCV and VCCV contexts.
|
Clusters |
Geminates |
| Condition |
Mean |
SD |
Mean |
SD |
| VCV |
108.3 |
24.1 |
115.8 |
23.6 |
| VCCV |
105.4 |
23.8 |
104.9 |
27.2 |
To test whether vowel duration differs systematically between singleton and geminate/cluster contexts, I fit separate linear mixed-effects models for clusters and geminates. The dependent variable was the duration of the preceding vowel (in ms). The fixed effect was Condition, coded as a factor with two levels: VCV (singleton baseline) and VCCV (geminate or cluster). Random intercepts for Participant and Item were included in both models. The results are summarised in Table 10. For clusters, the estimated difference between the VCCV and VCV conditions was (β = –3.45 ms), which was not statistically significant (t(8.00) = –0.437, p = 0.674). For geminates, the estimated difference was (β = –10.92) ms, which also did not reach significance (t(8.00) = –0.942, p = .374). Although both estimates are negative, suggesting a trend of vowel shortening before clusters and geminates, neither effect is statistically reliable.
Table 10: Mixed-effects models for preceding vowel duration in cluster and geminate contexts. The baseline (Intercept) corresponds to the singleton (VCV) condition.
| Fixed effect |
β |
SE |
t |
p |
| Clusters |
| Intercept (VCV) |
108.28 |
6.66 |
16.26 |
<0.001 |
| Condition (VCCV) |
–3.45 |
7.89 |
–0.44 |
0.674 |
| Geminates |
| Intercept (VCV) |
115.76 |
8.79 |
13.18 |
<0.001 |
| Condition (VCCV) |
–10.92 |
11.59 |
–0.94 |
0.374 |
In summary, the vowel-duration analysis does not provide statistically significant evidence for vowel shortening before medial clusters or geminates in Moroccan Arabic. Therefore, vowel duration does not offer additional support for a heterosyllabic representation from this phonetic cue.
4 Discussion and conclusion
This study investigated whether temporal stability patterns could be used to probe the syllabification of geminates, using Moroccan Arabic as a case study. The results show that both medial clusters and geminates in Moroccan Arabic exhibit a right-edge-to-anchor stability pattern, which indicates that a medial geminate is split into two units belonging to separate syllables. The results also show that there is no evidence for a significant difference between the temporal stability patterns of medial geminates and clusters, which suggests that both are syllabified similarly in Moroccan Arabic. More importantly, these findings suggest that this methodology can be used to test predictions of competing phonological hypotheses about the syllabification of geminates. This is supported by the fact that the temporal stability results align with established independent phonological evidence about the syllabification of medial clusters and geminates in Moroccan Arabic including versification and poetic meter (Dell & Elmedlaoui, 2002), phonotactic constraints such as the behavior of the schwa (Al Ghadi, 1990; Boudlal, 2001; Bensoukas & Boudlal, 2012; Noamane, 2021), and prosodic constraints such as foot binarity (Al Ghadi, 1994; Jebbour, 1996; Boudlal, 2001). The convergence of evidence from temporal stability with such independent phonological evidence supports the methodology as a useful diagnostic tool for the syllabification of geminates.
The findings of this study are compatible with a number of theoretical frameworks including Moraic Theory (Hayes, 1986, 1989; McCarthy & Prince, 1995; Davis, 1999), Prosodic Length (Skeletal/CV) analyses (McCarthy, 1979b; Leben, 1980; Tranel, 1991), and Two-Root Node (Segmental Length) analyses (McCarthy, 1988; Selkirk, 1990; Ringen & Vago, 2011); under these frameworks, a geminate has a heterosyllabic representation where it occupies two prosodic or timing units. On the other hand, the results are not consistent with approaches such as the Moraic Onset hypothesis, which treat geminates as tautosyllabic mono-segmental units (Mohanan, 1989; Topintzi, 2008, 2022).
The sonority analysis suggested that while both plateau and rising sonority clusters show right-edge stability, the pattern may be more pronounced for rising sonority clusters, though this result is not definitive given the post-hoc nature of the analysis and the limited number of cluster types. This finding has several implications. First, it contrasts with languages like Italian, where sonority profile can determine whether clusters show c-center or right-edge stability (Hermes et al., 2013). In Italian, sibilant-stop clusters (which violate sonority sequencing) show right-edge stability, while obstruent-liquid clusters (with rising sonority) show c-center stability. In Moroccan Arabic, both sonority profiles show right-edge stability, consistent with the results of Durvasula et al. (2021) for word-initial clusters in Jazani Arabic. On the other hand, this contrasts with Durvasula et al.’s 2021 results, where no sonority effects were observed. Interestingly, the apparent direction of the sonority effect in Moroccan Arabic—with rising sonority clusters showing stronger right-edge stability—is opposite to what might be expected if sonority distance facilitated tighter coordination (Hermes et al., 2013). However, it is important to note that the stimulus set did not include falling sonority clusters, and the number of cluster types was limited. Future work should include the full range of sonority profiles to fully characterize the role of sonority in the temporal coordination of Moroccan Arabic consonant sequences.
