Cue weighting in the perception and production of English stop voicing by Arabic-speaking learners: phonological context and immediate post-instruction adjustment

Introduction

English stop voicing is not signaled by a single acoustic property. In word-initial position, listeners rely most strongly on voice onset time (VOT), but onset F0 and vowel duration also contribute to robust categorization, especially when the timing cue is weak or contextually constrained (Dmitrieva et al., 2015; Hanson, 2009; Lisker, 1986; Lisker and Abramson, 1964). This is particularly clear in /sp/ onsets. In /sp/ onsets, English suppresses aspiration even though the stop remains voiceless, which means that speakers must produce a voiceless stop without the long-lag VOT that usually marks singleton /p/ (Kim, 2018; Klatt, 1975).

That configuration is hard for Arabic-speaking learners. Arabic contrasts stop voicing, but it does not use aspiration as a phonemic contrast in the English sense, and reported VOT patterns are shorter and differently organized across Arabic varieties than in English, with additional evidence that laryngeal implementation varies across Arabic consonant classes and dialects (Al-Deaibes et al., 2026; Kulikov, 2022; Mitleb, 1984a; Yeni-Komshian et al., 1977). Earlier work showed cross-language interference from Arabic to English in both production and intelligibility, and more recent work has continued to find difficulty with English voiceless stops and incomplete control of aspiration with increasing experience (Al-Damen et al., 2025; Alharbi et al., 2023; Flege and Port, 1981; Mitleb, 1985).

The cue-weighting literature in L2 speech is well established. Speech categories are typically supported by multiple cues whose relative importance shifts with language experience, development, and cue reliability, and broader work on speech perception has shown that category structure itself shapes how listeners respond to within-category variation and cross-category contrast (Holt and Lotto, 2006; Iverson and Kuhl, 1995; Nittrouer and Miller, 1997; Schertz and Clare, 2020). For that reason, the present contribution is narrower and more specific. It focuses on Arabic-speaking learners, examines perception and production in the same participants, isolates the concrete /sp/ context rather than speaking loosely about all s-stop clusters, and tests whether a brief pronunciation lesson is followed by immediate changes in cue weighting and phonetic realization within the same session.

The study is framed primarily by the revised Speech Learning Model, or SLM-r, because that model explicitly treats second-language phonetic categories as gradient, experience-sensitive, and internally structured by multiple cues rather than as all-or-none categories (Flege, 1995, 2007; Flege and Bohn, 2021). That orientation also builds on earlier work that treated L2 speech learning as phonetic approximation and category formation shaped by experience across perception and production (Bohn and Flege, 1997; Flege, 1980, 2003; Flege and Eefting, 1987). Related research has also shown that segmental learning contributes to broader accentedness outcomes in adult L2 speech (Flege et al., 1995). Other frameworks, including the Perceptual Assimilation Model, are also relevant to non-native speech learning, especially where cross-language similarity and category mapping are central concerns. The present study foregrounds SLM-r because its cue-based account is particularly well suited to a design that tracks multiple cues across perception and production within the same participants. At the same time, the present data should not be read as a decisive test of any one framework. The goal is more modest: to ask whether the observed patterns are consistent with cue-based restructuring as proficiency changes and as learners receive short, explicit instruction.

The perception-production relation also needs to be handled carefully. Parallel change across the two domains could reflect common representations, but it does not automatically establish directionality. Recent work on laryngeal systems in multilingual speakers makes that point clearly and argues for analyzing perception and production together while resisting stronger causal claims than the data allow (Bárkányi and Kiss, 2025). That caution is built into the present study.

Four research questions guide the study.

• How do Arabic-speaking learners weight VOT, onset F0, and vowel duration in the perception of English stop voicing, and how do those weights differ from native English controls?
• How are VOT, onset F0, and vowel duration realized in learners' production of singleton /p/-/b/ contrasts, and how do those realizations vary by proficiency and session?
• To what extent do learners suppress aspiration in /sp/ onsets, and does that control vary by proficiency and session?
• Are immediate posttest changes in perceptual cue weighting related to immediate posttest changes in production?

