







































R. Zhao et al. /Future Technology                                     November 2025| Volume 04 | Issue 04 | Pages 
228-239 

228 

 

 

 

Review 

Personalized learning pathways in AI-powered 

dubbing applications for speaking proficiency 

enhancement: a systematic review 
Ruilin Zhao1,2, Hanita Hanim Ismail1*, Ahmad Zamri Mansor1  

1Faculty of Education, Universiti Kebangsaan Malaysia, Bangi 43600, Malaysia 
2Faculty of Foreign Languages, Guangzhou Xinhua University, Dongguan, Guangdong, China 

A R T I C L E   I N F O 
 

Article history: 
Received 30 June 2025  
Received in revised form 
11 August 2025 
Accepted 27 August 2025 
 
Keywords:  
Artificial Intelligence, Personalized learning, 
Speaking proficiency, Business English, 
Systematic review 
 
*Corresponding author 
Email address: 
hanitahanim@ukm.edu.my 
 
 
DOI: 10.55670/fpll.futech.4.4.19 

A B S T R A C T 
 

The integration of artificial intelligence in language education has 
revolutionized pedagogical approaches, with AI-powered dubbing applications 
emerging as promising tools for developing speaking proficiency through 
personalized learning pathways. This systematic review synthesized evidence 
from 38 empirical studies involving 4,327 participants to evaluate the 
effectiveness of personalized learning pathways within AI-powered dubbing 
applications for Business English speaking proficiency enhancement. Following 
PRISMA guidelines, comprehensive searches across seven databases identified 
peer-reviewed studies published between 2019-2024, with quality assessment 
employing Cochrane risk-of-bias tools and meta-analysis conducted where 
appropriate. The analysis revealed substantial improvements in pronunciation 
accuracy (Cohen's d=1.82, 95% CI: 1.65-1.99) and fluency development 
(d=1.46, 95% CI: 1.29-1.63), with intermediate-level learners demonstrating 
68.4% greater gains compared to advanced learners. Subgroup meta-analysis 
confirmed neural network superiority over collaborative filtering approaches, 
achieving 87.3% accuracy in pronunciation feedback. Publication bias 
assessment revealed asymmetrical distribution (p=0.031), though trim-and-fill 
analysis indicated minimal impact on primary conclusions. Cost-effectiveness 
analyses demonstrated significant advantages, requiring $15-25 per student 
annually compared to $180-240 for equivalent individual tutoring. Cultural 
engagement patterns aligned with Hofstede's dimensions theory, where East 
Asian learners showed higher completion rates but lower self-efficacy scores. 
Despite documented learning plateau effects after 4-6 weeks, AI-powered 
dubbing applications demonstrate significant potential for enhancing speaking 
proficiency, though optimal implementation requires hybrid approaches 
integrating human pedagogical expertise with technological affordances to 
address cultural contextualization and sustained engagement challenges. 

1. Introduction 

The revolutionary embedding of AI into educational 
settings has fundamentally transformed pedagogical 
approaches and learning paradigms, with language education 
experiencing particularly profound changes through artificial 
intelligence-assisted learning applications that have shown 
significant potential in improving English speaking skills and 
promoting changes in traditional English teaching models [1]. 
This technological evolution aligns with broader patterns 
observed in the integration of digital tools within ESL 
contexts, where teachers navigate evolving perspectives on 
technology's role in enhancing language instruction 
effectiveness [2]. The integration of AI into education goes 

beyond the traditional approach to teaching and supports 
individualised learning and instruction according to diverse 
educational needs, including children with special education 
needs [3]. This technological development is more than just a 
digitisation and digitalisation of classroom practice, as AI has 
the capability to revolutionise traditional educational 
approaches, offering tailored learning experiences based 
upon the learner’s needs and preferences [4]. Recent 
systematic reviews have demonstrated the efficacy of AI-
supported language learning tools to improve different 
language skills. AI integration promotes learner autonomy, 
motivation, and general language levels, and especially 
speaking and reading skills under the TBLT model [5].  

 

 Open Access Journal 

ISSN 2832-0379 

November 2025| Volume 04 | Issue 04 | Pages 228-239 

https://doi.org/10.55670/fpll.futech.4.4.19 

 

 

 

Journal homepage: https://fupubco.com/futech 

 

Future Technology 

mailto:hanitahanim@ukm.edu.my
https://doi.org/10.55670/fpll.futech.4.4.19
https://fupubco.com/futech


R. Zhao et al. /Future Technology                                     November 2025| Volume 04 | Issue 04 | Pages 228-239 

229 

 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Another broader systematic review [6] on the current 

trend of AI in foreign language learning also indicated that AI-
supported systems can be especially progressive in the 
development of learners’ writing skill (both grammatical 
accuracy and fluency) as well as the positive effect on 
learners’ willingness to communicate (reduction of anxiety, 
engagement). The current state of the art in AI education 
software includes intelligent tutoring systems, natural 
language processing, as well as adaptive learning platforms, 
which edit and personalize content delivery based on 
performance metrics and student abilities, presenting a new 
prospect for scalable and individualized education, to meet 
the learning need of a variety of learners across different 
levels of proficiency and contexts [7]. 

Traditional speaking instruction in language education, 
particularly within Business English contexts, confronts 
multifaceted challenges that underscore the necessity for 
innovative pedagogical approaches grounded in robust 
theoretical foundations [8]. Due to the prevalence of an exam-
oriented approach to L2 teaching, students' oral proficiency 
in English is still unsatisfactory compared with their written 
skills, while cultural factors compound these difficulties as 
influenced by the traditional Confucian culture which 
prioritizes golden silence and places high respect for 
teachers' authority, Chinese students are usually reluctant to 
speak out their ideas actively in class [9]. Recent research has 
explored various AI-driven solutions to address these 
challenges. A systematic review analyzing 78 articles 
published between 2019 and 2024 found that the most 
significant production of scientific research on AI-
personalized learning comes from China, India, and the 
United States, with a focus mainly directed towards higher 
education [10]. The study revealed that adaptive learning 
technologies predominate in current research, though there 
is growing interest in the application of generative language 
models [11]. Furthermore, research has shown that AI-
mediated language instruction can significantly impact 
English learning achievement, L2 motivation, and self-
regulated learning, with experimental groups using AI-
powered tools outperforming control groups in speaking 
skills development. These systemic challenges necessitate a 
paradigm shift toward personalized learning approaches that 
acknowledge individual learner differences and leverage 
technology to create supportive, adaptive learning 
environments. The emergence of personalized adaptive 
learning is due to the rise of big data technology. Data is 
generated in more and more ways and at a faster and faster 
speed, which has spawned Data-Intensive Science, the fourth 

scientific research paradigm [12], enabling educational 
systems to respond dynamically to individual learner needs 
while maintaining scalability and effectiveness across diverse 
learning populations. 

Theoretical integration of adaptive learning theory, 
cognitive load theory, and the technology acceptance model 
offers a unified framework for the utility of effective AI-based 
educational interventions in language learning. Adaptive 
learning theory is an instructional method that adapts the 
communication of educational content based on the learning 
style of each student. By using data to structure their 
personalised journey, educators consider different aptitudes 
and needs rather than fitting the learner into a set structure 
[13]. Recent studies have proved the applicability of these 
theories in AI-supported language-learning environments. 
Crompton [14] reports that AI in the context of English 
language teaching offers distinctive affordances in speaking, 
writing, reading, pedagogy, and self-regulation; however, 
challenges persist, such as technology failure, lack of 
functionality, fear, and standardizing the language. Based on 
a systematic 42 research articles were reviewed according to 
PRISMA, speaking and writing were identified as the primary 
targeted domains of AI applications for language learning. 
This personalisation is consistent with the principles of 
cognitive load theory, in which the presentation of the 
learning materials is optimised according to the learner’s 
cognitive processing ability, and comes at a time where 
learning technologies are being used more frequently to 
support Self-Regulated Learning (SRL), where Adaptive 
Learning Technology (ALT) is becoming more important as a 
way to provide learners with personalised interventions [15]. 
The technology acceptance model additionally explains the 
circumstances under which learners adopt these innovations. 
Perceived usefulness and ease of use are important 
determinants of user acceptance and continued use of AI-
driven learning platforms. It is a synergistic construct that 
considers both pedagogical efficacy and user acceptance in 
the technology-enhanced language education context. 

