The impact of virtual reality environments on English language acquisition: Innovative immersive learning technologies for communication skills development

Abstract

In the rapidly evolving digital landscape, virtual reality (VR) has emerged as a transformative tool in education, particularly in language acquisition. Traditional language learning methods often struggle to provide immersive, interactive, and engaging experiences that foster real-world communication skills. VR environments are increasingly transforming English language acquisition by providing immersive and interactive learning experiences that enhance communication skills development. This paper explores the impact of immersive VR technologies on English language acquisition, focusing on communication skills development through dynamic and interactive learning experiences. A novel Dynamic Monarch Butterfly Optimized Bidirectional Gated Recurrent Unit (DMBO-BiGRU) model is presented to analyze learner interactions within VR environments, assessing communication patterns, language complexity, and learning progression. The dataset, collected from student interactions in VR settings, was preprocessed using noise filtering and normalization to ensure data quality. t-Distributed Stochastic Neighbor Embedding (t-SNE) is employed to extract key features, such as fluency, vocabulary usage, and contextual appropriateness. The DMBO-BiGRU model leverages adaptive learning strategies to provide tailored feedback and optimizes language acquisition based on learners’ communication skills. The effectiveness of the suggested method is demonstrated through experimental results, which indicate good accuracy and positive VR learning scores. The results underscore how immersive, customized, and interactive experiences provided by VR-integrated deep learning models hold the potential to transform language acquisition, showing accuracy of 97.2%, VR score of 97.52%, training gain of 98.49%, detection score of 92.6%, and a retention rate of 96% at iteration 13. This work contributes to the growing field of technology-enhanced education, emphasizing the role of VR and AI-driven methods in enhancing English communication skills.

Keywords

virtual reality (VR)English language communication skills immersive learning dynamic monarch butterfly optimized bidirectional gated recurrent unit (DMBO-BiGRU)

Introduction

Technological advancements have initiated a significant shift in language learning, with virtual reality (VR) emerging as a revolutionary educational tool. As the global demand for English language competence increases, educators and scholars are exploring new methods to enhance language learning.¹ VR presents an interactive and immersive learning atmosphere, providing students with a unique opportunity to engage with the language in a tangible and meaningful manner, as illustrated in Figure 1. The core component of VR is the role of technology in its environment.² VR refers to a computer-generated simulation that provides a synthetic, three-dimensional environment where users can engage with a digital landscape that appears realistic and immersive. Sensory components such as visual and, in some cases, haptic input enhance the user’s sense of presence.³ By offering interactive scenarios, VR allows for deeper cognitive and experiential learning compared to traditional digital learning approaches that rely on textbooks or screen-based materials. VR enables language learners to practice their skills through stable encounters and meaningful dialogues that replicate real-world experiences.⁴

The context in which communication is provided by the environment significantly influences language development. Traditional classroom-based language instruction often lacks real-world engagement, which can hinder the development of communication capabilities.⁵ However, through VR environments, language learners can apply their skills in everyday situations, such as online marketplaces, workplaces, social activities, or travel scenarios. These immersive environments encourage the use of natural language by giving students the opportunity to utilize their knowledge in real-world contexts.⁶ The adaptive nature of VR settings accommodates a variety of learning styles and levels. English language acquisition refers to the acquisition of skills in speaking, listening, reading, and writing.⁷ Traditional methods often focus on grammar practice and rote memorization, which may not foster natural language use. In contrast, VR-based learning promotes engagement and active participation.⁸ VR facilitates improvements in comprehension, fluency, and pronunciation through simulations of authentic conversations. Additionally, VR encourages self-observation and mitigates the language anxiety that can arise in traditional learning environments.⁹

These components collectively establish VR as a valuable tool for enhancing English language learning. VR fosters learning experiences that transcend academic knowledge by immersing students in an interactive, realistic environment.¹⁰ Such direct interactions allow language learners to develop their skills in conditions that reflect real-world situations. Furthermore, the personalized and adaptable nature of VR platforms ensures that learners receive tailored experiences according to their skill levels and learning paces.¹¹ As VR technology continues to advance and become more accessible, it can revolutionize English language instruction. This paper examines the advantages, challenges, and potential applications of VR environments in the context of learning the English language.¹² However, varying personal learning styles and technological accessibility may limit generalizations. Additionally, issues such as motion sickness and cognitive stress could negatively affect the user experience. A unique DMBO-BiGRU model is introduced to evaluate communication patterns, language complexity, and learning development through learner interactions within VR environments.