Literature review

English stop voicing as a multi-cue contrast

English listeners weight VOT heavily in the /b/-/p/ contrast, but they do not ignore the rest of the signal. Onset F0 and vowel duration systematically covary with laryngeal category, and both have been shown to support voicing judgments and phonetic realization in English; related work also shows structured talker- and context-sensitive covariation in the realization of English stop voicing more broadly (Chodroff and Wilson, 2017; Davidson, 2016; Dmitrieva et al., 2015; Hanson, 2009; Lisker, 1986; Lisker and Abramson, 1964). The present study distinguishes cue trading from cue reweighting. Cue trading refers to the use of one available cue when another cue is weak or conflicting in a particular perceptual decision, whereas cue reweighting refers to a change in the longer-run relative importance assigned to available cues across experience or instruction (Holt and Lotto, 2006; Nittrouer and Miller, 1997; Schertz and Clare, 2020). That distinction matters because the current design includes both neutral continua and cue-conflict stimuli.

Arabic stop systems and the relevance of dialect background

Arabic-speaking learners do not approach English stop voicing from a blank slate. Arabic varieties differ in detail, but the learner problem is still coherent: English long-lag aspiration lacks a direct phonemic counterpart, and the English pattern requires both a wider singleton VOT split and context-sensitive suppression in /sp/ onsets (Al-Damen et al., 2025; Flege and Port, 1981; Mitleb, 1984a,b, 1985; Yeni-Komshian et al., 1977). Related work on Arabic segments produced by English speakers likewise suggests that Arabic-English speech learning is sensitive to fine phonetic detail and cannot be reduced to broad phonological labels alone (Aldamen and Al-Deaibes, 2023). The learner sample in the present study included Jordanian Levantine, Palestinian Levantine, Egyptian Cairene, Saudi Najdi, Saudi Hijazi, Kuwaiti Gulf, and Iraqi Mesopotamian backgrounds. Those backgrounds were documented because dialect variation is relevant to the interpretation of laryngeal learning. They were not modeled as a separate inferential factor because the per-dialect cell sizes were small and were distributed across proficiency groups. The study therefore treats dialect background as an important sampling characteristic rather than as a factor that can be estimated reliably in the current dataset.

This also helps clarify what is special about the Arabic-English case. Learners whose first language already uses aspirated vs. unaspirated timing more similarly to English may still face new cue-weighting demands, but Arabic-speaking learners must first establish the role of long-lag aspiration itself and then learn where English suppresses it. That combination makes /sp/ especially informative.

Instruction and the perception-production interface

Pronunciation instruction can change both perceptual sensitivity and production when it directs attention to relevant acoustic structure and pairs that attention with guided practice (Bradlow et al., 1997; Saito, 2013; Saito and Plonsky, 2019). What is less often shown is whether the same learners who adjust perceptually also adjust phonetically in the same session, and whether the adjustment is strongest in the phonological contexts that are hardest to control. That gap is where the present study now situates itself.

Across these strands, prior work has often emphasized single cues, focused on production without matched perceptual evidence, or treated contextual suppression less directly, which makes it difficult to see how cue priorities and phonetic realization shift together within the same learners (Al-Damen et al., 2025; Flege and Port, 1981; Schertz and Clare, 2020). The present design addresses those limitations by modeling multiple perceptual cues, separating singleton voicing from /sp/ suppression, and examining perception and production within the same participants.

Method

Design

The study used a single-session, pretest-posttest design that combined perception and production within the same participants. Learners completed perception pretest, production pretest, a live pronunciation mini-lesson, perception posttest, and production posttest. Native controls completed the same task sequence but received a matched silent interval instead of instruction. Because the design does not include a learner no-instruction control group, posttest differences are interpreted as immediate post-instruction adjustment rather than durable learning.

Participants

Forty adults participated. The learner sample consisted of 30 Arabic-speaking undergraduates, with 10 learners in each proficiency band: Novice-High, Intermediate-High, and Advanced. The control group consisted of 10 native speakers of American English. Table 1 summarizes the participant characteristics. Learners reported Arabic as their first and dominant language, no speech or hearing disorder, and limited residence in an English-speaking country. Across the learner sample, residence abroad ranged from 0 to 3 months. For manuscript reporting, all participants were anonymized as P01 to P40. No names, initials, or identifying descriptors are used.