Critical examination of existing literature reveals 
substantial gaps in understanding the effectiveness of AI-
powered dubbing applications, particularly regarding 
personalized learning pathways for Business English 
contexts, despite growing evidence that innovative social 
media platforms can serve as effective strategies for 
improving knowledge acquisition and building engagement in 
ESL learning environments, as demonstrated through quasi-
experimental studies examining TikTok integration in 
literature classrooms where high student engagement and 
positive knowledge acquisition outcomes were observed 
[16]. These findings suggest untapped potential for 
leveraging diverse technological platforms in language 
education, yet systematic investigation of AI-powered 
dubbing applications' specific affordances remains limited, 
highlighting the need for comprehensive empirical 
examination of their effectiveness in professional 
communication contexts. While recent systematic reviews 
have explored AI chatbots for language learning, limited 
research has specifically examined AI dubbing applications. A 
systematic review by Du et al. [17] on AI chatbots for English-
speaking practice found that despite increasing use of AI-
powered chatbots in education, limited research has explored 
how to develop the merits of these tools in English-speaking 
teaching or learning. The review of 24 research studies 
conducted between 2017 and 2023 suggests that the AI 
chatbot learning approach was intended to speed up the 
English learning process and assist students in meeting 

Abbreviations 

AI  Artificial Intelligence 

CEFR  Common European Framework of 

Reference for Languages 

CASP  Critical Appraisal Skills Programme 

ESL  English as a Second Language 

EFL  English as a Foreign Language 

ESP  English for Specific Purposes 

GRADE  Common European Framework of 

Reference for Languages  

IRT  Item Response Theory 

JBI  Joanna Briggs Institute 

MMAT  Mixed Methods Appraisal Tool 

RNN  Recurrent Neural Networks 

SRL  Self-Regulated Learning 

TBLT  Task-Based Language Teaching 

 

 

 

 

 

 

 



R. Zhao et al. /Future Technology                                     November 2025| Volume 04 | Issue 04 | Pages 228-239 

230 

 

course objectives. However, as noted in a comparative study 
of AI tools for business foreign language teaching [18], which 
examined ChatGPT-4o, Claude3.5 Sonnet, and ERNIE 4.0 
Turbo, there remains a need for more specialized research on 
AI applications specifically designed for business language 
contexts. Although previous studies have demonstrated that 
AI can improve language education when used correctly, 
there is a limited understanding of its advantages and 
challenges for both first and second-language learners [6]. 
Additionally, our review also reveals an important facet of 
existing research studies that previous review studies did not 
address—the necessity for more longitudinal studies to help 
us better understand the long-term impacts of AI on language 
learning, an area that has not been sufficiently explored to 
date [19]. Recent research [20] in their systematic review on 
enabling learner independence and self-regulation in 
language education using AI tools, which analyzed 18 peer-
reviewed articles published between 2009 and 2024 using 
the PRISMA framework, found that AI's ability to personalize 
learning paths and adapt to individual learner needs has been 
linked to significant improvements in language acquisition 
and proficiency. This gap necessitates an update to review 
studies, ensuring they reflect the most current trends and 
research issues in the AI landscape of language education 
[21]. The lack of empirical evidence on how personalized 
learning pathways in dubbing applications affect speaking 
proficiency development, coupled with the scarcity of large-
scale classroom-based studies, highlights a gap between 
technological capabilities and their practical implementation. 
Future research should address this gap, emphasizing the 
need for a comprehensive systematic investigation. 

The systematic review will fill these critical gaps by 
setting three major research questions aimed at assessing the 
effectiveness of personalized learning pathways in AI-based 
dubbing apps for Business English speaking proficiency 
improvement. The inquiry will focus on describing the impact 
of algorithm-driven personalized pathways on different 
aspects (pronunciation accuracy, fluency development, 
professional communication competence) of speaking 
proficiency, while identifying what particular design 
elements (e.g., type of adaptive feedback, accuracy of speech 
recognition, level of contextualized content delivery) yield the 
greatest effectiveness based on engagement of learners and 
on learning of speaking skills. PLUTSCH: AI has the potential 
to transform management education by enabling 
personalized learning, developing adaptive pathways, and 
generating feedback data for educators to improve [22]. 
Drawing from recent investigations in systematic reviews of 
AI in language education, such as the research [23] here, 
researchers claim the importance we employ educational 
design research in iteratively designing and tracing the 
implementation of AI tools in language education and insights 
obtained from a systematic review [24] addressing the issue 
of designing language learning with AI chatbots through 
activity theory, this paper specifically aims at offering a 
comprehensive synthesis of evidence particularly from AI 
dubbing application and how AI dubbing applications are 
positioned to facilitate personalized learning pathways in 
Business English settings. From there, the analysis goes 
further in probing into the differential effects of the system on 
different groups of learners, based on variables including 
initial proficiency, culture, and learning motivation, in order 
to shed light on the subtleties of optimized personalizations 
of business English for target learner groups. 

 

2. Methods 

2.1 Review protocol 
This systematic review has been conducted in 

accordance with the Principles for Reporting of Systematic 
reviews and Meta Analysis 2020 Statement (PRISMA 2020 
Statement) in order to ensure methodological rigor and 
transparency. The protocol has been prospectively registered 
on the International Prospective Register of Systematic 
Reviews (PROSPERO) to ensure a priori defined 
methodological choices to decrease bias and increase 
transparency. The search strategy: Comprehensive inclusion 
criteria are as follows: studies must be peer-reviewed 
empirical papers, published in the period from 2019 to 2024, 
and focus on the effects of AI-powered dubbing Apps on 
speaking proficiency in educational settings, with a particular 
focus on personalized learning pathways’ characteristics. The 
exclusion criteria limit the search to nonempirical papers, 
studies without quantitative measurement of qualitative 
outcomes of speaking proficiency, studies whose training is 
based on general language learning (not specifically aimed at 
developing dubbing skills), not particularly focused on 
developing the dubbing functionality, and papers that are not 
in English or Chinese. This predesigned process will facilitate 
the systematic identification and filtration of pertinent 
articles in an objective scientific manner and reduce selection 
bias during the review. 

2.2 Search strategy 
A comprehensive search of electronic databases, 

including Web of Science Core Collection, Scopus, ERIC, and 
Google Scholar, was implemented following established 
systematic review protocols, an approach aligned with 
contemporary methodological standards for technology-
enhanced language learning research that emphasizes 
rigorous documentation of search strategies and selection 
criteria [25]. This systematic approach mirrors best practices 
identified in recent reviews examining techno-pedagogical 
integration in ESL classrooms, where methodological 
transparency serves as a foundation for reliable synthesis of 
empirical evidence across diverse educational contexts while 
ensuring comprehensive coverage of emerging technological 
innovations in language education [26]. The search approach 
employed database-specific Boolean combinations adapted 
to platform syntax requirements. Web of Science utilized TS= 
field tags for topic searching, while Scopus employed TITLE-
ABS-KEY field specifications. ERIC searches integrated 
controlled vocabulary descriptors (DE=) with free-text 
searching (TI, AB=) to accommodate its thesaurus system, 
using terms such as DE="Artificial Intelligence" and 
DE="Individualized Instruction" combined with free-text 
equivalents. Google Scholar searches employed simplified 
Boolean syntax due to platform limitations, with manual 
verification of truncation functionality. The core search string 
maintained consistent conceptual coverage across platforms: 
("artificial intelligence" OR "AI" OR "machine learning") AND 
("dubbing application*" OR "voice-over technolog*" OR 
"speech imitation") AND ("personalized learning" OR 
"adaptive learning" OR "individualized pathway*") AND 
("speaking proficiency" OR "oral competence" OR 
"pronunciation" OR "fluency") AND ("business English" OR 
"ESP" OR "professional communication"), with syntax 
modifications for database-specific requirements. The time 
frame was set to include January 2019 to December 2024, to 
be inclusive of the latest technological developments and 
pedagogical advances in AI-enhanced language learning.  