Related works

Recent advancements in immersive technologies and artificial intelligence (AI) have significantly influenced language education research. This section synthesizes key studies exploring the integration of virtual reality (VR), augmented reality (AR), and AI in language learning contexts. Shi et al.¹³ developed a VR-based English-speaking training and assessment system to examine the impact of cognitive style (field-dependent vs field-independent) and test settings (virtual vs real environments) on learning outcomes. Their findings revealed that field-independent learners exhibited superior performance in virtual testing environments, whereas field-dependent learners demonstrated better outcomes in traditional settings. The study highlighted the critical interplay between cognitive styles, environmental contexts, and technological interventions in shaping language acquisition efficacy. In a comparative analysis of vocabulary acquisition, Tai et al.¹⁴ investigated the effectiveness of mobile-rendered head-mounted displays (HMDs) for seventh-grade EFL learners. The experimental group using VR applications outperformed the control group (video-based learning) in both immediate retention and long-term recall of vocabulary. Despite the VR group’s success, the video group reported mixed perceptions of multimodal support, underscoring the need for further exploration of HMDs’ pedagogical implications.

Lu¹⁵ proposed a machine learning (ML) approach, termed the Re-defined Harris Hawks Optimized Intelligent Support Vector Machine (RHH-ISVM), to enhance English language learning in VR environments. By analyzing learners’ interaction patterns, language proficiency metrics, and communication behaviors, the model demonstrated high accuracy in skill classification and progress evaluation. This work emphasized the synergistic potential of immersive VR and adaptive ML algorithms for personalized language instruction. Chen et al.¹⁶ integrated AI and VR technologies to create a robot-assisted language learning (RALL) system for training English-speaking tour guides. During a 10-week intervention, learners engaged with a 3D VR environment and AI-driven robot interactions, resulting in measurable improvements in speaking fluency and learner autonomy. Participants acknowledged the system’s capacity to foster active participation and flexible learning pathways, validating RALL’s role in complementing traditional pedagogies.

Divekar et al.¹⁷ introduced the Cognitive Immersive Language Learning Environment (CILLE), combining Extended Reality (XR) and AI to facilitate non-dyadic multimodal interactions for Chinese as foreign language (CFL) learners. A 7-week mixed-methods study with university students revealed significant gains in vocabulary, comprehension, and conversational competence, demonstrating XR’s ability to deliver immersive experiences without reliance on intrusive hardware. Qiao and Zhao¹⁸ explored AI-driven interventions for enhancing self-regulation and L2 speaking skills among Chinese EFL learners. Their results indicated that AI-based training not only improved oral proficiency but also strengthened metacognitive strategies and learner autonomy, suggesting AI’s transformative potential in fostering independent language acquisition.

Lee¹⁹ compared the efficacy of printed materials and mobile AR games for narrative-driven language learning with South Korean undergraduates. While both media elicited positive engagement, print-based resources were perceived as more conducive to structured language practice. The study advocated for location-based AR applications to enrich contextualized learning experiences. Smutny²⁰ conducted a market analysis of VR educational applications (2019–2021), identifying history, art, medicine, and natural sciences as dominant domains. Notably, over 50% of sampled applications were freely accessible and English-centric, highlighting both opportunities and linguistic biases in VR’s educational adoption.

Ironsi²¹ investigated VR’s role in developing speaking skills through teacher and student perspectives. While VR enhanced oral communication outcomes, the study cautioned against technological overreliance and emphasized balancing immersive tools with pedagogical fundamentals. Beyond language-specific applications, Li et al.²² designed a distributed VR framework for entrepreneurship education, employing 3D point cloud processing and edge contour recognition to optimize collaborative virtual spaces. Their system reduced time overhead and information saturation, aligning with evolving demands for scalable, immersive learning infrastructures. Hu et al.²³ proposed a multimodal feature integration model to assess learner concentration in VR environments. By analyzing visual and interaction data, their method achieved superior accuracy in concentration recognition, correlating high immersion levels with improved learning outcomes.