Table 1. Participant characteristics by group.

Dialect backgrounds for the learner sample are described in Section 2.2 and Section 3.2. Native controls are not assigned Arabic dialect backgrounds.

Proficiency grouping was based on structured oral interviews rated independently by two trained raters over four dimensions: global comprehensibility, fluency, grammatical control, and interactional ability. Scores were assigned on a 1 to 5 scale in half-point steps and averaged across raters and dimensions. The resulting mean scores separated the learner groups cleanly: novice M = 2.83, range = 2.75 to 2.94; intermediate M = 3.57, range = 3.50 to 3.69; advanced M = 4.29, range = 4.06 to 4.44.

Materials

Production stimuli

Production prompts were embedded in the carrier sentence Say ___ once. The final stimulus set contained 10 target words: six singleton items and four /sp/ items. The singleton set contrasted /p/ and /b/ across high front, low front, and high back vowels. One phonotactically legal nonce form, poon, was included to balance the high back vowel context. Because nonce forms can be processed differently from real words, item identity was modeled in the analyses and the full item list is reported in Appendix A, Table A1. The cluster context was operationalized specifically as /sp/ through speak, span, spoon, and spoke. Each item was produced three times at each session.

Perception stimuli

Perception stimuli were built from two lexical bases, peak~beak and pack~back, and included 54 unique manipulated items. The set contained a seven-step neutral VOT continuum, F0-manipulated items, duration-manipulated items, and mixed multicue items. Exact cue values are given in Appendix A, Table A2. The neutral continuum used VOT values from 0 to 60 ms in 10 ms steps with constant F0 and duration. The manipulated sets sampled midrange VOT values where cue conflict would be informative, while onset F0 and vowel duration were varied to create aligned and conflicting combinations. Each unique stimulus was presented twice at pretest and twice at posttest.

Instructional protocol

The pronunciation lesson was delivered live by the same trained instructor using a standardized slide deck and scripted prompts. The lesson targeted singleton /p/ and /b/ timing, then /sp/ suppression. It included articulatory explanation, auditory modeling, guided perception practice, and short production practice with immediate oral corrective feedback. The same core content was delivered to all learner groups, with pacing adjusted only to allow extra clarification when needed. The nominal target duration was 20 min, and the archived session logs show mean instruction times of 20.4 min for the novice group, 20.8 min for the intermediate group, and 20.4 min for the advanced group. The protocol is reproduced in Appendix B.

Procedure

Sessions were completed individually and lasted about 71 min for learners and 69 min for native controls. The order was fixed at perception pretest, production pretest, instruction or matched interval, perception posttest, and production posttest. Randomization lists differed across participants and between the two sessions so that the task order changed while the target contrasts remained constant.

Acoustic measurement and reliability

Recordings were made in the same quiet room at 44.1 kHz and 16-bit resolution with the same head-mounted headset setup across participants. The archived records preserved the standardized setup note but not the hardware make and model, so those details are unavailable. Acoustic measurements were taken in Praat. VOT was measured from release to the onset of periodic voicing. Onset F0 was measured over the first 30 ms of the following vowel, and vowel duration was measured from vowel onset to vowel offset (Dmitrieva et al., 2015; Hanson, 2009; Klatt, 1975; Lisker and Abramson, 1964).

A total of 240 production tokens, or 10% of the dataset, were independently re-annotated by a second phonetician. Reliability was high for all three measures: VOT ICC(A, 2) = 0.992, onset F0 ICC(A, 2) = 0.976, and vowel duration ICC(A, 2) = 0.949 (Shrout and Fleiss, 1979). Production exclusions were rare. Of 2,400 production trials, 2,362 were retained, which corresponds to 1.58% excluded tokens. Perception missingness was even smaller, with valid responses on 8,607 of 8,640 trials, or 99.62%.

Statistical Analysis

Analyses were structured to keep laryngeal category and phonological context analytically distinct. Production data were analyzed with six linear mixed-effects models implemented in the statsmodels package in Python. Three models examined singleton /p/-/b/ contrasts with fixed effects of voicing category, group, timepoint, and their interactions. Three parallel models examined /sp/ suppression with fixed effects of context, group, timepoint, and their interactions. Participant was included as a random intercept in each model, and item was modeled as a variance component. Richer random-slope structures were attempted but were not estimable without singularity, so the final specifications retained the richest converged intercept structure, in line with the general logic of Barr et al. (2013). Full model outputs are provided in Appendix C, Tables C4–C9.

