Adaptive Artificial Intelligence for Personalized Pronunciation and Grammar Instruction in Multilingual African Classrooms: A Quasi- Experimental Study of Automatic Speech Recognition and Natural Language Processing

), regular classroom attendance, native or near-native proficiency in at least one African language, and minimal prior exposure to AI-assisted language-learning systems. These criteria ensured a relatively homogeneous sample while maintaining the linguistic diversity necessary for investigating adaptive AI in multilingual contexts. 5.3 Intervention The intervention extended over fourteen weeks and was organized into three phases comprising orientation and baseline assessment (Weeks 1–2), AI-supported instruction (Weeks 3–12), and post-intervention assessment (Weeks 13–14). The pronunciation intervention employed speaker-independent Automatic Speech Recognition technology capable of providing immediate, individualized feedback on learners' spoken production. During each instructional session, learners completed structured pronunciation exercises involving English and French phonological features before receiving real-time feedback identifying pronunciation errors and suggesting corrective practice. Speaker-independent ASR has consistently been shown to improve pronunciation accuracy because evaluation is based on standard target-language pronunciation rather than adaptation to individual accents (Neri et al., 2002). Continuous technology-mediated feedback also promotes learners' awareness of pronunciation errors and encourages repeated practice leading to sustained improvements in speaking performance (McCrocklin, 2021). Recent empirical evidence further demonstrates that ASR-supported pronunciation instruction significantly improves pronunciation accuracy, comprehensibility, and overall speaking proficiency when combined with iterative learner practice (Sun, 2023). The grammar intervention utilized an AI-powered Natural Language Processing engine capable of automatically identifying grammatical, syntactic, and lexical errors while generating contextualized corrective feedback. Rather than simply presenting corrected answers, the system explained grammatical rules and encouraged learners to revise their own writing through multiple feedback cycles. This instructional approach reflects contemporary intelligent language tutoring systems that integrate NLP techniques to provide flexible, learner-specific grammatical support (Amaral et al., 2011). Automated scoring and grammar analysis were informed by principles similar to those implemented in the Educational Testing Service's e-rater® engine, which provides reliable automated evaluation of linguistic accuracy and writing quality (Educational Testing Service, 2024). The combined AI intervention integrated both pronunciation and grammar components within a single adaptive learning environment. The system continuously analyzed learner IJEE IJEE www.ijee.io performance across activities and dynamically adjusted task difficulty, practice frequency, and instructional content according to individual learning progress. Such adaptive personalization represents a defining characteristic of modern AI-supported education and has been identified as a major contributor to improved learner engagement and academic achievement because instruction becomes increasingly responsive to individual learner needs (Bernacki et al., 2021). Across all experimental groups, peer-correction activities complemented automated feedback, encouraging collaborative reflection and reinforcing language learning through guided interaction. 5.4 Data Collection Instruments Data were collected using standardized instruments designed to assess multiple dimensions of second-language proficiency. Pronunciation performance was evaluated through read-aloud passages and structured speaking tasks scored using established pronunciation criteria addressing fluency, comprehensibility, pronunciation accuracy, and overall speaking quality. The assessment approach aligns with contemporary pronunciation research that combines expert ratings with technology-supported speech analysis to provide comprehensive evaluation of oral language development (McCrocklin, 2021; Sun, 2023). Grammar proficiency was measured using standardized language tests together with writing tasks analyzed through automated grammar assessment. Essays were evaluated using AI- supported scoring procedures comparable to those employed by the Educational Testing Service's