Collectively, these studies underscore VR, AR, and AI’s transformative potential in language education while highlighting contextual challenges such as cognitive diversity, technological accessibility, and pedagogical integration.

Methodology

The virtual reality environments were specifically designed to mimic real-world scenarios that language learners are likely to encounter, such as social gatherings, professional meetings, and everyday interactions in marketplaces. The environments were selected based on their ability to stimulate authentic language use and encourage engagement. The classification of communication patterns was based on several established linguistic features, including but not limited to fluency rates, vocabulary richness, and contextual appropriateness. We utilized a framework to categorize these patterns into low, medium, and high communication levels, which enabled us to assess the effectiveness of the VR interactions more accurately. Noise filtering and normalization are two steps in data preparation that improve data quality for precise learner assessment. t-SNE is used to extract important language properties, such as word richness, contextual appropriateness, and fluency. In immersive VR environments, the suggested DMBO-BiGRU model analyzes learner communication, offers feedback that is adjusted, and improves individualized language acquisition by utilizing adaptive learning techniques. Figure 2 shows the overall flow for VR in English language acquisition.

Dataset

The collection of learner interactions, linguistic characteristics, engagement metrics, and AI-generated feedback, the VR-LearnEng dataset facilitates English language learning in VR environments and analyzes the evolution of communication abilities. A learning success categorization (low, medium, high) is used to monitor progress, and it also contains learner demographics, VR session facts, language complexity, interaction metrics, and individualized feedback. A great resource for assessing and improving immersive language learning experiences, which consists of 1000 structured samples in CSV format, offers insightful information on vocabulary variety, fluency, contextual comprehension, and learning development.

Source: https://www.kaggle.com/datasets/ziya07/vr-learneng-dataset/data.

Data preprocessing

Data preprocessing is necessary for accuracy and reliability when measuring the impact of VR environments on English language learning. Z-score normalization is conducted on linguistic performance measures to normalize data and make comparisons more meaningful by bringing the data in view around zero with unit variance. The Wiener filter also enhances the measurement of response time and pronunciation quality by reducing noise in speech recordings and sensor inputs. By enhancing the robustness of the dataset, this preprocessing facilitates easier analysis of the outcomes of VR-based language learning with increased precision.

Z-score normalization

Z-score normalization is a statistical technique that standardizes the linguistic performance measures by transforming them into a function of the mean and standard deviation of the dataset. This method ensures that the data possesses a mean of zero and a standard deviation of one, which allows for meaningful comparison between various linguistic indicators. The formula for this transformation is given by:

W_{n e w} = \frac{W - μ}{σ}

(1)

where

W_{n e w}

is the new value,

W

is the original value,

μ

is the mean of the dataset, and

σ

is the standard deviation. This normalization is especially beneficial when comparing variables measured on different scales, as it mitigates the impact of outliers and ensures each indicator contributes equally to the analysis. Z-score normalization was chosen to standardize linguistic performance metrics due to its capacity to handle datasets with unknown value ranges, which is critical given the heterogeneous nature of VR interaction data. This technique ensures comparability between learners’ longitudinal performance records by converting raw scores into standard deviations from the cohort mean.

Weiner filter

The Wiener filter is employed to enhance the quality of the data collected by reducing background noise in the recorded interactions. This filter functions by estimating the desired signal from the noisy input based on statistical estimations of the signal and noise present. The Wiener filter optimally minimizes the mean square error between the estimated and the actual signal, thereby improving the clarity of spoken interactions in the VR environment. The filter model is represented in the frequency domain by the expression:

{| H_{W l e n e r} (x) |}^{2} = \frac{1}{{| G (x) |}^{2}} {(1 + \frac{1}{{| G (x) |}^{2} S N R (x)})}^{- 2}

(2)

where

S N R (x)

is the signal-to-noise ratio as a function of frequency,

G (x)

is the system transfer function, and

H_{W l e n e r} (x)

is the Wiener filter function in the frequency domain.