Perception cue weighting was estimated in two steps. First, separate participant-by-time logistic regressions over valid perception trials yielded coefficient estimates for VOT, onset F0, and vowel duration. Those coefficients served as each participant's cue weights at pretest and posttest. Second, each cue weight was analyzed with a linear mixed-effects model including group, timepoint, and their interaction, with a random intercept for participant. This approach made the independent cue contributions directly inspectable and aligns with the manuscript's focus on cue weighting rather than on endpoint accuracy alone. Full outputs appear in Appendix C, Tables C1–C3.

The perception-production relation was tested with two planned participant-level correlations: change in perceptual VOT weight with change in /sp/ suppression, and change in perceptual VOT weight with change in singleton /p/ VOT. Because only two linkage tests were planned, Holm adjustment was applied to that pair of p values. No large family of exploratory pairwise contrasts was conducted; interpretation instead centers on pre-specified model terms, effect sizes, and confidence intervals.

Results

Data integrity and descriptive orientation

The descriptive patterns appear in Table 2. Native controls showed the expected large singleton VOT split and stable /sp/ suppression at both sessions. Learners patterned by proficiency. The novice group entered with the shallowest perceptual VOT weighting, the smallest singleton VOT split, and almost no separation between singleton /p/ and /sp/ VOT. Intermediate and advanced learners were progressively closer to the native pattern, but the /sp/ context remained less stable than the singleton contrast.

Table 2. Group-level descriptive summary of perception weights and VOT means.

Perception weights are participant-level mean logistic coefficients by group and timepoint. Production values are group means in milliseconds from the analysis-ready dataset. Full inferential results are reported in Appendix C.

Figures 1–3 visualize the descriptive VOT patterns summarized in Table 2. Figure 1 shows VOT by group, context, and session, Figure 2 plots the contextual difference between the singleton average and the /sp/ cluster context, and Figure 3 plots singleton /p/-/b/ separation across sessions.

In Figures 1 extbfand 2, the cluster condition corresponds to the /sp/ context analyzed in the present study.

Perception cue weighting

The revised perception analysis confirms that the cue-weighting story is not only about VOT. At pretest, native controls had the highest VOT weight (M = 2.57), followed by the advanced group (M = 1.79), the intermediate group (M = 1.56), and the novice group (M = 0.92), as shown in Table 2. The VOT-weight model in Appendix C, Table C1, showed that all learner groups underweighted VOT at pretest relative to native controls: advanced b = −0.77, 95% CI [−1.12, −0.43], p < 0.001; intermediate b = −1.00, 95% CI [−1.35, −0.65], p < 0.001; novice b = −1.64, 95% CI [−1.99, −1.29], p < 0.001.

Native controls showed no reliable session change in VOT weighting, b = −0.16, 95% CI [−0.49, 0.18], p = 0.361. The learner groups, however, showed positive posttest shifts relative to the native pattern: advanced b = 0.66, 95% CI [0.19, 1.13], p = 0.006; intermediate b = 0.54, 95% CI [0.07, 1.01], p = 0.025; novice b = 0.97, 95% CI [0.50, 1.44], p < 0.001. Descriptively, the novice group rose from 0.92 to 1.74, the intermediate group from 1.56 to 1.95, and the advanced group from 1.79 to 2.30.

Secondary cues patterned differently. In the F0-weight model, the only reliable shift was a modest decline from pretest to posttest across groups, b = −0.28, 95% CI [−0.47, −0.09], p = 0.004, with no reliable group-specific interactions. In the duration-weight model, intermediate and novice learners entered with stronger duration weighting than native controls, intermediate b = 0.58, 95% CI [0.33, 0.82], p < 0.001, novice b = 0.68, 95% CI [0.44, 0.92], p < 0.001, and both groups then reduced duration weighting at posttest, intermediate b = −0.44, 95% CI [−0.75, −0.13], p = 0.005, novice b = −0.52, 95% CI [−0.83, −0.21], p = 0.001. The pattern suggests a shift toward stronger use of the primary timing cue in the lower-proficiency groups, not the disappearance of secondary cues.