e-rater® system, allowing objective measurement of grammatical accuracy, syntactic complexity, and overall writing quality (Educational Testing Service, 2024). Listening comprehension assessments examined learners' ability to recognize spoken words and comprehend complete sentences in authentic language contexts. Together, these instruments provided complementary measures of productive and receptive language skills, enabling comprehensive evaluation of the effectiveness of adaptive 5.5 Data Analysis Data analysis was conducted using SPSS Version 28. Descriptive statistics were first computed to establish baseline equivalence across the four instructional groups. The primary analyses employed Analysis of Covariance , with pre-test scores entered as covariates to control for initial differences and to estimate the effects of instructional condition on post-test performance. Additional analyses included hierarchical multiple regression to examine whether learner characteristics such as first-language background, digital literacy, and device access moderated treatment effects. Pearson product-moment correlation analyses were performed to investigate relationships among gains in pronunciation, grammar, listening comprehension, and overall speaking proficiency. Finally, effect sizes were reported using partial eta squared (η2) for ANCOVA and Cohen's d for pairwise comparisons to complement significance testing and facilitate interpretation of the practical magnitude of the observed intervention effects. These analytical procedures provided a comprehensive statistical evaluation of the study hypotheses while ensuring robust interpretation of the effectiveness of adaptive AI- assisted language instruction. Results Quantitative results Table 1 presents demographic characteristics and baseline equivalence across the four experimental conditions. Random assignment of intact classes resulted in comparable distributions of age, gender, L1 background, and prior exposure to English and French IJEE IJEE www.ijee.io language learning. One-way ANOVAs revealed no significant between-group differences on any pre-test measure (all ps> .05), confirming baseline equivalence essential for valid quasi- experimental inference. Table 1 Demographic Characteristics and Baseline Equivalence by Group Variable EG1: ASR Pronunciation (n = 30) EG2: NLP Grammar (n = 30) EG3: Combined AI (n = 30) CG: Traditional (n = 30) F/χ2 P Demographic Variables Age, M (SD) 17.4 (1.2) 17.6 (1.1) 17.3 (1.3) 17.5 (1.2) 0.42 .739 Gender, % female 56.7 53.3 60.0 56.7 0.32 .956 L1: Tone language, % 46.7 50.0 43.3 46.7 0.29 .963 Prior English/French tutoring, % 36.7 40.0 33.3 36.7 0.34 .952 Digital literacy, M (SD) 3.4 (0.8) 3.5 (0.7) 3.6 (0.9) 3.4 (0.8) 0.51 .677 Pre-test Measures Word Recognition Accuracy 61.8 (8.4) 62.5 (7.9) 62.1 (8.1) 61.9 (8.3) 0.05 .986 Accentedness 4.12 (0.58) 4.08 (0.61) 4.15 (0.55) 4.10 (0.59) 0.09 .966 Comprehensibility 3.42 (0.51) 3.38 (0.48) 3.45 (0.53) 3.40 (0.50) 0.12 .948 Grammar Proficiency 58.3 (9.2) 59.1 (8.8) 58.7 (9.0) 58.5 (9.1) 0.04 .989 Sentence Comprehension 4.08 (1.15) 4.12 (1.08) 4.05 (1.12) 4.10 (1.10) 0.03 .993 Note. EG = Experimental Group; CG = Control Group. Digital literacy measured on 5-point scale. Accentedness and comprehensibility rated on 9-point scales. Sentence comprehension measured on 10-point scale. A one-way ANCOVA was conducted to examine the effect of instructional condition on post-test word recognition accuracy, controlling for pre-test performance. The analysis revealed a significant main effect of group, F(3, 115) = 8.74, p< .001, partial η2 = .186, indicating that instructional condition accounted for approximately 18.6% of variance in word recognition accuracy after controlling for baseline ability. Pairwise comparisons with Bonferroni correction indicated that all three experimental groups significantly outperformed the control group (all ps< .05), with the combined AI group demonstrating the largest adjusted mean difference (see Table 2). IJEE IJEE www.ijee.io Table 2 ANCOVA Results for Pronunciation and Grammar Outcomes Outcome Source SS df MS F p Partial η2 Observed Power Word Recognition Accuracy Pre-test (covariate) 2,845.62 1 2,845.62 48.36 < .001 .296 1.000 Group 1,542.38 3 514.13 8.74 < .001 .186 .975 Error 6,765.45 115 58.83 Total 11,153.45 119 Accentedness Pre-test (covariate) 18.42 1 18.42 62.15 < .001 .351 1.000 Group 9.86 3 3.29 11.08 < .001 .224 .993 Error 34.08 115 0.30 Total 62.36 119 Comprehensibility Pre-test (covariate) 15.23 1 15.23 58.74 < .001 .338 1.000 Group 8.95 3 2.98 11.51 < .001 .231 .995 Error 29.82 115 0.26 Total 54.00 119 Grammar Proficiency Pre-test (covariate) 3,124.58 1 3,124.58 44.82 < .001 .280 1.000 Group 1,289.76 3 429.92 6.17 .001 .139 .930 Error 8,015.42 115 69.70 Total 12,429.76 119 Sentence Comprehension Pre-test (covariate) 58.42 1 58.42 52.36 < .001 .313 1.000 Group 28.65 3 9.55 8.56 < .001 .183 .973 Error 128.35 115 1.12 Total 215.42 119 Note. SS = Sum of Squares; MS = Mean Square. All analyses controlled for corresponding pre-test scores as covariates. Table 3 presents adjusted post-test means, standard errors, and pairwise comparisons for all outcome measures. The combined AI instruction group (EG3) consistently demonstrated the largest treatment effects across pronunciation measures, with adjusted mean word recognition accuracy of 78.4% representing a 16.3% improvement from pre-test levels. Effect sizes for pairwise comparisons between experimental groups and the control group ranged from medium (d = 0.62) to large (d = 1.24), with the combined intervention showing the most substantial effects. IJEE IJEE www.ijee.io Table 3 Adjusted Post-Test Means and Pairwise Comparisons Outcome Group Adjusted M SE 95% CI Cohen’s d vs. CG P Word Recognition Accuracy (%) EG1: ASR Pronunciation 75.2 1.40 [72.4, 78.0] 0.89 .002 EG2: NLP Grammar 65.8 1.40 [63.0, 68.6] 0.24 .412 EG3: Combined AI 78.4 1.40 [75.6, 81.2] 1.24 < .001 CG: Traditional 62.5 1.40 [59.7, 65.3] — — Accentedness (1–9) EG1: ASR Pronunciation 5.12 0.10 [4.92, 5.32] 0.94 .001 EG2: NLP Grammar 4.35 0.10 [4.15, 4.55] 0.31 .285 EG3: Combined AI 5.48 0.10 [5.28, 5.68] 1.18 < .001 CG: Traditional 4.18 0.10 [3.98, 4.38] — — Comprehensibility (1–9) EG1: ASR Pronunciation 4.68 0.09 [4.50, 4.86] 0.98 < .001 EG2: NLP Grammar 3.95 0.09 [3.77, 4.13] 0.28 .352 EG3: Combined AI 4.92 0.09 [4.74, 5.10] 1.15 < .001 CG: Traditional 3.72 0.09 [3.54, 3.90] — — Grammar Proficiency EG1: ASR Pronunciation 63.5 1.53 [60.5, 66.5] 0.38 .198 EG2: NLP Grammar 72.8 1.53 [69.8, 75.8] 0.98 < .001 EG3: Combined AI 75.4 1.53 [72.4, 78.4] 1.12 < .001 CG: Traditional 60.2 1.53 [57.2, 63.2] — — Sentence Comprehension (1–10) EG1: ASR Pronunciation 6.45 0.19 [6.07, 6.83] 0.86 .003 EG2: NLP Grammar 5.28 0.19 [4.90, 5.66] 0.28 .361 EG3: Combined AI 6.82 0.19 [6.44, 7.20] 1.02 < .001 CG: Traditional 4.95 0.19 [4.57, 5.33] — — Note. Means adjusted for pre-test covariates. CI = Confidence Interval. Cohen's d calculated IJEE IJEE www.ijee.io using pooled standard deviation. p values Bonferroni-corrected for multiple comparisons. The ANCOVA for grammar proficiency revealed a significant main effect of instructional condition, F(3, 115) = 6.17, p = .001, partial η2 = .139. As hypothesized, the NLP grammar instruction group (EG2) and combined AI group (EG3) demonstrated significantly greater improvement than the control group, whereas the pronunciation-focused group (EG1) did not differ significantly from controls on this measure. This pattern of results supports the specificity of treatment effects, with grammar instruction producing grammar gains and pronunciation instruction producing pronunciation gains, while the combined intervention produced benefits across both domains. Essay quality analysis revealed a 22.3% reduction in grammatical errors per 100 words for the combined AI group compared to 8.5% reduction for the control group. The magnitude of error reduction was significantly correlated with time-on-task in the AI system, r(58) = .48, p< .001, suggesting that engagement with adaptive exercises mediated treatment effectiveness. To test for synergistic effects of combined pronunciation and grammar training, a series of 2 (Pronunciation Training: Present vs. Absent) × 2 (Grammar Training: Present vs. Absent) ANCOVAs was conducted on global speaking proficiency composite scores. Results revealed significant main effects of pronunciation training, F(1, 115) = 15.24, p< .001, partial η2 = .117, and grammar training, F(1, 115) = 8.63, p = .004, partial η2 = .070, qualified by a significant interaction, F(1, 115) = 4.28, p = .041, partial η2 = .036. The interaction pattern indicated that the combined training group achieved speaking proficiency scores exceeding the additive prediction based on single-component effects, consistent with synergistic skill integration. Simple effects analysis confirmed that the presence of pronunciation training enhanced the effectiveness of grammar instruction (d = 0.42) and vice versa (d = 0.38), suggesting that metalinguistic awareness developed through simultaneous attention to phonological and syntactic form facilitated global language proficiency development. Table 4 Factorial Analysis of Combined Training Effects on Global Speaking Proficiency Source SS df MS F p Partial η2 Pre-test (covariate) 42.58 1 42.58 56.24 < .001 .328 Pronunciation Training (P) 11.52 1 11.52 15.24 < .001 .117 Grammar Training (G) 6.52 1 6.52 8.63 .004 .070 P × G Interaction 3.24 1 3.24 4.28 .041 .036 Error 86.94 115 0.76 Total 150.80 119 The ANCOVA for sentence comprehension yielded a significant group effect, F(3, 115) = 8.56, p< .001, partial η2 = .183. Notably, both pronunciation training groups (EG1 and EG3) demonstrated significant improvements in listening comprehension despite receiving no explicit listening instruction, supporting the theoretical prediction of transfer from productive to receptive skills. The grammar-only group (EG2) did not differ from controls on this measure, p = .361. IJEE IJEE www.ijee.io Pearson correlation analysis revealed a strong positive relationship between pronunciation gain scores (post-test minus pre-test accentedness ratings) and listening comprehension improvements, r(118) = .72, p< .001, 95% CI [.61, .81]. This correlation remained significant when controlling for overall English proficiency and time-on-task, partial r = .58, p< .001, suggesting that the relationship between pronunciation practice and listening development is not merely artifactual of general language ability or engagement. Table 5 Bivariate Correlations Among Gain Scores Variable 1 2 3 4 1. Pronunciation Gain 2. Grammar Gain .34 3. Listening Gain .72 .28 4. Global Speaking Gain .68 .52 .61 Note. N = 120. Gain scores calculated as post-test minus pre-test. p < .05, p < .01, p < .001. Hierarchical multiple regression analyses examined whether L1 background (tone vs. non- tone language), prior exposure to English and French, digital literacy, or device access moderated treatment effects. For pronunciation outcomes, a significant L1 × Group interaction emerged, β = .24, t(112) = 2.68, p = .009, indicating that tone language speakers benefited more from ASR training (adjusted M = 76.8%) than non-tone speakers (adjusted M = 71.2%), whereas no L1 difference was observed in the control condition. This pattern likely reflects the greater phonological distance between learner’s tone language backgrounds and the prosodic features of English and French, including English stress patterns and French intonation, creating larger room for improvement through explicit training. Digital literacy significantly predicted outcomes across all experimental groups, β = .31, t(112) = 3.54, p< .001, but did not interact with treatment condition, suggesting that while higher digital proficiency facilitated AI system use, the relative advantage of AI instruction was maintained across literacy levels. Device access (smartphone-only vs. laptop + smartphone) showed no significant main or interaction effects, indicating successful adaptation of the interface for mobile-constrained environments. Table 6 Hierarchical Regression: Moderation of Treatment Effects Predictor Step 1 β Step 2 β Step 3 β Pre-test Pronunciation .54 .52 .51 Group (AI vs. Control) .42 .38 .35 L1 Background (Tone Language) — .12 .15 Digital Literacy — .31 .29 Device Access — .05 .04 IJEE IJEE www.ijee.io Predictor Step 1 β Step 2 β Step 3 β Group × L1 — — .24 Group × Digital Literacy — — .08 Group × Device Access — — .06 R2 .29 .38 .42 ΔR2 — .09 .04 Note. Dependent variable: Post-test Word Recognition Accuracy. p< .05, p< .01, p< .001. Table 7 Summary of Hypothesis Testing (H1–H5) Hypothesis Statement Statistical Test Key Result Decision H1 Students receiving AI- assisted pronunciation training will show greater improvement in pronunciation outcomes than those receiving traditional instruction. ANCOVA (Table 2) Significant group effect for accentedness and comprehensibility (F ≈ 11.08–11.51, p< .001); large effect sizes (d ≈ 0.94–0.98). Supported H2 Students receiving AI- based grammar instruction will demonstrate greater improvement in grammar proficiency than those receiving traditional instruction. ANCOVA (Table 2) Significant group effect for grammar proficiency (F = 6.17, p = .001; d ≈ 0.98). Supported H3 Students receiving combined AI instruction (ASR + NLP) will demonstrate greater overall language improvement than those receiving single-mode AI or traditional instruction. Post-test adjusted