It should be noted that the Wiener filter in equation (2) reduces to the inverse filter in equation (3), provided that detector noise is absent $(S N R (ω) \to \infty)$ . The Wiener filter performs badly when it comes to reconstructing the associated power spectral density (PSD), despite offering an ideal reconstruction of the time-domain signal.

T_{\hat{w} \hat{w}}^{(w l e n e r)} (x) = {| H_{W l e n e r} (x) |}^{2} T_{y, y} (x) = {(1 + \frac{1}{{| G (x) |}^{2} S N R (x)})}^{- 2} T_{w, w,} (x) < T_{w, w} (x)

(3)

Here, $T_{w, w} (x)$ is the true concealed signal’s power spectral density (PSD), $T_{\hat{w} \hat{w}}^{(w l e n e r)} (x)$ is the Wiener filter estimate of the concealed signal’s PSD, and $T_{y, y} (x)$ is the observed signal’s PSD. As a result, the concealed signal is consistently underestimated by the Wiener filter.

The Wiener filter was implemented using a 512-sample Hamming window with 50% overlap to mitigate ambient noise from VR headset microphones, particularly addressing the 60 Hz electrical hum prevalent in laboratory environments. The filter’s frequency response was calibrated using noise profiles captured during system idle states, ensuring adaptive suppression of device-specific artifacts.

Feature extraction using t-SNE

A useful dimensionality reduction method for displaying high-dimensional data is t-SNE. t-SNE can extract useful characteristics from learner interactions, speech patterns, and engagement metrics while exploring how VR environments affect English language learning. t-SNE assists in identifying student groups according to competency, immersion levels, or engagement patterns by translating intricate linguistic and behavioral data into a lower-dimensional environment. Groups of neighboring data points are called neighbors, and the t-SNE approach models their distribution. The initial high-dimensional space is treated as a Gaussian distribution, while the low-dimensional output space is represented as a t-distribution. This four-step process intends to find the transformation that helps distribute data points more uniformly in low-dimensional conditions by reducing the gaps between all points and transferring high-dimensional space to low-dimensional space.

The highly dimensional Euclidean distance between each point of data in equation (4) is converted into a conditional probability of resemblance or the conditional probability $o_{i | j}$ between data points $w_{j}$ and $w_{i} .$

o_{i | j} = \frac{\exp (- \frac{{‖ w_{j} - w_{i} ‖}^{2}}{2 σ_{j}^{2}})}{\sum_{l \neq j} \exp (- \frac{{‖ w_{j} - w_{i} ‖}^{2}}{2 σ_{j}^{2}})}

(4)

where

σ_{j}

is the Gaussian variance centered on data point

w_{j}

, controlling the distribution’s spread,

o_{i | j}

is the conditional probability that point

w_{i}

is a neighbor of point

w_{j}

w_{j}

, and

w_{i}

are the high-dimensional data points, and

‖ w_{j} - w_{i} ‖

is the Euclidean distance between points

w_{j}

and

w_{i}

The formula for calculating $σ_{j}$ is provided. Set $o_{i | j}$ to $0$ as the user is interested in modeling pairwise similarities. For the low-dimensional counterparts, $z_{j} a n d z_{i}$ , of the highly dimensional data points, $w_{j} a n d w_{i}$ , a similar conditional probability, denoted as $r_{i | j}$ , could be computed in equation (5), set $r_{i | j}$ to 0.

r_{i | j} = \frac{\exp (- {‖ z_{j} - z_{i} ‖}^{2})}{\sum_{l \neq j} \exp (- {‖ z_{j} - z_{i} ‖}^{2})}

(5)

Here, $‖ z_{j} - z_{i} ‖$ are the low-dimensional counterparts of $w_{j}$ and $w_{i}$ and $r_{i | j}$ is the conditional probability that point $z_{i}$ is a neighbor of point $z_{j}$ .

Using the gradient descent approach, SNE finds the smallest sum of the conditional probability differences by minimizing the $L K$ distance. To minimize congestion and maximize SNE, t-SNE uses a heavy-tailed distribution, whereas SNE’s cost function focuses on the local architecture of the map’s data. However, optimizing the function is extremely challenging. The SNE cost function is displayed below in equation (6).

