Proficiency, phonological context, and distributional control in Arabic L1 learners' English stop voicing

1 Introduction

Voice onset time (VOT) remains one of the most informative acoustic dimensions for studying stop voicing because it links fine-grained laryngeal timing to phonological contrast, cross-linguistic typology, and second-language speech learning. The basic typological distinction among prevoicing, short-lag, and long-lag categories is well established, but current work increasingly treats VOT categories as distributions rather than as single mean targets (Lisker and Abramson, 1964; Cho and Ladefoged, 1999; Abramson and Whalen, 2017; Cho et al., 2019). This distributional view is especially important in L2 phonetics because a learner may show the expected direction of a contrast while still retaining excessive dispersion, unstable category boundaries, or substantial cross-category overlap.

The speech learning model (SLM) and its revised version, SLM-r, provide the main theoretical frame for the present study. The original model proposes that L2 sounds are learned in relation to perceived similarity with established L1 categories; similar sounds may undergo equivalence classification, whereas sufficiently new sounds may support the formation of distinct L2 categories (Flege, 1987, 1995). The SLM-r retains this contrastive logic while emphasizing lifelong plasticity, bidirectional interaction between L1 and L2 phonetic subsystems, and distributional learning from input (Flege and Bohn, 2021). The Perceptual Assimilation Model and its L2 extension also place cross-language similarity at the center of learning, especially in the perceptual assimilation of unfamiliar contrasts to native categories (Best et al., 2001; Best and Tyler, 2007). Together, these models motivate an analysis that connects production distributions, perceptual boundary placement, and learner experience without assuming that any single measure directly exhausts phonetic category learning.

Arabic-English stop voicing provides a strong empirical test of these claims. Earlier Arabic-English work showed that Arabic learners encounter difficulty with English long-lag aspirated voiceless stops because Arabic voiceless stops usually occupy a short-lag region, while voiced stops often exhibit robust prevoicing (Flege and Port, 1981). Subsequent work on Arabic and Arabic–English speech has documented both broad cross-variety similarities and important dialect-specific differences (Yeni-Komshian et al., 1977; Khattab, 2002; Al-Tamimi et al., 2021; Al-Tamimi and Khattab, 2018; Kulikov, 2022; Al-Deaibes et al., 2025, 2026; Aldamen et al., 2026a,b). For Jordanian Arabic, Mitleb (2001) is directly relevant because the present learner group consisted of Jordanian Arabic L1 speakers. The Arabic baseline data in the present study therefore provide a necessary local check on the assumed L1 phonetic system rather than relying only on pan-Arabic generalizations.

Arabic phonotactics also matter for the /sp/ condition. It is not sufficient to state that Arabic lacks onset clusters categorically, because Arabic dialects differ in how clusters are licensed, repaired, or syllabified. Work on Arabic syllabification shows that dialectal structure and moraic organization condition the treatment of long segments and consonant sequences, while work on Ammani Arabic loanword phonology illustrates how English forms can be adapted within local phonological constraints (Watson, 2007; Abu Guba, 2021). The present study therefore treats English /sp/ clusters as a context in which learners must coordinate two learning problems: the English-specific phonetic target for /p/ and the context-conditioned suppression of aspiration after /s/.

English contextual VOT patterns make this coordination nontrivial. In stressed singleton onsets, General American English /p/ is normally realized with long-lag positive VOT, whereas /b/ is short-lag or near-zero. After /s/, however, English voiceless stops show markedly shorter VOT, and learners who overgeneralize aspiration to /sp/ produce a pattern that is phonetically interpretable but contextually non-native (Klatt, 1975; Lisker and Abramson, 1967; Mielke and Nielsen, 2018). Similar L2 difficulties with /s/-stop clusters have been documented in other language pairs, suggesting that aspiration control involves more than a global increase in voiceless-stop VOT (Flege and Eefting, 1987; Schwartz, 2023).