means (Table 3) Combined AI group showed the highest scores across all outcomes (d ≈ 1.02–1.24). Supported IJEE IJEE www.ijee.io Hypothesis Statement Statistical Test Key Result Decision H4 Pronunciation, grammar, and listening gains will be positively associated with global speaking improvement. Correlation analysis (Table 5) Significant correlations between speaking gain and pronunciation (r = .68***), grammar (r = .52***), and listening (r = .61***). Supported H5 Learner characteristics (e.g., digital literacy and L1 background) will moderate the effect of AI- assisted language learning. Hierarchical regression (Table 6) Digital literacy predicted pronunciation gain (β ≈ .29–.31); significant Group × L1 interaction (β = .24, p< .01). Partially Supported 7. Discussion This study examined whether adaptive AI integrating Automatic Speech Recognition and Natural Language Processing could improve language learning among multilingual African learners. Overall, the evidence suggests that AI-assisted instruction enhanced pronunciation, grammatical competence, listening comprehension, and speaking performance more effectively than conventional teacher-led instruction. More importantly, integrating ASR and NLP within a single adaptive learning environment produced stronger learning gains than using either technology independently; indicating that personalized multimodal feedback can support complementary aspects of second-language acquisition. These findings extend existing Computer-Assisted Language Learning research by demonstrating the effectiveness of integrated AI within multilingual African classrooms, a context that remains underrepresented in current literature. 7.1 ASR-Assisted Instruction Enhanced Pronunciation Development Learners who received ASR-supported pronunciation instruction achieved substantially greater improvements in pronunciation accuracy, comprehensibility, and speaking fluency than those receiving conventional instruction. This finding is consistent with previous studies showing that Computer-Assisted Pronunciation Training enables learners to benefit from immediate corrective feedback and repeated opportunities for self-correction (McCrocklin, 2021; Sun, 2023). It also supports the earlier work of Neri et al. (2002), who argued that speech recognition technologies are most effective when integrated with sound pedagogical practice rather than replacing teacher guidance. The effectiveness of the ASR intervention can be explained through Swain's (1985) Output Hypothesis. Producing spoken language, receiving immediate diagnostic feedback, and revising pronunciation repeatedly encourage learners to notice gaps between their production and target-language norms. In multilingual African classrooms, where teachers often have limited opportunities to provide individualized pronunciation coaching because of large class sizes, adaptive ASR systems effectively supplement classroom instruction by delivering continuous personalized practice. This suggests that AI-supported pronunciation training offers a practical solution to one of the most persistent challenges in language education across resource-constrained multilingual settings. IJEE IJEE www.ijee.io 7.2 Adaptive NLP Strengthened Grammatical Competence The grammar-focused intervention also produced meaningful improvements in learners' grammatical accuracy compared with conventional instruction. This finding agrees with research demonstrating that NLP-based grammar systems improve language learning by providing immediate, individualized, and context-sensitive feedback rather than simply identifying errors (Amaral et al., 2011; Wang et al., 2024). Similar conclusions have been reported in studies of automated language assessment, including Educational Testing Service's e-rater® system, where formative AI feedback supports continuous revision and language development. These improvements may be explained by the adaptive nature of the NLP system. Instead of presenting correct answers alone, the platform encouraged learners to revise their own writing through multiple feedback cycles, promoting deeper grammatical understanding and self-regulated learning. Such iterative learning reflects contemporary theories of personalized instruction, where continuous feedback enables learners to construct grammatical knowledge progressively rather than through isolated correction. Consequently, adaptive NLP appears to complement classroom teaching by extending opportunities for independent practice beyond the limitations of teacher-provided feedback. 