D = \sum_{j} L K (O_{j} ‖ R_{j}) = \sum_{j} \sum_{i} o_{i | j} \log \frac{o_{i | j}}{r_{i | j}}

(6)

where

L K

is the Kullback-Leibler divergence and

D

is the cost function that measures the difference between distributions

O_{j}

and

R_{j}

Finally, the Gaussian variance $o_{j}$ , centered on each high-dimensional data point $w_{j}$ , is selected. This distribution’s entropy increases as $o_{j}$ increases. SNE determines the value of $o_{j}$ using a binary search, producing an $o_{j}$ with a fixed confusion, as shown in equation (7).

p e r p (σ_{j}) = 2^{G (σ_{j})}

(7)

when

G (σ_{j})

is the Shannon entropy of

σ_{j}

, expressed in bits and

p e r p (σ_{j})

is the perplexity value, a measure of effective nearest neighbors, and the Shannon entropy of

o_{j}

expressed in bits is denoted by

G (o_{j})

in equation (8).

G (o_{j}) = \sum_{i} o_{i | j} \log_{2} o_{i | j}

(8)

DMBO-BiGRU

A deep learning method called the DMBO-BiGRU model was designed to check and evaluate learning interactions in VR settings. This model optimizes hyperparameters of a BiGRU network using the MBO method, a metaheuristic technique inspired by nature. BiGRU learner captures the past and future dependence in the conversation, making it an excellent option for sequential data processing. The model improves training efficiency and convergence speed using MBO, resulting in learning progression, communication patterns, and accuracy when evaluating the complexity of the English language. DMBO-BiGRU VR-based learning environment handles multimodal data, including speech, text, gestures, and engagement levels. By identifying speech flow, consistency, and interaction frequency, it analyzes communication structures and provides information about the participation of the learner. In addition, it uses linguistic characteristics including sentence structure, vocabulary prosperity, and syntax pattern to assess the complexity of the English language. By tracking the performance pattern over time and by offering analog reactions, the model monitors further learning progress. BiGRU’s ability to respond to various types of learning behaviors is better than MBO incorporation, which guarantees optimal parameter selection. In its final stages, DMBO-BiGRU makes it easy to assess the language of learners in immersive virtual reality settings in a clever, flexible, and data-operated manner. Algorithm 1 shows the working procedure of the proposed DMBO-BiGRU model.

BiGRU

BiGRU is used to investigate how VR environments affect learning English. Through the utilization of VR immersive learning experiences, the BiGRU model analyzes advances in vocabulary retention, pronunciation, and understanding by processing sequential language learning data. Recurrent neural networks, or RNNs, use sequence data as input and process it repeatedly along the sequence’s development path. Each node is connected to the others by a chain. Bidirectional and long short-term memory networks (LSTM) are two common varieties of recurrent neural networks. Nevertheless, it has issues with gradient expansion and disappearance. The issues of disappearing gradients and explosion of gradients can be resolved using LSTM, an enhanced RNN-based variant. It performs well in the domains of machine translation, language modeling, and speech recognition. In addition, it is applied to a variety of time series attribute issues. Understanding bidirectional GRU requires knowledge of gated recurrent units (GRUs). A simpler variant of the LSTM network is the GRU. The input, forget, and output gates constitute an LSTM. The update gate and the reset gate are the only two gates in GRU. The input and forget gates of an LSTM function are similar to the update gate. It decides which data should be maintained as current and which should be left behind. When determining the current time, the reset gate decides if a certain amount of the previous data is unnecessary. Due to having fewer gating points, GRU executes computations faster than LSTM. The following is the formula for calculating the GRU equations (9)–(14).

y_{s} = σ (X_{y} . [g_{s - 1}, w_{s}])

(9)

where

g_{s - 1}

denotes the hidden state from the previous time step,

y_{s}

represents the update gate value at time step s,

X_{y}

is the weight matrix for the update gate, and

w_{s}

indicates the current input data at time step s.

q_{s} = σ (X_{q} . [g_{s - 1}, w_{s}])

(10)

where

q_{s}

represents the reset gate value at time step s, and

X_{q}

denotes the weight matrix for the reset gate.

g_{s} = (1 - y_{s}) * g_{s - 1} + y_{s} * {\tilde{g}}_{s}

(11)

where

{\tilde{g}}_{s}

is the candidate hidden state based on the current input, and

g_{s}

represents the updated hidden state at time step s.