A distributional approach is required because group means may hide the structure of learning. Talkers differ systematically in VOT implementation, and stop categories show covariation within and across speakers (Davidson, 2016; Chodroff and Wilson, 2017; Chodroff et al., 2019). In L2 speech, individual differences in phonological processing, cue use, and experience further shape how learners use acoustic dimensions in production and perception (Darcy et al., 2015; Schertz et al., 2015; Gorba and Cebrian, 2021). These issues are also central to pronunciation research more generally, where L2 proficiency, intelligibility, accentedness, and pronunciation development do not map onto one another in a simple one-to-one manner (Munro and Derwing, 1995; Piske et al., 2001; Major, 2001; Saito and Plonsky, 2019).

The present study tests whether proficiency-related learning in Arabic L1 learners' English stop voicing is better captured by the distributional organization of VOT than by mean values alone. Proficiency is treated as a continuous predictor based on a non-pronunciation placement test, which avoids direct circularity with VOT production but remains an imperfect proxy for phonetic experience (Hulstijn, 2012). Following exemplar and statistical-learning accounts of phonology, the main prediction is that learning will be expressed through gradually reorganized distributions and context-conditioned control rather than abrupt category replacement (Pierrehumbert, 2003). The analysis addressed four questions:

• Do Arabic L1 learners of English differ from General American English controls in the mean and distributional properties of singleton /p/ and /b/ VOT?
• Does English proficiency predict longer singleton /p/ VOT, reduced within-speaker dispersion, and reduced /p/-/b/ overlap?
• Does phonological context modulate /p/ realization, and does proficiency predict stronger aspiration suppression in /sp/ clusters?
• Do production-based distributional patterns and perception-probe results converge in showing proficiency-related reduction in /p/-/b/ category competition and improved context-sensitive aspiration control?

2 Methods

2.1 Participants and language background

Sixty-five adults participated. The learner group consisted of 45 Jordanian Arabic L1 speakers of English (23 women, 22 men; age M = 20.58 years, SD = 1.49, range = 18.1–23.8). All learners reported Jordanian Arabic as their L1, formal-schooling exposure to English in Jordan, no residence in an English-speaking country, normal hearing, and no history of speech or language disorder. Their Jordanian Arabic backgrounds were Ammani urban (n = 19), Central Jordanian (n = 7), Urban Palestinian-Jordanian (n = 7), North Jordanian (n = 7), and Southern Jordanian (n = 5). Learners reported low acceptability of complex Arabic onset clusters on a 1–5 scale (M = 1.59, SD = 0.45), consistent with the view that onset-cluster treatment is a phonotactic issue for these learners rather than a uniform property of all Arabic varieties. The control group consisted of 20 native speakers of General American English (11 women, 9 men; age M = 22.71 years, SD = 1.86, range = 19.3–26.0). Controls reported no knowledge of Arabic or other prevoicing-based stop systems. Both groups were recorded in the same sound-treated university phonetics laboratory; the phrase “same university” refers to the recording and recruitment site, while the target English variety for acoustic comparison was General American English. Participant characteristics are summarized in Table 1.

Table 1. Participant, proficiency, exposure, and variety information.

As shown in Table 1, the learner sample was deliberately homogeneous in broad learning context but heterogeneous in proficiency. Learner dialect backgrounds were: Ammani urban = 19, Central Jordanian = 7, Urban Palestinian-Jordanian = 7, North Jordanian = 7, Southern Jordanian = 5. Their main English input profiles were: American English media = 25, Mixed English media = 14, British English classroom materials = 6.

2.2 Proficiency measure

English proficiency was measured with the university language center placement test. The test included reading, listening, and grammar components scored from 0 to 100 and did not include a speaking or pronunciation component. The composite proficiency score was the mean of the three component scores (M = 60.96, SD = 12.58, range = 36.80–90.00), and the reported internal reliability for the test version was alpha =0.86. CEFR-equivalent classifications were A2 (n = 5), B1 (n = 17), B2 (n = 17), and C1 (n = 6). The absence of a pronunciation component reduces circularity with the acoustic dependent variables, although the construct should be understood as general English proficiency rather than a direct measure of phonetic experience (Hulstijn, 2012).