7.3 Integrated ASR and NLP Produced Synergistic Learning Benefits The most important finding of this study is that learners receiving the combined AI intervention consistently outperformed those receiving only pronunciation instruction, only grammar instruction, or conventional teaching. Rather than producing merely additive effects, integrating ASR and NLP appeared to generate synergistic improvements across pronunciation, grammar, listening comprehension, and overall speaking performance. This outcome supports Bloom's (1984) proposition that individualized instruction substantially enhances learning and demonstrates how adaptive AI can approximate important characteristics of one-to-one tutoring. It also aligns with Bernacki et al. (2021) and Shemshack and Spector (2020), who argue that adaptive educational technologies are most effective when they personalize multiple dimensions of learning simultaneously. By integrating spoken and written language support within one adaptive environment, the intervention enabled learners to strengthen interconnected language skills, resulting in broader communicative competence than could be achieved through isolated instructional approaches. These findings therefore suggest that future AI-supported language learning systems should adopt integrated rather than single-skill instructional designs. 7.4 Pronunciation Practice Facilitated Listening Development An important secondary finding was that learners who improved their pronunciation also demonstrated significant gains in listening comprehension, despite not receiving explicit listening instruction. This relationship suggests that productive and receptive language skills develop interactively rather than independently. This finding is consistent with Krashen's (1985) Input Hypothesis and Swain's (1985) Output Hypothesis. Repeated pronunciation practice likely strengthened learners' phonological representations, making spoken input easier to recognize and interpret during listening tasks. Similar transfer effects have been reported by Sun (2023), who found that improvements in pronunciation frequently extend to broader communicative competence. The present study extends this evidence to multilingual African classrooms, indicating that pronunciation instruction may simultaneously strengthen learners' ability to perceive and process spoken language. 7.5 Implications for Multilingual African Language Education The findings have important implications for language education in multilingual African contexts characterized by linguistic diversity, large class sizes, and limited instructional IJEE IJEE www.ijee.io resources. Rather than replacing teachers, adaptive AI functioned as an instructional support system that provided continuous personalized feedback, allowing learners to practice independently while enabling teachers to focus on higher-order instructional activities. These findings reinforce recommendations by UNESCO (2021) and the OECD (2023), which advocate the responsible integration of AI to improve educational quality and equity. They also respond to calls by Fu et al. (2025) and Abano and Estremera (2026) for greater empirical evidence from African educational contexts. The present study demonstrates that adaptive AI can effectively support multilingual learners even in resource-constrained environments, suggesting considerable potential for expanding equitable access to personalized language instruction across Africa. 7.6 Ethical Considerations and Future AI Development Although adaptive AI proved effective overall, the moderating influence of learners' first- language background indicates that AI systems may not benefit all linguistic groups equally. Differences in pronunciation outcomes suggest that speech recognition algorithms remain sensitive to linguistic variation, highlighting the importance of culturally responsive AI design. This observation supports concerns raised by Baker et al. (2023) regarding algorithmic bias in educational AI and reinforces the ethical principles proposed by Floridi and Cowls (2019), UNESCO (2021), and Molenaar and Sleegers (2023), who emphasize fairness, transparency, inclusiveness, and collaborative governance. Expanding speech and language datasets to include greater representation of African languages and accents should therefore become a priority for future AI development. Such efforts will help ensure that adaptive language-learning technologies remain equitable, culturally responsive, and effective across diverse multilingual educational contexts. 