{\tilde{g}}_{s} = \tanh (X \cdot [g_{s}^{*} g_{s - 1} + y_{s} * {\tilde{g}}_{s}])

(12)

where

X

represents the weight matrix for the input data.

T (w) = \frac{1}{1 + f^{- w}}

(13)

T^{'} (w) = \frac{f^{- w}}{{(1 + f^{- w})}^{2}} = T (w) (1 - T (w))

(14)

Here, $f$ is the scaling factor, $T^{'} (w)$ is the derivative of the sigmoid function, and $T (w)$ is the sigmoid function output.

The sigmoid activation function, denoted as $σ$ , compresses all output data between $0 a n d 1$ . Equations (13) and (14) show the sigmoid function and its derivation formula. Update gate $y_{s}$ and reset gate $q_{s}$ have weights denoted by $X_{y}$ and $X_{q}$ , respectively. The prior information is represented by $g_{s - 1}$ , which also includes the current information ${\tilde{g}}_{s}$ . Past information $g_{s - 1}$ and current input generate the current information. Bidirectional GRU and GRU have a similar model structure. Two types of time series exist: positive and reverse. The final output is a combination of the results for the reverse time sequence and the end state of the positive time sequence. The model is able to simultaneously use knowledge from the past and the future. Subnetworks for forward status and backward status which represent forward transmission and backward transmission, respectively, are located within the network shown in Figure 3.

DMBO

In VR environments, English language acquisition is enhanced through immersive learning experiences. When it comes to language learning, VR exhibits traits that set it apart from traditional methods. Learners engage with dynamic, interactive scenarios that mimic real-world communication. The structured yet flexible learning process in VR has an impact on the development of effective language acquisition strategies. Exploration and exploitation are two crucial techniques that guide the learning process, much like in other adaptive learning systems. VR-based learning utilizes these two techniques: exploring new linguistic contexts and adjusting learned skills. Equations (15) and (16) are used to initialize the two phases of learning, referred to as $M_{s p 1} and M_{s p 2}$ , with regard to language exploration and refinement.

M_{s p 1} = c e i l (o . M_{p})

(15)

M_{s p 2} = M_{p} - M_{s p 1}

(16)

There are several solutions for the entire population, represented by $M_{p} and p$ represents the ratio of $M_{s p 1}$ and $M_{p}$ . $p$ is important in the MBO search procedure. Consequently, two subpopulations are impacted by $p$ adjustment. The resultant solutions are from $M_{s p 1}$ if $p$ value is large, and from $M_{s p 2}$ otherwise. In the original MBO, the value of $p$ was set at $5 / 12$ . One benefit of the solution migration concept is that it produces innovative solutions. It is accomplished by using equations (17) and (18).

w_{j, i}^{s + 1} = w_{q_{1}, i}^{s}

(17)

w_{j, i}^{s + 1} = w_{q_{2}, i}^{s}

(18)

The symbols $w_{q 1} and w_{q 2}$ represent the new solutions at iteration $s + 1$ . The $i t h$ parameter of the new solution is represented by $w_{j, i}^{s + 1}$ , whereas the $i t h$ components of the two solutions are indicated by $w_{q_{1}, i}^{s}$ and $w_{q_{1}, i}^{s}$ . Using equation (19), the value of $q$ is calculated to identify whether the provided new solution is related with $M_{s p 1} or M_{s p 2}$ .

q = r a n d \cdot p e r i

(19)

The rand is a number extracted from a distribution, and the variable peri determines the period of migration to regulate MBO functioning. The peri value in the original MBO was $1.2$ . Equation (17) is used to construct the $i t h$ component if $q \leq p$ ; if not, equation (18) is utilized. Adjustment for solutions is intended for both exploration and exploitation. Here, the direction of the search process is determined by a pseudo-random number, represented by $r n d$ , and there are two different scenarios. In the first scenario, the existing solution is modified to provide a new solution as shown in equation (20) if $r n d \leq p$ .