2.3 Stimuli and tasks

The English production task used real words in three target conditions: singleton /b/, singleton /p/, and /p/ in /sp/ clusters. Each word appeared in the carrier phrase Say—again and was repeated five times, producing 110 English reading trials per participant. The full English stimulus list appears in Appendix A, Table A1. The design also included an Arabic baseline production task for learners, using Jordanian Arabic words with initial /b/, /t/, and /k/; these items appear in Appendix B, Table B1. The Arabic baseline was included to document the learners' local L1 VOT system and to avoid treating Arabic as a phonetically uniform system. Table 2 summarizes the production conditions and usable token counts.

Table 2. Production tasks, target conditions, and usable VOT tokens.

Two perception probes were included. First, a /b/-/p/ identification task presented a synthetic continuum from −20 to 80 ms VOT in 10 steps, with four repetitions per step for each participant (2,600 trials). The derived boundary was the VOT value at which /p/ responses began to dominate the response function. Second, a context-naturalness task presented singleton /p/ and /sp/ words with contextually appropriate and overaspirated VOT values; participants rated naturalness from 1 to 7 (1 = very unnatural, 7 = very natural). This task contained 16 trials per participant (1,040 trials). These perception probes were used to triangulate production patterns, not to replace the production analysis as the primary outcome.

2.4 Recording, acoustic measurement, and reliability

Recordings were made in a sound-treated phonetics laboratory using a head-mounted condenser microphone positioned at a consistent mouth-to-microphone distance. Speech was recorded directly in Praat at 44.1 kHz and 16-bit resolution. No downsampling was applied for VOT measurement. Archival segmented copies were peak-normalized to −3 dBFS, but all VOT measurements were made on the unaltered waveform.

VOT was defined as the interval between the stop release burst and the onset of periodic voicing; negative values were assigned when voicing preceded the burst. Measurements were made manually in Praat from the waveform and wideband spectrogram. Tokens were excluded when unclear burst or voicing boundaries, background noise, or creaky onset prevented reliable measurement. Across the English reading data, exclusions were low: 6–27 tokens per group-by-condition cell, with rates from 0.75% to 2.50%. The most common exclusion reason was unclear voicing onset (n = 38), followed by ambiguous burst (n = 28), creaky voice onset (n = 25), and background noise (n = 9).

A second trained phonetician remeasured 1,762 tokens, representing 24.99% of included tokens. Inter-rater agreement for VOT was high, ICC(A,1) =0.94; the mean absolute difference was 6.03 ms and the median absolute difference was 4.96 ms. Additional double-coding was used for /sp/ cluster repair coding (Cohen's kappa = 0.81) and acoustic usability coding (kappa = 0.83).

2.5 Distributional and statistical measures

For each participant and condition, the analysis computed mean VOT, within-speaker standard deviation, coefficient of variation, the overlap area between kernel density estimates of singleton /p/ and singleton /b/, and the proportion of singleton /p/ tokens falling inside the participant-specific /b/ VOT range. The overlap index ranges from 0 to 1, with higher values indicating greater production-level distributional overlap. Kernel density estimates were computed separately for each participant's singleton /p/ and singleton /b/ VOT distributions using Gaussian kernels. The overlap index was calculated as the shared area under the two estimated density curves across the observed VOT range. The proportion-in-range measure was included as a simpler token-level check on whether /p/ productions intruded into a speaker's /b/ range. These measures separate category placement from category precision and cross-category separation, which is central to the distributional interpretation of SLM-r (Flege and Bohn, 2021; Pierrehumbert, 2003).

Inferential models were fitted in R with lme4 (Bates et al., 2015). Mixed-effects models included random intercepts for participants and lexical items wherever repeated tokens were analyzed. Random slopes were evaluated according to a predefined rule: theoretically justified slopes were attempted first; slopes were retained when the model converged without singular fit and removed when they produced singularity or non-convergence. This approach balances Type I error control and power in mixed-effects modeling (Barr et al., 2013; Matuschek et al., 2017). For the context-by-proficiency model, a context-by-participant random slope was attempted but not retained because the sensitivity run produced a singular fit; participant and item intercepts were retained. Confidence intervals and standardized effect sizes are reported to support interpretation beyond significance testing (Lakens, 2013). A power-sensitivity check indicated that the observed design, with 45 learners, 20 controls, and 110 reading tokens per participant, had approximately 80% power to detect standardized effects of about 0.42 in the group design and about 0.39 in learner-only proficiency models, which is appropriate for medium effects but not for small effects in mixed models (Kumle et al., 2021).