The descriptive identification curves that correspond to this perceptual pattern appear in Figure 4.

Production results for singleton /p/-/b/

The singleton VOT model in Appendix C, Table C4, showed the expected voicing effect in the native baseline. Voiceless singleton stops had much longer VOT than voiced singletons, b = 65.15, 95% CI [63.71, 66.60], p < 0.001. Learners produced smaller singleton VOT separations at pretest than native controls, and the reduction scaled with proficiency: advanced b = −11.32, 95% CI [−13.36, −9.27], p < 0.001; intermediate b = −30.91, 95% CI [−32.96, −28.87], p < 0.001; novice b = −48.65, 95% CI [−50.69, −46.60], p < 0.001.

The posttest interactions show that immediate adjustment was strongest in the groups that started farthest from the native pattern. Relative to native controls, the voiceless singleton VOT split increased at posttest by 2.97 ms in the advanced group, 10.05 ms in the intermediate group, and 21.55 ms in the novice group. Those estimates were reliable for all three learner groups, although the effect was much larger in the two lower-proficiency groups. Descriptive means in Table 2 show the same movement: novice singleton /p/ rose from 34.8 to 56.8 ms, intermediate from 49.7 to 59.8 ms, and advanced from 66.5 to 70.6 ms.

The secondary cues aligned with that pattern but did not mirror it mechanically. In the singleton F0 model, onset F0 was higher after /p/ than after /b/ in the native baseline, b = 23.86, 95% CI [22.51, 25.21], p < 0.001. The novice group showed an even larger F0 difference at pretest, b = 3.35, 95% CI [1.45, 5.25], p = 0.001, but there was no reliable learner-specific posttest change. In the singleton duration model, vowels were longer after /b/ than after /p/ in the native baseline by about 50 ms. Intermediate and novice learners showed smaller duration splits at pretest, and the novice group increased that split at posttest, b = −9.92, 95% CI [−17.56, −2.27], p = 0.011. This suggests that duration still functioned as part of the learner system, especially at lower proficiency, but it was not the main locus of immediate change.

Production results for /sp/ aspiration suppression

The /sp/ model sharpened the contextual story. In native controls, /sp/ tokens had much shorter VOT than singleton /p/ tokens, b = −63.14, 95% CI [−64.62, −61.67], p < 0.001, which is the expected suppression pattern. Learners showed attenuated suppression at pretest: advanced b = 17.46, 95% CI [15.37, 19.54], p < 0.001; intermediate b = 42.43, 95% CI [40.35, 44.52], p < 0.001; novice b = 60.38, 95% CI [58.30, 62.47], p < 0.001. In other words, the lower the proficiency level, the more /sp/ VOT drifted toward singleton /p/ rather than toward the native cluster target.

Posttest change again depended on group. The three-way interactions were negative for all learner groups, which means that /sp/ suppression increased relative to the native pretest baseline: advanced b = −6.86, 95% CI [−9.80, −3.92], p < 0.001; intermediate b = −13.49, 95% CI [−16.43, −10.55], p < 0.001; novice b = −29.33, 95% CI [−32.27, −26.39], p < 0.001. Table 2 shows the descriptive counterpart. Mean /sp/ VOT fell from 32.0 to 24.2 ms in the novice group, from 29.0 to 25.1 ms in the intermediate group, and from 20.8 to 17.6 ms in the advanced group, while native controls remained stable at about 15 ms.

The /sp/ F0 model also showed context sensitivity. Native controls had lower onset F0 in /sp/ than in singleton /p/, b = −19.08, 95% CI [−20.35, −17.81], p < 0.001. Intermediate and novice learners showed less contextual lowering at pretest, and both groups showed more native-like posttest adjustment, intermediate b = −5.93, 95% CI [−8.47, −3.39], p < 0.001, novice b = −9.86, 95% CI [−12.40, −7.32], p < 0.001. By contrast, vowel duration did not show a targeted /sp/ posttest effect. The contextual duration difference in the native baseline was small, and none of the three-way interactions for duration were reliable. Here the main immediate adjustment was carried by VOT, with supporting movement in onset F0.