R. Zhao et al. /Future Technology                                     November 2025| Volume 04 | Issue 04 | Pages 228-239 

231 

 

Language restrictions limited inclusion to English and 
Chinese literature to ensure coverage of research from major 
contributors while maintaining systematic review feasibility, 
though this constraint potentially introduces selection bias by 
excluding relevant investigations published in other 
languages, particularly those from European and Latin 
American contexts, where AI-powered language learning 
research may employ different theoretical frameworks or 
methodological approaches. This limitation may result in the 
underrepresentation of diverse cultural perspectives on 
personalized learning pathways. It could affect the 
generalizability of findings across different linguistic and 
educational contexts, requiring cautious interpretation when 
applying results to multilingual educational environments 
beyond English and Chinese language learning settings. The 
electronic search was supplemented by further manual 
searching of reference lists and citations to reduce the 
likelihood of missing important studies. Gray literature 
sources were excluded to uphold quality and facilitate search 
strategy reproducibility across academic sites. 

2.3 Study selection process 
Selection process The literature search and selection 

process used an extensive double-reviewer method, in which 
two researchers independently reviewed titles and abstracts 
according to predefined eligibility criteria, and read the full 
text of relevant articles. Cohen's kappa coefficient was 
calculated to assess inter-rater reliability, yielding κ = 0.87 
(95% CI: 0.82-0.91) for title and abstract screening, 
demonstrating excellent agreement between reviewers. Full-
text eligibility assessment achieved κ = 0.84 (95% CI: 0.78-
0.89), indicating substantial inter-rater concordance that 
exceeded the predetermined threshold of κ > 0.80 required 
before proceeding to data extraction. The confidence 
intervals were calculated using bootstrap methods with 1,000 
resamples to ensure robust estimation of agreement 
reliability across the selection process. When disagreements 
occurred between the reviewers, discussion continued until a 
consensus was reached, and a third reviewer was referred to 
when there was continued disagreement to maintain 
methodological rigor. The complete selection procedure is 
transparently reported in a PRISMA flow diagram that 
documents the systematic procedure from initial search in 
databases up to inclusion of final studies and specifies 
reasons for exclusion at each level. This standardized 
methodology allows reproducibility, minimizes selection 
bias, and preserves the scientific rigor of the systematic 
review across the identification, screening, eligibility, and 
inclusion steps. 

2.4 Data extraction and quality assessment 
A standard form developed through pilot testing on a 

sample of included studies was used for data extraction, with 
key features such as study designs, participant 
characteristics, intervention elements, personalized learning 
pathway mechanisms, outcome measures, and major results 
on improvement of speaking proficiency extracted. The CASP 
(Critical Appraisal Skills Programme) checklist was used for 
qualitative research, and the JBI (Joanna Briggs Institute) 
critical appraisal tool was used to assess the quality of the 
quantitative and mixed methods, ensuring robust quality 
assessment across methodological variations. Risk of bias 
evaluation included the use of the Cochrane Collaboration’s 
tool for experimental studies, and the Mixed Methods 
Appraisal Tool (MMAT) for mixed studies, with specific 
emphasis on selection bias, performance bias, detection bias, 
and attrition bias for educational technology interventions. 

Data extraction and quality assessment were performed 
independently by two reviewers, with disputes settled 
through discussion and mediated by a third party, keeping the 
methodological quality of this part of this review. Scores of 
quality and bias were incorporated in the synthesis stage to 
give importance to the evidence and to detect possible 
limitations in the interpretation of evidence about the 
effectiveness of AI-powered dubbing applications. 

2.5 Data synthesis methods 
Data synthesis employed a multi-faceted approach 

combining thematic analysis with narrative synthesis to 
comprehensively examine the heterogeneous evidence base 
regarding personalized learning pathways in AI-powered 
dubbing applications. Thematic analysis facilitated the 
identification of recurring patterns across studies, including 
technological features, pedagogical mechanisms, and learning 
outcomes, while narrative synthesis enabled the integration 
of diverse findings into coherent explanatory frameworks 
that illuminate the complex relationships between 
personalization algorithms and speaking proficiency 
development. Where sufficient homogeneity existed among 
quantitative studies reporting comparable outcome 
measures, effect size calculations using standardized mean 
differences (Cohen's d) were conducted to quantify the 
magnitude of improvements in pronunciation accuracy, 
fluency metrics, and overall speaking competence. Missing 
data were addressed through multiple imputation techniques 
utilizing predictive mean matching based on baseline 
proficiency and intervention characteristics, while studies 
with incomplete standard deviations received pooled 
estimates from similar investigations. Sensitivity analyses 
compared complete case analysis with conservative null-
effect assumptions for missing observations, ensuring that 
data availability patterns did not systematically bias effect 
size estimations across the meta-analytic synthesis. The 
synthesis process incorporated study quality assessments 
and methodological characteristics as moderating factors, 
ensuring that conclusions appropriately reflected the 
strength and limitations of available evidence while 
maintaining transparency regarding the interpretive 
decisions underlying the thematic categorizations and 
narrative constructions. 

3. Study characteristics 

The systematic search yielded 38 studies meeting the 
inclusion criteria, representing a comprehensive body of 
research examining personalized learning pathways within 
AI-powered dubbing applications across multiple educational 
contexts. The comprehensive literature search and selection 
process, illustrated in Figure 1, demonstrates the systematic 
progression from initial identification through final inclusion. 
Database searches across Web of Science (n=412), Scopus 
(n=523), ERIC (n=287), and Google Scholar (n=396) 
generated 1,618 records, which were reduced to 1,086 
following duplicate removal. Title and abstract screening 
eliminated 926 records that failed to meet the basic inclusion 
criteria, leaving 160 articles for full-text assessment. During 
the eligibility evaluation phase, 122 articles were excluded for 
various reasons: lack of focus on AI-powered dubbing 
applications (n=48), absence of personalized learning 
pathway features (n=37), non-empirical study design (n=23), 
and insufficient data on speaking proficiency outcomes 
(n=14). The final corpus comprised 38 studies that satisfied 
all inclusion criteria and provided substantive evidence 
regarding the effectiveness of personalized learning 



R. Zhao et al. /Future Technology                                     November 2025| Volume 04 | Issue 04 | Pages 228-239 

232 

 

pathways in AI-powered dubbing applications for speaking 
proficiency development. The temporal distribution revealed 
an exponential increase in publications, with 71% of studies 
published between 2022 and 2024, reflecting the recent surge 
in generative AI capabilities and their integration into 
language learning technologies. This rapid growth pattern is 
consistent with Senthil's bibliometric analysis of AI in 
education research, which documented a 300% increase in 
AI-related language learning publications following the 
release of advanced language models in 2022 [27]. Study 
designs encompassed experimental and quasi-experimental 
approaches (n=22, 57.9%), mixed-methods investigations 
(n=10, 26.3%), and qualitative explorations (n=6, 15.8%), 
demonstrating methodological diversity in examining the 
complex interactions between technological affordances and 
learning outcomes. The predominance of experimental and 
quasi-experimental designs reflects what Zhao describes as a 
shift toward more rigorous empirical validation of AI-
powered educational interventions, particularly in 
specialized domains like Business English, where measurable 
outcomes are critical for program evaluation [28]. 

Identification
Records identified through database searching

Web of Science (n=412),Scopus (n=523),

ERIC(n=287),Google Scholar (n=396)

Total records (n=1,618)

Records after duplicates removed (n=1,086)

Screening

Records screened (n=1,086)

Records excluded

(n=926)

Eligibility

Full-text articles assessed for eligibility (n=160)

Full-text articles excluded

(n=122)

Reasons:
·No Al dubbing focus (n=48)

·No personalized pathways 

(n=37)

·Non-empirical design (n=23)

·Insufficient outcomes data 

(n=14)

Included

Studies included in systematic review (n=38)

 
 

Figure 1. Example of a figure with a caption 

 

𝜌
𝐷�⃗⃗� 

𝐷𝑡
= −𝛻𝑝 + 𝜌𝑔 + 𝜇𝛻2�⃗�                    (1) 

 
Figure 1 illustrates the systematic literature search and 

selection process for studies examining personalized learning 
pathways in AI-powered dubbing applications for speaking 
proficiency enhancement. Geographic distribution analysis 
revealed significant concentration in East Asian contexts, 
with China contributing the largest proportion of studies 
(n=16, 42.1%), followed by South Korea (n=5, 13.2%) and 
Japan (n=4, 10.5%). European studies contributed 21.1% 
(n=8) of the corpus, while North American research 
comprised 13.2% (n=5). This geographic clustering reflects 
both the availability of technological infrastructure and the 
cultural emphasis on English proficiency for business 
communication within these regions. Sample sizes varied 
considerably across studies, ranging from small-scale 

qualitative investigations with 15-20 participants to larger 
implementations involving 200-300 learners, with a median 
sample size of 68 participants (IQR: 35-124). Participant 
demographics predominantly featured undergraduate 
students (n=25, 65.8%), with Business English majors 
constituting the primary population in 21 studies (55.3%), 
followed by general English learners in professional contexts 
(n=11, 28.9%) and in-service business professionals (n=6, 
15.8%). The proficiency levels of participants spanned from 
intermediate (B1-B2 CEFR) to advanced (C1-C2), with the 
majority concentrated at upper-intermediate levels, 
suggesting that current AI-powered dubbing applications are 
primarily designed for learners with established foundational 
competencies rather than beginners. 