w_{j, i}^{s + 1} = w_{b e s t, i}^{s}

(20)

The $i t h$ parameter of the best solution generally in use for the whole population is $w_{j, i}^{s + 1}$ , whereas the $i t h$ parameter of the innovative solution is $w_{b e s t, i}^{s}$ . Equation (21) yields a new solution if $r n d > p$ .

w_{j, i}^{s + 1} = w_{q_{3}, i}^{s}

(21)

The randomly selected solution, represented by the symbol $q_{3}$ , is derived from $M_{s p}$ , and its $i t h$ parameter is represented by the symbols $w_{q_{3}, i}^{s}$ . New solutions have been added as shown in equation (22) after establishing $r n d > a r$ .

w_{j, i}^{s + 1} = w_{j, i}^{s + 1} + α \times (c_{w_{i}} - 0.5)

(22)

The adjustment rate is represented by $a r$ , and a step is indicated by $d w$ , taking consideration of the Lévy flights formula found in equation (23).

d w = L e v y (w_{i}^{s})

(23)

The weighting factor $α$ is calculated using equation (24).

α = \frac{T_{\max}}{s^{2}}

(24)

The maximum value for the step is represented by $T_{\max}$ , where $s$ is the current iteration and a given solution proceeds one step at a time. The trade-off between exploration and exploitation is exploited by the parameter $α$ . The length of the search step varies according to its value. The original MBO technique has a lot of problems. Due to its reliance on the weighting element, it first lacks adequate power throughout the exploration process. It is also limited in its ability to strike a balance between exploration and exploitation.

The number of DMBO iterations

([20, 50, 100])

determines how well the model optimizes learning parameters in VR-driven English language acquisition. Although more iterations improve accuracy and customization, it can additionally raise computing expenses. Effective and flexible language learning experiences are guaranteed by a well-rounded strategy. Table 1 shows the hyperparameter for DMBO-BiGRU.

Table 1.

Hyperparameter for DMBO-BiGRU.

Hyperparameter	Search range / Values
Hidden units	$[64, 128, 160, 192, 256]$
Number of BiGRU layers	$[1, 2, 3]$
Learning rate	$[0.0001, 0.001, 0.01]$
Dropout rate	$[0.1, 0.2, 0.3, 0.4, 0.5]$
Batch size	$[32, 64, 128]$
Epochs	$[30, 50, 100]$
Optimizer	[Adam, RMSprop, SGD]
Loss function	[Crossentropy, MSE]
Migration factor (DMBO)	$[0.2, 0.4, 0.6]$
Population size (DMBO)	$[20, 30, 50]$
Number of iterations (DMBO)	$[20, 50, 100]$

Experimental findings

The proposed DMBO-BiGRU model is utilized to assess the effects of VR environments on English language learning and is compared against existing techniques such as Artificial Neural Networks (ANN)²⁴ and Fuzzy-assisted Double Deep Q-Learning Networks (F-D2QLN). 25 Performance metrics include accuracy, VR score, training gain, detection score, and retention rate. Table 2 outlines the experimental setup.

Table 2.

Experimental setup for the suggested DMBO-BiGRU method.

Characteristics	Specifications
Programming language	Python 3.11
Operating system	Windows 11
Hardware	Intel i7 Core CPU, 16 GB RAM
Libraries	TensorFlow/Keras, Scikit-Learn

Three classes, designated as 0, 1, and 2, constitute the confusion matrix in Figure 4, which illustrates how well a model predicts various categories. Class 1 is incorrectly categorized as class 2 three times; however, classes 0 and 2 have excellent accuracy (80 and 117 correct predictions, respectively), indicating that the model successfully separates these groups. Such a matrix can display how successfully a machine learning model categorizes language competence levels or learning outcomes in the context of exploring how VR environments affect English language acquisition. Although there could be places for improvement at intermediate competence levels (class 1), the low misclassification indicates that VR-based language instruction could enhance specific language abilities.