Figures 5, 6 show the descriptive context-sensitive patterns for onset F0 and vowel duration in the singleton and /sp/ contexts.

Perception-production linkage

The participant-level linkage analyses support a connection between perceptual and productive adjustment. Change in perceptual VOT weight correlated positively with change in the /sp/ suppression index, r = 0.440, 95% CI [0.149, 0.661], Holm-adjusted p = 0.009. It also correlated positively with change in singleton /p/ VOT, r = 0.427, 95% CI [0.133, 0.652], Holm-adjusted p = 0.009. These estimates are reported in Appendix C, Table C10. Learners who shifted more strongly toward VOT in perception also tended to show larger immediate production changes in the predicted direction.

Discussion

Cue weighting in perception

The revised analyses clarify the perceptual story. Arabic-speaking learners did not simply perform worse than native controls. They weighted the same cue set differently. VOT was already part of the learner system, but it carried less weight at pretest, especially for the novice group, while duration carried more weight in the intermediate and novice groups. That profile is consistent with the view that early L2 categories can be internally structured around a nonnative cue hierarchy rather than around a total absence of cue use (Flege, 1995, 2007; Flege and Bohn, 2021; Holt and Lotto, 2006; Schertz and Clare, 2020).

The cue-conflict design matters here. Because onset F0 and vowel duration were manipulated independently, the data support a more precise claim than a single-continuum design would allow. The immediate posttest change was not just a steeper identification slope along a single continuum. It was a redistribution of weighting across cues, strongest for VOT and duration, and concentrated in the learner groups that had entered with the least native-like weighting profile.

Production across singleton and /sp/ contexts

The production results show why the context question needed to be separated from the singleton voicing question. Learners improved their singleton /p/-/b/ contrast with proficiency, but /sp/ remained a distinct challenge. That challenge is not trivial. It requires speakers to maintain a voiceless category while suppressing the long-lag timing pattern that English encourages elsewhere. The present data make clear that a learner can produce a fairly good singleton contrast and still struggle in /sp/.

This is also where the Arabic-English comparison is most informative. Prior work has shown that Arabic speakers often produce reduced English VOT contrasts and that the phonetic organization of stop voicing and temporal structure differs across Arabic and English (Al-Damen et al., 2025; Flege and Port, 1981; Mitleb, 1984a,b, 1985; Yeni-Komshian et al., 1977). The current results extend that line by showing that the difficulty is not uniform across contexts. The /sp/ pattern remains the hardest place to control, even when the singleton contrast moves toward the target.

Immediate post-instruction adjustment

The lesson appears to have mattered, but only in the limited sense that this design can support. The strongest immediate posttest changes occurred in the same domains that were weakest at pretest: perceptual VOT weighting, singleton /p/ VOT, and /sp/ suppression in the novice and intermediate groups. That pattern fits the broader pronunciation-instruction literature, which has repeatedly shown benefits when learners are directed to relevant phonetic structure and then asked to practice it with focused feedback (Bradlow et al., 1997; Saito, 2013; Saito and Plonsky, 2019).

At the same time, those changes should be interpreted conservatively. The learner groups were not compared with a no-instruction learner control group, and all posttests occurred in the same session. Some session effects may therefore reflect rapid adaptation, practice, or heightened attention. The small session-wide drop in F0 weighting supports that caution. What can be claimed confidently is that the immediate posttest pattern was structured rather than random and that it was strongest in the contexts and groups where the system had most room to shift.

Perception-production relations

The participant-level correlations strengthen the perception-production argument without making it overly strong. The data now show more than parallel group means. Participants who increased their perceptual reliance on VOT also tended to increase singleton /p/ VOT and strengthen /sp/ suppression. That is consistent with the idea that perception and production draw on linked phonetic representations, which is central to SLM-r and is also in line with recent work urging tighter integration of the two domains in laryngeal research (Bárkányi and Kiss, 2025; Flege, 1995, 2007; Flege and Bohn, 2021).