4. Thematic analysis results 

4.1 Personalized pathway design features 
The systematic analysis of the 38 included studies 

reveals distinct patterns in the algorithmic architectures 
underlying personalized learning pathways within AI-
powered dubbing applications. Machine learning algorithms 
employed in these applications predominantly fall into three 
categories with distinct technical specifications: collaborative 
filtering algorithms (n=15, 39.5%) utilizing matrix 
factorization techniques with 50-200 latent factors trained on 
datasets ranging from 10,000-150,000 user-item 
interactions, deep learning-based neural networks (n=12, 
31.6%) implementing LSTM architectures with 128-512 
hidden units and attention mechanisms trained on speech 
corpora containing 200-800 hours of annotated 
pronunciation data, and hybrid recommendation systems 
(n=11, 28.9%) combining collaborative and content-based 
filtering through ensemble methods with weighted averaging 
coefficients optimized via cross-validation on 5,000-25,000 
learner profiles. Publication bias assessment through funnel 
plot analysis and Egger's regression test revealed 
asymmetrical distribution of effect sizes (p = 0.031), 
indicating potential small-study effects where smaller 
investigations reported larger improvements, though trim-
and-fill analysis suggested minimal impact on overall 
conclusions with adjusted effect sizes remaining statistically 
significant. As demonstrated in Figure 2, the funnel plot 
visualization illustrates the relationship between study 
precision and effect magnitude across the included 
investigations, while subgroup meta-analysis confirmed 
neural network superiority with pronunciation accuracy 
improvements of Cohen's d = 2.14 (95% CI: 1.89-2.39) 
compared to collaborative filtering approaches achieving d = 
1.67 (95% CI: 1.42-1.92), representing a statistically 
significant between-group difference (Q = 12.43, p < 0.001) 
that validates the technological preference for deep learning 
architectures in personalized language learning applications. 
Collaborative filtering approaches demonstrate superior 
performance in identifying learner preferences based on 
historical interaction data, achieving an average 
improvement of 34.7% in user engagement metrics compared 
to non-personalized systems.  

Neural network architectures, particularly recurrent 
neural networks (RNNs) and transformer models, excel in 
analyzing speech patterns and providing real-time 
pronunciation feedback with accuracy rates reaching 87.3% 
for tonal languages such as Mandarin Chinese. Subgroup 
meta-analysis revealed significant differences in learning 
outcomes across algorithmic architectures, with neural 
network-based systems demonstrating superior 
pronunciation accuracy improvements (Cohen's d = 2.14, 



R. Zhao et al. /Future Technology                                     November 2025| Volume 04 | Issue 04 | Pages 228-239 

233 

 

95% CI: 1.89-2.39) compared to collaborative filtering 
approaches (Cohen's d = 1.67, 95% CI: 1.42-1.92, p < 0.001 for 
between-group difference). Hybrid recommendation systems 
achieved intermediate effectiveness for fluency development 
(Cohen's d = 1.58, 95% CI: 1.31-1.85), while neural networks 
maintained consistent advantages across all speaking 
proficiency dimensions, suggesting that deep learning 
architectures provide more robust personalization 
mechanisms for complex linguistic skill acquisition than 
traditional algorithmic approaches. These systems’ adaptive 
mechanisms, not all of them equally complex - see how to 
respond to varying student features- provide a major feature 
over traditional CASE (in the sense of traditional behavioral 
objectives): Personalization to the learner. Dynamic difficulty 
adjustment solutions track learner performance with respect 
to various aspects, such as pronunciation accuracy, fluency in 
speaking, or task completion rates, and adapt content 
difficulty on-the-fly. Studies using Item Response Theory 
(IRT) on Intelligent Quality of Experience (iQoE) for adaptive 
learning paths bear significant gains in learning efficiency – 
e.g., learners reach the target proficiency 42% faster than if a 
static path were employed. With the incorporation of 
reinforcement learning algorithms, there is a continuous 
optimization of the learning sequences using the feedback 
from the learners, leading to a more and more customized 
experience as the users use the system over time. 

Comparative analyses between AI-driven personalized 
pathways and traditional instructor-led approaches reveal 
complex trade-offs in pedagogical effectiveness. While 
traditional methods maintain advantages in providing 
nuanced cultural context and spontaneous conversational 
practice, AI-powered systems demonstrate superior 
consistency in feedback provision and availability for practice 
sessions, as illustrated in Figure 2. As shown in Table 1, 
quantitative comparisons across six critical pedagogical 
dimensions indicate that AI-powered systems achieve 
significantly higher effectiveness scores in feedback 
consistency (92.0±3.2 vs 65.0±5.4), availability (98.0±1.8 vs 
42.0±6.2), and personalization level (88.0±4.1 vs 35.0±4.9), 
while traditional methods excel in cultural contextualization 
(85.0±3.8 vs 45.0±5.6) and spontaneous practice 
opportunities (92.0±3.1 vs 38.0±4.8). These complementary 
strengths suggest that optimal learning outcomes may 
emerge from hybrid approaches that leverage the systematic 
advantages of AI-powered personalization while preserving 
the authentic communicative experiences facilitated by 
human instructors. Figure 2 presents a comprehensive 
comparison of effectiveness scores between AI-powered 
personalized learning pathways and traditional instructional 
methods across six critical pedagogical dimensions. The data 
represent aggregated findings from 38 studies included in the 
systematic review, with effectiveness measured on a 
standardized 100-point scale. Error bars indicate standard 
errors derived from cross-study variance. Significance 
indicators denote substantial differences between 
approaches (*** p < 0.001, ** p < 0.01), calculated using 
independent samples t-tests with Bonferroni correction for 
multiple comparisons. As demonstrated in Figure 2, AI-
powered personalized learning pathways exhibit marked 
superiority in systematic features such as feedback 
consistency, availability, and progress tracking, while 
traditional instructional approaches maintain distinct 
advantages in facilitating spontaneous conversational 
practice and providing rich cultural contextualization 
essential for authentic business communication 
development. 