Task-style checks used local speech rate from the carrier phrase and semi-spontaneous speech samples from a picture-description and short-interview task. These checks do not convert the controlled reading task into natural speech, but they allow the interpretation of reading-task VOT to be qualified against speech rate and semi-spontaneous production.

3 Results

3.1 Descriptive production patterns

Table 3 presents participant-level VOT descriptives and distributional indices. Figure 1 visualizes the token-level VOT distributions by group and context. Controls produced the expected General American English pattern: short-lag singleton /b/, long-lag singleton /p/, and short-lag /sp/ cluster /p/. Learners also distinguished singleton /b/ and /p/, but their singleton /p/ values were shorter, more dispersed, and more overlapped with /b/ than controls' productions. Learner /sp/ cluster VOT values were shorter than learner singleton /p/ values but substantially longer than control /sp/ values, indicating incomplete aspiration suppression in the cluster context.

Table 3. Participant-level VOT means, variability, and distributional overlap by group.

3.2 Proficiency, singleton /p/ VOT, and within-speaker dispersion

The learner-only model for singleton /p/ showed a positive proficiency effect on mean VOT: each one-point increase on the 0–100 proficiency scale predicted a 0.42 ms increase in singleton /p/ VOT, 95% CI [0.24, 0.60], standardized beta = 0.49 (Table 4). Across the observed proficiency range, this corresponds to a predicted increase of approximately 22 ms, a difference large enough to be phonetically meaningful for the acquisition of long-lag aspiration. Figure 2 shows the participant-level relationship between proficiency and singleton /p/ mean VOT.

Table 4. Proficiency-related model estimates for production outcomes.

Proficiency also predicted within-speaker dispersion. Each one-point increase in proficiency predicted a 0.19 ms decrease in the within-speaker standard deviation of singleton /p/ VOT, 95% CI [−0.27, −0.11], standardized beta = −0.53. Thus, learning was expressed not only as movement toward a longer-lag /p/ target but also as tighter and more stable production. Figure 3 shows this reduction in dispersion across the proficiency continuum.

3.3 Distributional overlap between singleton /p/ and /b/

Learners showed substantially greater /p/-/b/ distributional overlap than controls (Table 3). The learner group mean overlap index was 0.29, compared with 0.06 for controls, and the proportion of /p/ tokens falling inside the speaker-specific /b/ range was 0.27 for learners, compared with 0.04 for controls. Proficiency predicted lower overlap: each one-point increase in proficiency predicted a 0.006 decrease in the overlap index, 95% CI [−0.008, −0.004], standardized beta = −0.57 (Table 4). A parallel token-level intrusion measure showed the same direction, with proficiency negatively correlated with the proportion of /p/ tokens inside the speaker-specific /b/ range, r = −0.39.

Figure 4 visualizes the participant-level overlap pattern. Higher-proficiency learners generally had better-separated singleton /p/ and /b/ distributions, although residual overlap remained for several speakers.

3.4 Phonological context and aspiration suppression in /sp/ clusters

The context model examined singleton /p/ and /sp/ cluster /p/ in one model rather than separate tests. Learners produced shorter VOT in /sp/ clusters than in singleton onsets, but the magnitude of aspiration suppression varied with proficiency. As shown in the final row of Table 4, the context-by-proficiency interaction was negative, estimate = −0.21, 95% CI [−0.35, −0.07], standardized beta = −0.32. This interaction indicates that higher proficiency predicted stronger context-conditioned suppression in /sp/ clusters beyond the general increase in singleton /p/ aspiration.