Still, the data do not establish directionality. The present design cannot say whether perceptual change drove production change, whether production practice sharpened perception, or whether both responded to the same short-term instructional focus. The safer conclusion is that the two domains changed together in a way that is consistent with shared restructuring.

Pedagogical implications

The present findings point to a clear instructional priority for Arabic-speaking learners of English: stop voicing should be taught as a cue system rather than as a single pronunciation rule. In practice, that means instruction should not stop at telling learners that English /p/ is aspirated and /b/ is voiced. Learners need guided attention to the fact that English listeners rely most heavily on VOT, while onset F0 and vowel duration also contribute to the signal in structured ways (Dmitrieva et al., 2015; Hanson, 2009; Lisker, 1986; Lisker and Abramson, 1964). The current results suggest that lower-proficiency learners may enter instruction with a less target-like cue hierarchy, including heavier reliance on duration, so teaching should explicitly redirect attention toward timing at release and voicing onset. A useful pedagogical sequence would begin with perceptual contrasts between singleton /p/ and /b/, move to controlled noticing of release-to-voicing delay, and then connect that perceptual work to short, monitored production practice. This approach is compatible with research showing that pronunciation instruction is more effective when it draws learners' attention to relevant phonetic structure and pairs that attention with guided practice and feedback (Bradlow et al., 1997; Saito, 2013; Saito and Plonsky, 2019).

The results also show that pronunciation teaching should include phonological context much earlier and more systematically than many classroom routines currently do. Learners may improve their singleton /p/-/b/ contrast while still struggling with /sp/ onsets, because /sp/ requires suppression of a pattern that English otherwise encourages in singleton voiceless stops. That makes /sp/ words pedagogically valuable, not peripheral. Classroom materials should therefore contrast forms such as peak and speak, or pack and span, so learners can experience that English voicelessness is not realized identically across contexts. Instructors can make this concrete by combining auditory models, slowed demonstrations, visual timing cues, and immediate feedback on whether extra aspiration is being inserted after /s/. The goal is not to overload learners with technical terminology, but to help them develop context-sensitive control. This is especially important for Arabic-speaking learners, whose first-language background does not provide a direct counterpart to the English long-lag aspiration pattern or to its suppression in /sp/ onsets (Al-Damen et al., 2025; Flege and Port, 1981; Mitleb, 1984a,b, 1985; Yeni-Komshian et al., 1977).

A further implication concerns how perception and production should be organized in pedagogy. The present study supports treating them as mutually informative components of the same instructional sequence rather than as separate skills taught on different days or in different units. Learners who shifted more strongly toward VOT in perception also tended to show larger productive adjustment in singleton /p/ and /sp/ suppression, which suggests that classroom gains may be stronger when listening tasks, metalinguistic explanation, and production practice are deliberately linked. This does not require laboratory-style training. Even brief classroom routines can incorporate this logic: learners can first identify which member of a contrast they hear, then explain what they noticed, then produce the same forms with focused corrective feedback. For teachers, the practical lesson is that improvement is likely to be strongest when instruction targets the parts of the system that are least stable, especially lower-proficiency learners' cue priorities and context-sensitive timing control. In that sense, effective pronunciation teaching is not just about getting students to say the sound correctly. It is about helping them reorganize which parts of the acoustic signal they attend to, how they interpret them, and how they reproduce them under different phonological conditions (Bárkányi and Kiss, 2025; Flege, 1995, 2007; Flege and Bohn, 2021; Holt and Lotto, 2006; Schertz and Clare, 2020).

Limitations and future directions

Several limitations should shape how the present findings are interpreted. First, the design was single-session and did not include a learner no-instruction control group, so the instructional component can only be interpreted as immediate posttest adjustment within the same session, not as evidence of stable phonetic learning. Some portion of the change may reflect short-term adaptation, task familiarity, or heightened attention after explicit focus on the target contrasts. Second, the production cluster context was limited to /sp/. For that reason, the manuscript does not extend its claims to all s-stop clusters, because other clusters may place somewhat different demands on timing and cue coordination. Third, the learner sample included several Arabic dialect backgrounds, but the number of participants within each dialect was too small to support reliable dialect-specific modeling. The study therefore treats dialect background as an important sampling characteristic rather than an inferential factor. Fourth, one singleton item was a nonce form used to balance vowel context. Item identity was modeled statistically, but lexical generalization remains narrower than it would be in a larger all-real-word design. Fifth, the archived methodological records preserved the standardized recording setup but not the hardware make and model. Reporting that gap directly is preferable to implying a degree of methodological completeness that the archive cannot support.