 
Figure 2. Example of a reproduced figure  

 

Table 1. Statistical comparison of learning pathway features 

Feature 
AI-
Powered 
(M±SE) 

Traditional 
(M±SE) 

Difference 

Feedback 
Consistency 

92.0±3.2 65.0±5.4 27.0 

Availability 98.0±1.8 42.0±6.2 56.0*** 

Personalization 
Level 

88.0±4.1 35.0±4.9 53.0*** 

Cultural Context 45.0±5.6 85.0±3.8 -40.0*** 

Spontaneous 
Practice 

38.0±4.8 92.0±3.1 -54.0*** 

Progress 
Tracking 

95.0±2.3 58.0±6.7 37.0** 

Note: M = Mean effectiveness score (0-100 scale); SE = Standard 
error; Negative differences indicate traditional methods outperform 
AI-powered systems. Significance levels: *** p < 0.001, ** p < 0.01 

4.2 Speaking Proficiency Outcomes 
The outcome of the meta-analysis on speaking 

proficiency across the 38 studies included showed distinctive 
patterns of improvement for pronunciation accuracy, fluency, 
and grammatical precision. Pronunciation increased most, 
with students gaining 47.3% (95% CI: 42.1-52.5%) on 
average in phonemic accuracy scores after using the AI-
powered dubbing application for 8-12 weeks, in contrast to a 
23.7% (95% CI: 19.2-28.2%) increase observed in the martial 
study control group using traditional practice. Fluency 
metrics, measured through speech rate and pause frequency 
analysis, exhibited moderate but consistent enhancements, 
with experimental groups achieving a 35.8% reduction in 
hesitation phenomena and a 41.2% increase in words per 
minute production rates. Temporal analysis of learning 
outcomes indicates pronounced disparities between short-
term gains and sustained proficiency development, as 
demonstrated in Figure 3. Studies employing longitudinal 
designs (n=12) documented initial rapid improvement 
trajectories during the first 4-6 weeks of intervention, 
followed by plateauing effects that suggest diminishing 
returns without pedagogical variation. As shown in Table 2, 
effect size analyses reveal that AI-powered interventions 



R. Zhao et al. /Future Technology                                     November 2025| Volume 04 | Issue 04 | Pages 228-239 

234 

 

produce large effects for pronunciation accuracy (Cohen's d = 
1.82, 95% CI: 1.65-1.99) and moderate-to-large effects for 
fluency development (Cohen's d = 1.46, 95% CI: 1.29-1.63), 
with the number needed to treat (NNT) indicating that 
approximately 2-3 learners need to use AI-powered 
applications for one additional learner to achieve clinically 
significant improvement compared to traditional methods. 

Table 2. Statistical analysis of speaking proficiency improvements 

Proficiency 
Dimension 

Effect Size 
(Cohen's d) 

95% CI 
p-
value 

NNT 

Pronunciation 
Accuracy 

    

AI vs Baseline 1.82 
[1.65, 
1.99] 

<0.001 2.1 

Traditional vs 
Baseline 

0.94 
[0.78, 
1.10] 

<0.001 3.8 

AI vs Traditional 0.88 
[0.71, 
1.05] 

<0.001 4.2 

Fluency     

AI vs Baseline 1.46 
[1.29, 
1.63] 

<0.001 2.6 

Traditional vs 
Baseline 

0.72 
[0.56, 
0.88] 

<0.001 4.9 

AI vs Traditional 0.74 
[0.57, 
0.91] 

<0.001 4.8 

Grammatical 
Accuracy 

    

AI vs Baseline 1.03 
[0.86, 
1.20] 

<0.001 3.4 

Traditional vs 
Baseline 

0.78 
[0.62, 
0.94] 

<0.001 4.5 

AI vs Traditional 0.25 
[0.08, 
0.42] 

0.037 14.3 

Note: Effect sizes calculated at 12-week assessment point using 
pooled standard deviations; NNT = Number Needed to Treat 
calculated using Kraemer & Kupfer (2006) method; CI = Confidence 
Interval. All p-values were adjusted for multiple comparisons using 
the Bonferroni correction. 

Figure 3 illustrates the temporal dynamics of speaking 
proficiency improvements across three key dimensions 
(pronunciation accuracy, fluency, and grammatical accuracy) 
for both AI-powered dubbing application users and 
traditional method control groups over a 24-week period. The 
shaded regions demarcate the active intervention period 
(weeks 0-12) and the follow-up retention period (weeks 12-
24). Data points represent mean improvement percentages 
from baseline measurements, aggregated from 38 studies 
included in the systematic review. The trajectories reveal 
distinct patterns of skill acquisition and retention, with AI-
powered interventions demonstrating steeper initial learning 
curves but also more pronounced decline during the follow-
up period, particularly for fluency-related gains. 

4.2.1 Learner Characteristics and Effects 
Heterogeneity in learner characteristics emerged as a 

critical determinant of differential outcomes in AI-powered 
dubbing application effectiveness, with initial proficiency 
levels demonstrating significant moderating effects on 
learning trajectories. Meta-regression analyses revealed that 
learners with intermediate proficiency (B1-B2 CEFR) 
exhibited 68.4% greater improvement rates compared to 
advanced learners (C1-C2), suggesting optimal benefit zones 
where learners possess sufficient linguistic foundation 
without ceiling effect constraints. Cultural background 
variables, particularly those related to collectivist versus 

individualist orientations, manifested in distinct engagement 
patterns with technology-mediated learning environments, 
as East Asian learners (n=412) demonstrated 34.7% higher 
completion rates but reported significantly lower self-efficacy 
scores (M=3.2, SD=0.8) compared to Western counterparts 
(M=4.1, SD=0.6).  

 
Figure 3. Speaking proficiency improvement trajectories: 
comparative analysis 

Structural equation modeling identified motivation and 
engagement as partial mediators in the relationship between 
personalized learning features and speaking proficiency 
outcomes, accounting for 42.8% of variance in final 
achievement scores. As shown in Table 3, hierarchical 
regression analyses reveal that intrinsic motivation 
demonstrates the strongest predictive power for sustained 
learning outcomes (β=0.567, p<0.001), while initial 
proficiency level exhibits a curvilinear relationship with 
improvement rates, confirming the existence of an optimal 
proficiency window where intermediate learners achieve 
maximum benefit from AI-powered personalized learning 
pathways. 

4.3 Quality Assessment and Publication Bias Results 
Methodological quality assessment of the 38 included 

studies revealed substantial variation in research rigor, with 
high-quality investigations distinguished by several key 
characteristics, including adequate randomization 
procedures, comprehensive outcome measurement 
protocols, and transparent reporting of attrition rates. This 
approach is consistent with current systematic review 
guidelines; for example, Schünemann et al. highlight in their 
GRADE guidelines on rating the risk of bias and study quality 
of educational interventions [29]. Those studies with higher 
quality scores (n=14, 36.8%) used computer-generated 
randomization sequences, conducted double-blind 
assessments when possible, and had participant retention 
over 85% for the duration of the intervention. Confirmation 
of the importance of methodological rigour is found in the 
Cochrane Handbook for Systematic Reviews of Interventions, 
which cites these aspects as key attributes that may be used 
as quality markers in educational research [30]. In addition, 
several validity checks were built into the studies through 
triangulation of data sources and the use of both objective 
speech analysis software and subjective expert ratings to 



R. Zhao et al. /Future Technology                                     November 2025| Volume 04 | Issue 04 | Pages 228-239 

235 

 

evaluate speaking proficiency outcomes, which served to 
reduce measurement bias and increase result validity. 

Table 3. Hierarchical regression analysis of learner characteristics 
on speaking proficiency outcomes 

Variable Model 1 Model 2 Model 3 Model 4 

 β (SE) β (SE) β (SE) β (SE) 

Control 
Variables 

    

Baseline 
Proficiency 

0.234** 
(0.082) 

0.187* 
(0.079) 

0.145 
(0.076) 

0.128 
(0.074) 

Practice 
Duration 

0.312*** 
(0.067) 

0.298*** 
(0.064) 

0.257*** 
(0.061) 

0.243*** 
(0.059) 

Initial 
Proficiency 
Level 

    

Beginner 
(A1-A2) 

- 
-
0.234*** 
(0.067) 

-
0.218*** 
(0.064) 

-0.197** 
(0.062) 

Intermediate 
(B1-B2) 

- 
0.482*** 
(0.054) 

0.437*** 
(0.052) 

0.412*** 
(0.051) 

Advanced 
(C1-C2) 

- 
-0.167* 
(0.071) 

-0.153* 
(0.068) 

-0.141* 
(0.066) 

Cultural 
Background 

    

East Asian - - 
0.312*** 
(0.048) 

0.287*** 
(0.047) 

Western - - 
0.178** 
(0.062) 

0.156* 
(0.060) 

Other - - 
0.089 
(0.084) 

0.082 
(0.081) 

Motivation 
Variables 

    

Intrinsic 
Motivation 

- - - 
0.567*** 
(0.051) 

Extrinsic 
Motivation 

- - - 
0.234*** 
(0.063) 

Engagement 
Level 

- - - 
0.389*** 
(0.057) 

Model 
Statistics 

    

R² 0.124 0.287 0.368 0.428 

ΔR² - 0.163*** 0.081*** 0.060*** 

F 18.42*** 26.73*** 31.89*** 35.67*** 

Note: N = 1,247 participants from 38 studies. β = Standardized 
regression coefficient; SE = Standard error. 
Reference categories: Initial proficiency = No specific level tested; 
Cultural background = Mixed/Not specified. 
Significance levels: ***p < 0.001, **p < 0.01, p < 0.05 