Some phonetic studies have inferred a heterosyllabic representation for geminates based on phonetic cues. Most of these accounts use shortening of the vowel preceding a geminate, relative to a singleton, as evidence that the first part of the geminate functions as a coda of the preceding syllable. Such conclusions have, for instance, been drawn for Tashlhiyt Berber (Ridouane, 2010), Lebanese Arabic (Khattab & Al-Tamimi, 2014), Italian (Esposito & Di Benedetto, 1999), and Japanese (Smith, 1995; Burroni et al., 2025). In contrast, the present study found no significant vowel shortening before Moroccan Arabic medial geminates or clusters, suggesting that vowel duration may not be a reliable cue to syllabification in this language. This aligns with cross-dialectal variation within Arabic, where phonetic cues to gemination, for instance, can differ substantially (Khattab & Al-Tamimi, 2014).
Other studies report phonetic patterns that are compatible with a heterosyllabic parse, although they do not explicitly address the issue of syllabification. For example, Zeroual et al. (2015) show, using EMA data from Moroccan Arabic, that medial geminate plosives are produced as two overlapping consonantal gestures, as their closing and opening movements share similar temporal and kinematic properties with those of clusters. In Italian, Di Benedetto et al. (2021) show that medial geminated plosives, in some cases, exhibit double bursts, and that the second part of a medial geminate is often produced with more stability and strength than the first. In addition, a number of studies report geminate/singleton duration ratios exceeding 2 in languages such as Turkish (Lahiri & Hankamer, 1988), Pattani Malay (Abramson, 1987), Japanese (Kawahara, 2005), Persian (Hansen, 2004), and Polish (Rojczyk & Porzuczek, 2019). Although the findings of these studies are consistent with the view that geminates occupy two timing units, they rely on indirect phonetic cues and are not specifically designed to test competing syllabification parses. By contrast, the temporal stability approach used in this paper provides a more direct diagnostic of internal coordination, i.e., how the two parts of a geminate are timed relative to each other and to adjacent vowels, which allows the syllabification of geminates to be evaluated empirically rather than inferred from secondary cues.
It should be noted that the temporal stability methodology only partially identifies syllabification patterns when a right-edge stability pattern is observed. While a c-centre stability pattern for medial geminates would clearly indicate that the geminate is syllabified as an onset of the following vowel, a right-edge stability pattern leads to a more limited conclusion: the first timing unit of the geminate cannot be syllabified as an onset of the following vowel, and the second timing unit must be an onset of the following vowel. While we can determine that the first timing unit of the geminate is not in the onset of the following vowel, the temporal stability method does not determine whether it is a coda of the preceding vowel (VCi.CiV) or forms a degenerate syllable (V.Ci.CiV).
Several questions arise from this study that deserve future investigation. First, it remains to be seen whether we should always expect geminates to pattern similar to clusters in terms of their syllabification or temporal stability patterns. For instance, in languages with c-center-to-anchor intervals for clusters (e.g., Italian, Polish), would their geminates also exhibit a c-center-to-anchor pattern? Second, this study focused on lexical (underlying or contrastive) geminates, but other types of geminates, such as assimilated and concatenated (also called ‘fake’) geminates, also deserve future investigation. Although previous acoustic studies suggest little phonetic difference between these types of geminates (Ridouane, 2010; Kotzor et al., 2016), their temporal stability patterns might still be different.
Finally, the position of geminates within words could influence their phonological representation. Previous work on Arabic dialects, for instance, have shown that word-final geminates are, like word-medial ones, argued to be moraic (Davis & Ragheb, 2014). Word-initial geminates, on the other hand, may have a different representation. Ham (2001) showed that word-initial geminates in Bernese have different durational properties from word-medial and word-final geminates, suggesting a non-moraic onset representation for initial geminates. These possible differences between geminates in each position could be further investigated using the temporal stability methodology. Moroccan Arabic would be an ideal test case for future studies, since it exhibits all three types of geminates (lexical, assimilated, and concatenated) in all positions (Noamane, 2019, 2020, 2021), which allows for a comprehensive investigation of these questions.