Figure 5 visualizes this diverging developmental pattern. Singleton /p/ increased with proficiency, whereas /sp/ cluster /p/ remained comparatively short and became more clearly separated from singleton /p/ at higher proficiency. The pattern supports the interpretation that development was not simply a uniform increase in aspiration, but an increasingly context-sensitive control of English /p/ realization.

3.5 Perception, task-style checks, and individual differences

The perception probes aligned with the production results. Learners' mean /b/-/p/ perceptual boundary was later than that of General American English controls, indicating that learners required longer VOT before identifying tokens as /p/. Within learners, higher proficiency was associated with a lower, more English-like boundary (r = −0.63), and learners with greater /p/-/b/ production overlap tended to have later perceptual boundaries (r = 0.35). Learners also rated overaspirated /sp/ tokens as relatively natural, whereas controls rated such tokens lower, consistent with residual difficulty in context-sensitive aspiration control. Figure 6 shows the proficiency-boundary relationship, and Table 5 summarizes the perception and task-style checks.

Table 5. Perception and task-style checks by group.

Individual differences were substantial. Learner singleton /p/ mean VOT ranged from 27.54 to 53.83 ms, within-speaker SD ranged from 12.26 to 24.16 ms, and /p/-/b/ overlap ranged from 0.08 to 0.66. These ranges show that the learner group cannot be reduced to a single mean pattern. The Arabic baseline also confirmed local L1 structure: learners produced Arabic /b/ with robust prevoicing (M = −74.67 ms, SD = 15.62), while Arabic /t/ and /k/ were short-lag positive (M = 21.60 ms and 29.64 ms, respectively). This local baseline supports the interpretation that English singleton /p/ requires movement into a long-lag region not used contrastively for the learners' Jordanian Arabic voiceless stops.

4 Discussion

The study examined whether Arabic L1 learners' English stop voicing is better characterized by distributional organization than by mean VOT alone. The results support that view. Learners produced a reliable singleton /b/-/p/ contrast, but singleton /p/ remained shorter, more variable, and more overlapped with /b/ than in General American English controls. With increasing proficiency, singleton /p/ moved toward longer-lag VOT, dispersion decreased, /p/-/b/ overlap decreased, the perceptual /b/-/p/ boundary shifted downward, and aspiration suppression in /sp/ clusters improved. These outcomes show that L2 stop learning involves several partly separable dimensions: category placement, category precision, cross-category separation, perceptual boundary alignment, and context-sensitive phonological control.

The findings refine the SLM-r interpretation of Arabic-English VOT learning. English /p/ is not simply acquired by shifting the learner mean toward native-speaker values. Less proficient learners showed broader and more overlapped /p/ distributions, while higher-proficiency learners showed tighter and better-separated distributions. This pattern is consistent with the SLM-r claim that L2 phonetic categories are probabilistic distributions that reorganize with experience (Flege and Bohn, 2021). The results also fit the logic of the original SLM, where a new or difficult L2 phone can initially be assimilated to an L1-based region of phonetic space and then separate gradually with experience (Flege, 1987, 1995). At the same time, the interpretation should remain cautious: production overlap is not identical to perceptual equivalence classification. The perception probes strengthen the interpretation by showing related boundary effects, but direct perceptual assimilation designs would be needed to test PAM-style assimilation patterns more explicitly (Best et al., 2001; Best and Tyler, 2007).

The Jordanian Arabic baseline is important for interpreting the source of learner difficulty. Learners' Arabic /b/ showed robust prevoicing, while Arabic /t/ and /k/ showed short-lag positive VOT. This profile is compatible with prior work on Arabic VOT and Jordanian Arabic in particular, while also acknowledging dialectal variation across Arabic varieties (Mitleb, 2001; Al-Tamimi et al., 2021; Al-Tamimi and Khattab, 2018; Kulikov, 2022; Al-Deaibes et al., 2025, 2026). The L2 task therefore required learners to build a long-lag English /p/ category while controlling the relationship between that category and a short-lag /b/ category in English. The fact that learner /b/ was near the English short-lag region but learner /p/ was compressed relative to controls is consistent with the cross-language asymmetry described in earlier Arabic-English work (Flege and Port, 1981; Khattab, 2002).