These constraints also point to several productive directions for future research. The most immediate need is a longitudinal design that includes delayed posttests and a learner comparison group without instruction, so that short-term adjustment can be separated from more durable restructuring. Such work would make it possible to test whether the cue-weighting shifts observed here are transient or whether they remain stable after the instructional focus has faded. A second priority is broader contextual sampling in production. The present study shows that /sp/ is highly informative, but future work should examine whether similar patterns hold across other s-stop sequences and across a wider range of lexical and prosodic environments. A third priority is larger-scale sampling of Arabic dialect backgrounds. Arabic stop systems are similar enough to motivate the present learner grouping, but dialect-specific differences in VOT organization and temporal structure remain theoretically important and could become analytically visible in a larger dataset (Mitleb, 1984a,b, 1985; Yeni-Komshian et al., 1977). That would help clarify which parts of the learning problem are broadly shared across Arabic-speaking learners and which may vary with dialectal phonetic experience.

Future research should also deepen the perception-production link in ways that go beyond the present participant-level correlations. The current data support the view that perceptual and productive change are related, but they do not establish directionality. A stronger design would track learners across multiple sessions, model individual trajectories, and test whether change in perceptual cue weighting precedes, follows, or co-develops with productive control. That would speak more directly to accounts in which L2 phonetic categories are internally structured, gradient, and jointly reflected in perception and production (Bárkányi and Kiss, 2025; Flege, 1995, 2007; Flege and Bohn, 2021; Schertz and Clare, 2020). It would also be useful to expand the instructional dimension itself by comparing different kinds of pronunciation training, such as perceptual-only training, production-only practice, and integrated cue-focused instruction. A program of that kind would help determine not only whether learners change, but which instructional components are most effective for reshaping cue priorities and for improving context-sensitive control of English stop voicing.

Conclusion

This study provides a more precise account of English stop voicing in Arabic-speaking learners by examining perception and production together and by treating stop voicing as a multi-cue phenomenon rather than a single-parameter contrast. The findings show that learners differed from native English controls not only in the size of their stop-voicing contrasts, but also in how they distributed perceptual weight across acoustic cues. VOT remained the primary cue throughout, yet lower-proficiency learners assigned it less weight at pretest and relied more heavily on secondary structure, especially vowel duration. In production, learners moved toward more target-like realization of the singleton /p/-/b/ contrast as proficiency increased, but this development did not eliminate contextual difficulty. The /sp/ environment remained the clearest site of instability, which shows that English stop voicing cannot be fully understood through singleton contrasts alone.

The immediate posttest results add an important qualification to that account. The largest shifts appeared in the groups and phonological environments that had shown the weakest pretest control, especially in perceptual VOT weighting, singleton voiceless timing, and /sp/ aspiration suppression. This pattern is consistent with rapid cue reorganization following explicit pronunciation instruction, but the present design supports only a short-term interpretation. The findings do not demonstrate durable restructuring, nor do they justify strong causal claims about instruction independent of task repetition or short-term adaptation. What they do show is that learner performance was responsive in systematic ways and that the strongest adjustment occurred where the gap between learner and native patterns was greatest.

A further contribution of the study lies in the clearer treatment of the perception-production relation. The participant-level linkage analyses showed that learners who increased their reliance on VOT in perception also tended to show larger productive adjustment in the predicted direction. This strengthens the empirical basis for viewing perception and production as linked components of phonetic development, while still leaving directionality unresolved. The more defensible interpretation is that both domains changed together in a manner consistent with shared, gradient restructuring. In that sense, the study offers a narrower, more transparent, and more strongly supported account of English stop voicing in Arabic-speaking learners.

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.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/flang.2026.1799496/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Yike Yang, Hong Kong Shue Yan University, Hong Kong SAR, China

Reviewed by: Bashar M. Farran, University of Missouri, United States

Zsuzsanna Bárkányi, The Open University, United Kingdom