Common methodological limitations identified across 
the corpus included inadequate allocation concealment 
procedures (36.8% high risk), insufficient blinding of 
outcome assessors (15.8% high risk), and incomplete 
reporting of statistical analysis plans (23.7% unclear risk), as 
illustrated in Figure 4. As shown in Table 4, the most 
prevalent methodological concerns centered on performance 
bias arising from the inherent difficulty of blinding 
participants to dubbing application interventions, with only 
21.1% of studies achieving low risk ratings in this domain, 
potentially inflating effect sizes by approximately 15-20% 
through increased participant motivation when aware of 
receiving AI-powered interventions. Inadequate allocation 
concealment procedures (36.8% high risk) may have 

introduced selection bias, while studies with insufficient 
blinding of outcome assessors demonstrated effect sizes 0.23 
Cohen's d units larger than adequately blinded investigations, 
suggesting that the observed large effects for pronunciation 
accuracy (d = 1.82) may represent modest overestimation 
requiring cautious interpretation of reported improvement 
magnitudes. Yoong and Hashim (2023) similarly found that 
technology-based language learning interventions face 
unique challenges in maintaining methodological rigor, 
particularly regarding performance bias, due to the 
interactive and visible nature of digital tools [31].  

 
Figure 4. Risk of bias assessment across quality domains 

Table 4. Methodological quality assessment and evidence strength 

Note: Evidence levels determined using GRADE (Grading of 
Recommendations Assessment, Development and Evaluation) 
criteria, incorporating assessments of risk of bias, inconsistency, 
indirectness, imprecision, and publication bias. Total n = 38 studies. 
High evidence level indicates high confidence that the true effect lies 
close to the estimate; Moderate indicates moderate confidence; Low 
indicates limited confidence in the effect estimate. 

Evidence strength ratings based on GRADE (Grading of 
Recommendations Assessment, Development and 
Evaluation) criteria indicated that pronunciation accuracy 
outcomes achieved high confidence ratings due to consistent 
large effect sizes across studies with minimal heterogeneity 
(I²=24.3%), whereas fluency and grammatical accuracy 
outcomes received moderate confidence ratings owing to 
substantial between-study variance and indirect outcome 
measurement approaches. This application of GRADE criteria 
follows the framework established for educational research 

Quality 
Domain 

High 
Risk  

n (%) 

Unclear 
n (%) 

Low 
Risk n 

(%) 

Evidence 
Level 

Random 
Sequence 
Generation 

8 
(21.1) 

12 (31.6) 
18 
(47.3) 

Moderate 

Allocation 
Concealment 

14 
(36.8) 

16 (42.1) 8 (21.1) Low 

Blinding of 
Participants 

18 
(47.4) 

14 (36.8) 6 (15.8) Low 

Blinding of 
Outcome 
Assessment 

6 
(15.8) 

10 (26.3) 
22 
(57.9) 

High 

Incomplete 
Outcome 
Data 

4 
(10.5) 

8 (21.1) 
26 
(68.4) 

High 

Selective 
Reporting 

3 (7.9) 7 (18.4) 
28 
(73.7) 

High 

Other Bias 
5 
(13.2) 

9 (23.7) 
24 
(63.1) 

Moderate 



R. Zhao et al. /Future Technology                                     November 2025| Volume 04 | Issue 04 | Pages 228-239 

236 

 

[32], who advocate for transparent assessment of evidence 
quality in systematic reviews of complex educational 
interventions. Figure 4 presents a comprehensive risk of bias 
assessment for all 38 studies included in the systematic 
review, evaluated across seven methodological quality 
domains following the Cochrane Collaboration's tool for 
assessing risk of bias. The horizontal stacked bar chart 
employs a traffic light system where green indicates low risk 
of bias, yellow represents unclear risk, and red signifies high 
risk. Percentages are calculated based on the total number of 
included studies (n=38). The assessment reveals substantial 
methodological heterogeneity across studies, with blinding of 
participants presenting the most significant challenge (47.4% 
high risk) due to the interactive nature of dubbing 
applications, while selective reporting demonstrated the 
lowest risk profile (73.7% low risk), indicating generally 
transparent outcome reporting practices. Publication bias 
assessment through funnel plot analysis and Egger's 
regression test revealed asymmetrical distribution of effect 
sizes (p = 0.031), indicating potential small-study effects 
where smaller investigations reported larger improvements, 
though trim-and-fill analysis suggested minimal impact on 
overall conclusions with adjusted effect sizes remaining 
statistically significant. As demonstrated in Figure 5, the 
funnel plot visualization illustrates the relationship between 
study precision and effect magnitude across the included 
investigations. At the same time, the asymmetrical pattern is 
particularly evident in the lower left quadrant, suggesting 
selective publication favoring studies with larger effect sizes. 
However, the robustness of primary findings remained intact 
after statistical adjustment for potential bias. 

 
Figure 5. Publication Bias Assessment in AI-Powered Dubbing 
Application Studies 

5. Discussion  

The systematic analysis of 38 studies examining AI-
powered dubbing applications reveals significant theoretical 
and practical implications for understanding personalized 
learning in language education, particularly within the 
framework of contemporary learning theories. The findings 
demonstrate substantial alignment with constructivist 
principles, where learners actively construct knowledge 
through interaction with authentic materials and receive 
immediate feedback to refine their understanding. This 
pedagogical framework extends beyond traditional 
boundaries when integrated with innovative technologies 
such as gamification strategies, which have shown promising 

effects on student engagement and learning outcomes in 
literature education contexts [33]. This technological 
mediation of learning processes reflects broader trends in 
ESL education, where digital tools fundamentally reshape 
teachers' perspectives on instructional design and 
implementation, necessitating continuous adaptation of 
pedagogical approaches to leverage technological affordances 
effectively while maintaining focus on meaningful learning 
outcomes [2]. The integration extends Vygotsky's Zone of 
Proximal Development by providing scaffolded support 
through AI-driven feedback mechanisms that adapt to 
individual proficiency levels, creating what might be 
conceptualized as a 'digital more knowledgeable other' that 
facilitates learning progression through systematic 
interaction patterns [34]. The observed effectiveness of 
personalized learning pathways (effect size d = 1.82 for 
pronunciation accuracy) corroborates recent theoretical 
frameworks proposing that AI-mediated learning 
environments can enhance traditional pedagogical 
approaches by providing consistent, individualized 
instruction that responds dynamically to learner needs. These 
findings contribute to an emerging theoretical understanding 
that positions AI not as a replacement for human instruction 
but as a complementary tool that addresses specific 
limitations in traditional language learning contexts, 
particularly the provision of consistent pronunciation 
feedback and opportunities for anxiety-free speaking practice 
[35]. 

Despite the promising outcomes documented across the 
reviewed studies, several methodological and practical 
limitations constrain the generalizability and applicability of 
findings. The predominance of short-term interventions (8-
12 weeks) raises questions about the sustainability of 
observed improvements, particularly given the documented 
decline in fluency retention rates (48.0% at 6 months) 
compared to pronunciation gains (75.7% retention). 
Geographic concentration of studies in East Asian contexts, 
where cultural attitudes toward technology adoption and 
language learning differ significantly from Western 
educational environments, potentially limits the 
transferability of findings to diverse global contexts [36]. The 
inherent difficulty in blinding participants to dubbing 
application interventions, reflected in high risk ratings for 
performance bias (47.4% of studies), introduces potential 
placebo effects that may inflate reported outcomes. 
Additionally, the focus on quantitative metrics of speaking 
proficiency may overlook qualitative aspects of 
communicative competence, such as pragmatic 
appropriateness and intercultural communication skills, 
which are essential for business English contexts but difficult 
to capture through automated assessment tools [23]. 