8. Conclusion This study examined the effectiveness of an adaptive Artificial Intelligence (AI) framework integrating Automatic Speech Recognition and Natural Language Processing to provide personalized pronunciation and grammar instruction in multilingual Nigerian classrooms. Using a quasi-experimental pre-test–post-test control group design, the study evaluated whether AI-assisted pronunciation instruction, AI-assisted grammar instruction, and their combined implementation improved learners' pronunciation, grammatical proficiency, listening comprehension, and speaking performance more effectively than conventional teacher-led instruction. The findings demonstrate that adaptive AI-assisted instruction significantly enhanced language-learning outcomes across all measured domains. Although both ASR-based pronunciation training and NLP-supported grammar instruction produced meaningful improvements, the integrated ASR–NLP intervention consistently generated the greatest gains, indicating that combining complementary AI technologies provides broader benefits than implementing either approach independently. These results suggest that adaptive learning systems capable of delivering continuous, individualized feedback can effectively support communicative language development in multilingual classrooms where opportunities for personalized instruction are often limited. The study contributes to Computer-Assisted Language Learning by providing empirical evidence from multilingual African classrooms, an educational context that remains underrepresented in AI-in-education research. It further demonstrates that AI-assisted language learning is most effective when grounded in established pedagogical principles rather than technology alone. By integrating adaptive learning, automated feedback, and peer- supported learning within a single instructional framework, the study illustrates how AI can complement classroom teaching while supporting learner engagement, self-regulated learning, and individualized instruction. IJEE IJEE www.ijee.io The findings also have practical implications for language education. Adaptive AI should be viewed as a tool that supports rather than replaces teachers by extending personalized learning opportunities beyond what is feasible in large multilingual classrooms. Investments in AI-enabled learning infrastructure, teacher capacity building, and culturally responsive educational technologies may therefore improve language instruction, particularly in resource-constrained educational systems. At the same time, the observed influence of learners' linguistic backgrounds and digital literacy highlights the importance of designing AI systems that are equitable, inclusive, and representative of diverse African languages and speech varieties. This study is not without limitations. The research was conducted within a single educational setting using a relatively modest sample and a fourteen-week intervention, which may limit the generalizability and long-term interpretation of the findings. Future studies should involve larger multi-site and cross-national samples, longer intervention periods, indigenous African languages, and emerging generative AI technologies to further examine the effectiveness, fairness, and sustainability of AI-assisted language learning across multilingual educational contexts. Overall, the study provides evidence that integrating ASR and NLP within an adaptive learning framework offers a practical and effective approach to improving pronunciation, grammar, listening, and speaking skills in multilingual African classrooms. As AI technologies continue to evolve, their responsible and pedagogically informed integration has considerable potential to promote more personalized, inclusive, and effective language education while contributing to the advancement of AI-supported language learning in diverse educational settings. IJEE IJEE www.ijee.io References Abano, R. C., & Estremera, M. L. (2026). Teaching languages with machines: A systematic review on the effectiveness of AI in ESL learning. International Journal of Technology in Education, 9(2), 431–446. https://doi.org/10.46328/ijte.5196 Amaral, L., Meurers, D., & Ziai, R. (2011). Analyzing learner language: Towards a flexible natural language processing architecture for intelligent language tutors. Computer Assisted Language Learning, 24(1), 1–16. https://doi.org/10.1080/09588221.2010.520674 