The /sp/ findings show that phonological context is a separate component of learning. Learners did not merely increase /p/ VOT everywhere. Higher proficiency predicted longer singleton /p/ but stronger suppression in /sp/ clusters, a pattern that matches English contextual timing rather than a global aspiration strategy. This result is consistent with classic evidence that English stop VOT depends on context and that /s/-stop clusters show reduced aspiration (Klatt, 1975; Lisker and Abramson, 1967; Mielke and Nielsen, 2018). It is also compatible with recent work showing that /s/-stop clusters can reveal whether learners have acquired a shared phonological space rather than isolated segmental targets (Schwartz, 2023). The perception naturalness results add an important qualification: learners still accepted overaspirated /sp/ tokens more readily than controls, suggesting that context-conditioned phonological control lags behind basic singleton /p/ category movement.

Individual differences were not peripheral noise. Learners varied widely in mean VOT, within-speaker dispersion, overlap, perceptual boundary location, and /sp/ repair rates. This is expected in L2 phonetic learning, where talker-specific phonetic organization, cue weighting, cognitive factors, and input histories all affect production and perception (Davidson, 2016; Chodroff and Wilson, 2017; Chodroff et al., 2019; Darcy et al., 2015; Schertz et al., 2015). The present data show why those differences matter: two learners may have similar mean /p/ VOT but differ in overlap with /b/, or may show good singleton aspiration while still accepting overaspirated /sp/ clusters. A distributional analysis captures these distinctions more clearly than a mean-only analysis.

The proficiency measure has both strengths and limitations. Because the placement test did not include speaking or pronunciation, the proficiency-VOT relationships cannot be dismissed as circular. However, a reading-listening-grammar composite is a broad index of English ability, not a direct measure of phonetic experience, phonological awareness, or English input quality. The learner sample also shared a classroom-based, non-immersion educational context. The observed effects therefore show that distributional VOT learning can occur without residence in an English-speaking country, but they do not establish how immersion, richer conversational input, or targeted pronunciation instruction would alter the trajectory (Piske et al., 2001; Alharbi et al., 2023; Gorba and Cebrian, 2021; Saito and Plonsky, 2019).

The controlled reading task is another limitation. Carrier-phrase reading supports clean acoustic measurement and balanced lexical comparisons, but it may elicit careful speech and cannot by itself establish everyday pronunciation patterns. The speech-rate and semi-spontaneous checks partly address this issue by showing comparable group differences across task styles, but future work should include larger spontaneous samples and analyze rate, prosodic position, lexical frequency, and communicative setting more extensively. Such work would also help connect acoustic category structure to intelligibility and perceived accentedness, which are influenced by more than VOT alone (Munro and Derwing, 1995; Major, 2001).

Overall, the study supports a modest but important theoretical contribution. It provides a production-centered and perception-supported operationalization of SLM-r for Arabic-English stop voicing: proficiency is associated with longer-lag singleton /p/, reduced within-category dispersion, reduced between-category overlap, lower perceptual VOT boundary values, and better context-conditioned aspiration suppression. The central implication is not simply that higher-proficiency learners are more native-like. Rather, the depth of learning is visible in the internal structure of their phonetic distributions and in their ability to condition those distributions by phonological context.

5 Conclusion

Arabic L1 learners of English reorganized English stop voicing through gradual, distribution-level change. Higher proficiency predicted longer singleton /p/ VOT, tighter /p/ distributions, reduced /p/-/b/ overlap, more English-like perceptual boundaries, and stronger /sp/ aspiration suppression. Mean VOT therefore captured only part of the learning trajectory. The combination of production distributions, perception probes, Arabic baseline data, and context-sensitive analysis shows that L2 phonetic development is best understood as progressive control over probabilistic categories and their phonological conditioning.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Mark Antoniou, Western Sydney University, Australia

Reviewed by: Motasim Almobaddal, Imam Muhammad ibn Saud Islamic University, Saudi Arabia

Wafaa Alshangiti, King Abdulaziz University, Saudi Arabia