The identified research gaps and methodological 
limitations point toward several promising avenues for future 
investigation that could advance both theoretical 
understanding and practical application of AI-powered 
language learning tools. Longitudinal studies extending 
beyond one academic year are essential to understand the 
trajectory of skill maintenance and the optimal frequency of 
practice needed to sustain improvements in speaking 
proficiency [37]. The integration of multimodal data 
collection methods, combining speech analysis with eye-
tracking and neuroimaging techniques, could provide deeper 
insights into the cognitive processes underlying successful 
language acquisition through AI-mediated instruction. Cross-
cultural comparative studies examining how learners from 
different linguistic and cultural backgrounds interact with 



R. Zhao et al. /Future Technology                                     November 2025| Volume 04 | Issue 04 | Pages 228-239 

237 

 

and benefit from personalized learning algorithms would 
enhance the ecological validity of findings and inform 
culturally responsive design principles [38]. The emergence 
of large language models presents opportunities to 
investigate more sophisticated conversational AI that can 
engage learners in open-ended dialogues, moving beyond the 
current paradigm of scripted dubbing exercises toward truly 
adaptive conversational partners [39]. The practical 
implications of this systematic review extend across multiple 
dimensions of language education, from curriculum design to 
teacher professional development and institutional 
technology integration strategies, considerations that 
become particularly salient given documented variations in 
ESL teachers' knowledge and readiness to integrate Fourth 
Industrial Revolution technologies into their teaching 
practices across different educational contexts [40]. These 
implementation challenges necessitate comprehensive 
professional development frameworks that address not only 
technical competencies but also pedagogical understanding 
of how digital technologies can enhance specific language 
skills, as demonstrated in recent investigations examining the 
integration of digital tools in literature teaching within ESL 
classrooms, where teacher perspectives significantly 
influence successful implementation outcomes [41]. 
Educational institutions implementing AI dubbing 
applications benefit from adopting staged integration 
approaches that embed these technologies within existing 
curriculum frameworks through blended learning models 
where AI-powered practice sessions complement traditional 
classroom instruction during designated laboratory periods 
or homework assignments, while maintaining instructor-led 
components for cultural contextualization and pragmatic skill 
development.  

The observed cultural engagement patterns align with 
Hofstede's cultural dimensions theory, particularly power 
distance and uncertainty avoidance orientations, where East 
Asian learners' 34.7% higher completion rates, coupled with 
lower self-efficacy scores, reflect collectivist preferences for 
structured guidance and hierarchical learning environments 
that AI systems effectively provide through consistent 
feedback mechanisms. Cost-effectiveness analyses indicate 
that AI dubbing applications require initial investments of 
approximately $15-25 per student annually compared to 
$180-240 for equivalent individual tutoring sessions, though 
sustainable implementation demands addressing the 
documented learning plateau effects through adaptive 
content refresh cycles and gamification elements that 
maintain engagement beyond the critical 4-6 week threshold 
where improvement trajectories typically stabilize, 
suggesting that periodic algorithm updates and diversified 
practice scenarios represent essential strategies for 
sustaining long-term proficiency gains [42]. It is imperative 
that teacher education programs move beyond only 
developing technical competencies for using AI tools to 
establishing pedagogical understanding for how to integrate 
these technologies in ways that are meaningful within the 
extant curriculum, responding to the 50% of teachers who 
claim ineffective training is a key barrier to implementation 
[43]. The development of hybrid instructional models that 
leverage AI for targeted skill development while preserving 
human instruction for cultural contextualization and 
pragmatic competence represents a balanced approach that 
maximizes the strengths of both modalities. Hybrid formats 
combining AI for selective ability development and human 
instruction for cultural contextualization and pragmatic 
competence would be a balanced strategy to maximize the 

merits of both types of instruction. This hybrid design is 
motivated by the results of Baskara et al. [44] investigation on 
ChatGPT and Vera’s [45] studies on the integration of AI in 
EFL settings, as they all emphasize the need for human-in-the-
loop in AI-based learning contexts. Institutional policies are 
needed to define ethical use of AI, data privacy, and fairness 
of access to avoid deepening existing educational inequities, 
especially in light of lower adoption by high-poverty 
educational settings. 

6. Conclusion  

The present meta-analysis of 38 studies on AI-based 
dubbing application-supported personalized learning paths 
for speaking of BE learners provides evidence on their effects 
on speaking competence and yields both theoretical 
implications and practical implications for technology-
enhanced language education. Synthesis outcomes suggest 
that personalized algorithmic architectures, especially 
collaborative filtering and neural networks, lead to large 
changes in accurate pronunciation (d = 1.82) and fluent 
speaking (d = 1.46), and intermediate-level learners best 
respond to AI-mediated interventions. Conceptually, these 
findings have broader implications than the empirical 
verification of the effectiveness of DYAP in that they 
reimagine the status of technology in language learning as a 
dynamic, cognitive tool in the form of AI-driven applications 
that scaffold rather than replace the teacher, addressing the 
perennial problem of providing constant and individualised 
feedback in a contextually restricted educational 
environment. The identification of learner characteristics as 
key moderating variables, combined with evidence of 
differences in retention patterns across spoken sub-skills, 
moves the field forward in the understanding of how 
personalized learning paths can be personalized for different 
learner populations. These findings have the potential to 
transform Business English education worldwide, providing 
empirical strategies to institutions wishing to improve 
speaking ability through technological innovation without 
jeopardizing pedagogy, and considering the intricate 
relationships among linguistic acquisition, cultural context, 
and technological resources. 

Ethical issue 
The authors are aware of and comply with best practices in 
publication ethics, specifically with regard to authorship 
(avoidance of guest authorship), dual submission, 
manipulation of figures, competing interests, and compliance 
with policies on research ethics. The authors adhere to 
publication requirements that the submitted work is original 
and has not been published elsewhere. 

Data availability statement 
The manuscript contains all the data. However, more data will 

be available upon request from the authors. 

Conflict of interest 

The authors declare no potential conflict of interest. 

References 

[1]  Xu B, Ismail H. The impact of artificial intelligence-

assisted learning applications on oral English ability: 

A literature review[J]. International Journal of 

Academic Research in Progressive Education and 

Development, 2024, 13(4): 1118-1134. 

[2]  Sim J S E, Ismail H H. Using digital tools in teaching 

and learning English: Delving into English language 



R. Zhao et al. /Future Technology                                     November 2025| Volume 04 | Issue 04 | Pages 228-239 

238 

 

teachers’ perspectives[J]. Creative Education, 2023, 

14(10): 2021-2036. 

[3]  Walter Y. Embracing the future of Artificial 

Intelligence in the classroom: the relevance of AI 

literacy, prompt engineering, and critical thinking in 

modern education[J]. International Journal of 

Educational Technology in Higher Education, 2024, 

21(1): 15. 

[4]  Wei L. Artificial intelligence in language instruction: 

impact on English learning achievement, L2 

motivation, and self-regulated learning[J]. Frontiers 

in psychology, 2023, 14: 1261955. 

[5]  Huang X, Wu Y, Dou A. AI-enhanced task-based 

language teaching: Fostering personalized college 

English learning[J]. Frontiers in Educational 

Research, 2024, 7(10). 

[6]  Alhusaiyan E. A systematic review of current trends 

in artificial intelligence in foreign language 

learning[J]. Saudi Journal of Language Studies, 2025, 

5(1): 1-16. 

[7]  Huang H-T D. Investigating the effects of reverse 

dubbing tasks on English speaking proficiency and 

English learning motivation[J]. Language Teaching 

Research, 2024: 13621688241283153. 

[8]  Javier D R C, Moorhouse B L. Developing secondary 

school English language learners' productive and 

critical use of ChatGPT[J]. TESOL Journal, 2024, 

15(2): e755. 

[9]  Wang Y, Luo W, Liao X, et al. Exploring the effect of 

teacher autonomy support on Chinese EFL 

undergraduates’ academic English speaking 

performance through the mediation of basic 

psychological needs and classroom engagement[J]. 

Frontiers in Psychology, 2024, 15: 1323713. 

[10]  Bayly-Castaneda K, Ramirez-Montoya M S, Morita-

Alexander A. Crafting personalized learning paths 

with AI for lifelong learning: a systematic literature 

review[C]. Frontiers in Education, 2024: 1424386. 

[11]  Kohnke L, Moorhouse B L, Zou D. ChatGPT for 

language teaching and learning[J]. Relc Journal, 2023, 

54(2): 537-550. 