Baker, R. S., Hawn, M. A., & Lee, S. (2023). Algorithmic bias: The state of the situation and policy recommendations. In Organisation for Economic Co-operation and Development , OECD Digital Education Outlook 2023: Towards an effective digital education ecosystem. OECD Publishing. https://doi.org/10.1787/c74f03de-en Bernacki, M. L., Greene, M. J., & Lobczowski, N. G. (2021). A systematic review of research on personalized learning: Personalized by whom, to what, how, and for what purpose(s)? Educational Psychology Review, 33(4), 1675–1715. https://doi.org/10.1007/s10648-021-09615-8 Bloom, B. S. (1984). The 2-sigma problem: The search for methods of group instruction as effective as one-to-one tutoring. Educational Researcher, 13(6), 4–16. https://doi.org/10.3102/0013189X013006004 Educational Testing Service. (2024). e- rater® automated scoring engine. https://www.ets.org Fathali, S., & Emadi, A. (2022). Purpose, outcome, and methodology in computer-assisted language learning publications: A systematic review. The JALT CALL Journal, 18(2), 211–237. https://doi.org/10.29140/jaltcall.v18n2.603 Floridi, L., & Cowls, J. (2019). A unified framework of five principles for AI in society. Harvard Data Science Review, 1(1). https://doi.org/10.1162/99608f92.8cd550d1 Fu, Y., Weng, Z., & Wang, J. (2025). Examining AI use in educational contexts: A scoping meta-review and bibliometric analysis. International Journal of Artificial Intelligence in Education, 35(3), 1388–1444. https://doi.org/10.1007/s40593-024-00442-w Godwin-Jones, R. (2021). Emerging technologies for language learning. Language Learning & Technology, 25(3), 1–16. https://www.lltjournal.org Krashen, S. D. (1985). The input hypothesis: Issues and implications. Longman. https://openlibrary.org/works/OL1981212W/The_input_hypothesis McCrocklin, S. M. (2021). Pronunciation learning strategies and technology-mediated feedback in second language pronunciation development. In J. Levis & M. Reed (Eds.), The handbook of second language pronunciation (pp. 445–460). Wiley- Blackwell. https://doi.org/10.1002/9781119808775.ch28 Molenaar, I., & Sleegers, P. (2023). Multi-stakeholder collaboration and co-creation: Towards responsible application of AI in education. In Organisation for Economic Co-operation and Development , OECD Digital Education Outlook 2023: Towards an effective digital education ecosystem. OECD Publishing. https://doi.org/10.1787/c74f03de-en Mustafa, M. Y., Tlili, A., Lampropoulos, G., Huang, R., Altinay, F., Altinay, Z., Sharma, R. C., Agyemang, B., Burgos, D., & Kinshuk. (2024). A systematic review of literature reviews on artificial intelligence in education : A roadmap to a future research agenda. Smart Learning Environments, 11, Article 59. https://doi.org/10.1186/s40561-024-00350-5 Neri, A., Cucchiarini, C., & Strik, H. (2002). Feedback in computer assisted pronunciation training: Technology push or demand pull? In Proceedings of the International Symposium on Automatic Detection of Errors in Pronunciation Training. https://www.isca-archive.org IJEE IJEE www.ijee.io Organisation for Economic Co-operation and Development. (2023). OECD digital education outlook 2023: Pushing the frontiers with AI, blockchain and robots. OECD Publishing. https://www.oecd.org/education/digital-education-outlook/ Shemshack, A., & Spector, J. M. (2020). A systematic literature review of personalized learning terms. Smart Learning Environments, 7, Article 33. https://doi.org/10.1186/s40561-020-00140-9 Sun, W. (2023). The impact of automatic speech recognition technology on second language pronunciation and speaking skills of EFL learners: A mixed methods investigation. Frontiers in Psychology, 14, Article 1210187. https://doi.org/10.3389/fpsyg.2023.1210187 Swain, M. (1985). Communicative competence: Some roles of comprehensible input and comprehensible output in its development. In S. Gass & C. Madden (Eds.), Input in second language acquisition (pp. 235–253). Newbury House. UNESCO. (2021). AI and education: Guidance for policy-makers. https://unesdoc.unesco.org/ark:/48223/pf0000376709 Vygotsky, L. S. (1978). Mind in society: The development of higher psychological processes. Harvard University Press. https://www.hup.harvard.edu/books/9780674576292/mind- in-society Wang, S., Wang, F., Zhu, Z., Wang, J., Tran, T., & Du, Z. (2024). Artificial intelligence in education: A systematic literature review. Expert Systems with Applications, 252, Article 124167. https://doi.org/10.1016/j.eswa.2024.124167