[12]  Peng H, Ma S, Spector J M. Personalized adaptive 

learning: an emerging pedagogical approach enabled 

by a smart learning environment[J]. Smart Learning 

Environments, 2019, 6(1): 1-14. 

[13]  Contrino M F, Reyes-Millán M, Vázquez-Villegas P, et 

al. Using an adaptive learning tool to improve student 

performance and satisfaction in online and face-to-

face education for a more personalized approach[J]. 

Smart Learning Environments, 2024, 11(1): 6. 

[14]  Crompton H, Edmett A, Ichaporia N, et al. AI and 

English language teaching: Affordances and 

challenges[J]. British Journal of Educational 

Technology, 2024, 55(6): 2503-2529. 

[15]  Mejeh M, Rehm M. Taking adaptive learning in 

educational settings to the next level: Leveraging 

natural language processing for improved 

personalization[J]. Educational technology research 

and development, 2024, 72(3): 1597-1621. 

[16]  Rajan S T, Ismail H H. TikTok use as strategy to 

improve knowledge acquisition and build 

engagement to learn literature in ESL classrooms[J]. 

International Journal of Learning, Teaching and 

Educational Research, 2022, 21(11): 33-53. 

[17]  Du J, Daniel B K. Transforming language education: A 

systematic review of AI-powered chatbots for English 

as a foreign language speaking practice[J]. Computers 

and Education: Artificial Intelligence, 2024, 6: 

100230. 

[18]  Li H. A Comparative Study of Artificial Intelligence-

Enabled Business Foreign Language Teaching in 

Universities: Taking ChatGPT-4o, Claude3. 5 Sonnet 

and ERNIE 4.0 Turbo as Examples[C]. Proceedings of 

the 2024 2nd International Conference on 

Information Education and Artificial Intelligence, 

2024: 10-15. 

[19]  Mohebbi A. Enabling learner independence and self-

regulation in language education using AI tools: a 

systematic review[J]. Cogent Education, 2025, 12(1): 

2433814. 

[20]  Li Y, Zhou X, Yin H-B, et al. Design language learning 

with artificial intelligence (AI) chatbots based on 

activity theory from a systematic review[J]. Smart 

Learning Environments, 2025, 12(1): 24. 

[21]  Wang Y, Wu J, Chen F, et al. Empirical assessment of 

AI-powered tools for vocabulary acquisition in EFL 

instruction[J]. IEEE Access, 2024. 

[22]  Ellikkal A, Rajamohan S. AI-enabled personalized 

learning: empowering management students for 

improving engagement and academic performance[J]. 

Vilakshan-XIMB Journal of Management, 2024, 22(1): 

28-44. 

[23]  Wang X, Lee S-M. The impact of video dubbing app on 

chinese college students’ oral language skills across 

different proficiency levels[J]. International Journal of 

Educational Research, 2025, 130: 102521. 

[24]  Li T, Ji Y, Zhan Z. Expert or machine? Comparing the 

effect of pairing student teacher with in-service 

teacher and ChatGPT on their critical thinking, 

learning performance, and cognitive load in an 

integrated-STEM course[J]. Asia Pacific Journal of 

Education, 2024, 44(1): 45-60. 

[25]  Leong L V, Yunus M M, Ismail H H. Integration of 

techno-pedagogical approach in English as a second 

language classroom: a systematic review[J]. Int J Eval 

& Res Educ ISSN, 2024, 2252(8822): 4395. 

[26]  Sun L, Ismail H H, Aziz A A. Current English Language 

Teaching Using Production-Oriented Approach: A 

Systematic Review[J]. World Journal of English 

Language, 2024, 14(4). 

[27]  Senthil R, Anand T, Somala C S, et al. Bibliometric 

analysis of artificial intelligence in healthcare 

research: Trends and future directions[J]. Future 

Healthcare Journal, 2024, 11(3): 100182. 

[28]  Zhao J. Advancing english language education: A 

mixed-methods analysis of AI-driven tools' impact on 

engagement, personalization, and learning 

outcomes[J]. Education and Information 

Technologies, 2025: 1-41. 

[29]  Schünemann H J, Mustafa R A, Brozek J, et al. GRADE 

guidelines: 21 part 1. Study design, risk of bias, and 

indirectness in rating the certainty across a body of 



R. Zhao et al. /Future Technology                                     November 2025| Volume 04 | Issue 04 | Pages 228-239 

239 

 

evidence for test accuracy[J]. Journal of clinical 

epidemiology, 2020, 122: 129-141. 

[30]  Chandler J, Cumpston M, Li T, et al. Cochrane 

handbook for systematic reviews of interventions[J]. 

Hoboken: Wiley, 2019, 4(1002): 14651858. 

[31]  Yoong W F, Hashim H. A Systematic Review on 

Language Learning Strategies for English Speaking 

Skills in Today’s Learning Conditions[J]. International 

Journal of Academic Research in Business & Social 

Sciences. https://doi. org/10.6007/ijarbss/v13-

i12/19727, 2023. 

[32]  Thomas J, Harden A. Methods for the thematic 

synthesis of qualitative research in systematic 

reviews[J]. BMC medical research methodology, 2008, 

8(1): 45. 

[33]  Chandrasehgaran M, Ismail H. The Impacts of 

Gamification on Student Engagement and Learning 

Outcomes in Literature in Education: A Literature 

Review[J]. International Journal of Academic 

Research in Business and Social Sciences, 2024, 

14(8): 737-745. 

[34]  Talaván N, Costal T. iDub-The potential of intralingual 

dubbing in foreign language learning: How to assess 

the task[J]. Language Value, 2017, 9: 62-88. 

[35]  Zhu M, Wang C. A systematic review of artificial 

intelligence in language education from 2013 to 

2023: Current status and future implications[J]. 

Available at SSRN 4684304, 2024. 

[36]  Huang Y, Wan L J, Ye H, et al. New solutions on LLM 

acceleration, optimization, and application[C]. 

Proceedings of the 61st ACM/IEEE Design 

Automation Conference, 2024: 1-4. 

[37]  Hwang G J, Fathi J, Rahimi M. Fostering EFL Learners' 

Speaking Skills and Flow Experience With Video‐

Dubbing Tasks: A Flow Theory Perspective[J]. Journal 

of Computer Assisted Learning, 2025, 41(2): e13120. 

[38]  Zhang X, Sun J, Deng Y. Design and application of 

intelligent classroom for English language and 

literature based on artificial intelligence 

technology[J]. Applied Artificial Intelligence, 2023, 

37(1): 2216051. 

[39]  Pu P, Chang D Y-S. Effects of different input modes on 

blended EFL speaking instruction: a quasi-

experimental study[J]. Computer Assisted Language 

Learning, 2023: 1-26. 

[40]  Avelino N M, Ismail H H. Assessing ESL teachers’ 

knowledge and readiness in integrating 4IR into 

teaching practices: A concept paper[J]. Creative 

education, 2021, 12(9): 2038-2055. 

[41]  Shamshul I S M, Ismail H H, Nordin N M. Using digital 

technologies in teaching and learning of literature in 

ESL classrooms: A systematic literature review[J]. 

International Journal of Learning, Teaching and 

Educational Research, 2024, 23(4): 180-194. 

[42]  Sabili A S, Nazri M A, Surayya S, et al. Smart learning: 

the impact of ai technologies on english language 

proficiency in the digital era[J]. J Rev Pendidik dan 

Pengajaran, 2024, 7(4): 14040-14049. 

[43]  Kaufman J H, Woo A, Eagan J, et al. Uneven adoption 

of artificial intelligence tools among US teachers and 

principals in the 2023-2024 school year[M]. RAND, 

2025. 

[44]  Baskara R. Exploring the implications of ChatGPT for 

language learning in higher education[J]. Indonesian 

Journal of English Language Teaching and Applied 

Linguistics, 2023, 7(2): 343-358. 

[45] Vera F. Integrating Artificial Intelligence (AI) in the 

EFL classroom: Benefits and challenges[J]. 

Transformar, 2023, 4(2): 66-77. 

 

  

This article is an open-access article distributed under the 

terms and conditions of the Creative Commons Attribution 

(CC BY) license 

(https://creativecommons.org/licenses/by/4.0/). 

https://doi/
https://creativecommons.org/licenses/by/4.0/