The distribution of language skill levels among students is depicted in Figure 5, where the majority (611 learners) are in the intermediate group, followed by the advanced (369 learners), and the beginners (20 learners) category is the smallest. This distribution implies that learners at intermediate and advanced levels could discover VR-based learning more popular or more successful when it comes to VR environments and English language acquisition. The low percentage of beginners could indicate that beginning language learners who need more fundamental assistance can benefit from immersive VR experiences for language learning, or it could have restricted access to VR resources.

A model’s ability to differentiate between various degrees of language skill in a VR-assisted English language learning environment is demonstrated by the Precision-Recall (PR) curve in Figure 6. The model appears to correctly identify intermediate and advanced learners with minimal misclassification, as indicated by the high AUC values (1.00 for these learners). However, the beginner class’s lower AUC (0.85) suggests that it could be challenging to differentiate starting learners, either as a result of overlapping linguistic traits or fewer data points. This implies that even while VR environments help intermediate and advanced learners very well, additional effort could be required to improve language learning in the early stages.

The average VR learning scores for each level of English competence are displayed in Figure 7, with no difference between beginners (85.0), intermediate learners (84.5), and advanced learners (84.7). It indicates that regardless of skill level, virtual reality (VR) environments offer a consistently successful learning experience. The minimal variations imply that the advantages of VR-based language learning are comparable for all students, confirming its potential as an inclusive language learning assistant. However, an additional investigation could determine if certain VR elements have a greater impact on various skill levels.

Accuracy

The accuracy of grammar, vocabulary, pronunciation, and sentence structure in VR-based English language learning is referred to as accuracy. It assesses how effectively students adhere to accepted language use norms. Through immersive voice recognition, writing exercises, and interactive learning, virtual reality improves accuracy. Accuracy performance is shown in Figure 8 and Table 3. The ANN performed well in pattern recognition and classification, demonstrated by its 96.37% accuracy rate. The F-D2QLN’s slightly lower accuracy of 95.6% demonstrated its adaptability to dynamic learning conditions with some minor drawbacks. With an accuracy of 97.2%, the suggested DMBO-BiGRU performed better than both, proving its greater capacity to capture temporal dependencies in language acquisition. Therefore, it is a very useful tool for evaluating learning outcomes in virtual reality environments.

Table 3.

Overall values for the proposed and the existing methods.

Method	Accuracy (%)	VR score (%)	Training gain (%)	Detection score (%)
ANN²⁴	96.37	97.37	98.34	85.47
F-D2QLN²⁵	95.6	—	—	—
DMBO-BiGRU [proposed]	97.2	97.52	98.49	92.6

VR score

The VR score is a numerical indicator that evaluates how VR environments affect learning English. Vocabulary retention, pronunciation accuracy, understanding, engagement levels, and participation in the VR environment are some of the criteria utilized to assess students’ success.

Training gain

The improvement in English language proficiency that students attain following exposure to VR-based instruction in comparison to their starting skill level is referred to as training gain.

Detection score

The detection score indicates how well the system can evaluate and categorize students’ replies or interactions inside the VR environment.

The outcomes indicate that in determining how VR environments impact English language learning, the DMBO-BiGRU model is superior to the ANN, as illustrated in Figure 9 and Table 3. The ability of the model to determine the effectiveness of VR-based learning is expressed in the VR score, which is slightly greater for DMBO-BiGRU (97.52) than for ANN (97.37). Moreover, DMBO-BiGRU (98.49) performs better than ANN (98.34) in training gain, which signifies learning progressions over time. Moreover, DMBO-BiGRU has a significantly superior detection score (92.6) compared to ANN (85.47), which signifies accuracy in identifying key language learning patterns. This means that DMBO-BiGRU is a better choice for assessing learning development in immersive settings since it provides a more accurate and reliable assessment of VR-based language learning.

Retention rate

In the context of VR environments’ effects on English language learning, retention rate is the proportion of acquired vocabulary, language skills, or understanding that students can successfully retain and use after a certain amount of time. Throughout several iterations, the suggested DMBO-BiGRU model outperforms the F-D2QLN model in terms of retention rates when used in VR environments for English language learning, as shown in Figure 10 and Table 4. At iteration 1, DMBO-BiGRU outperforms F-D2QLN by showing more engagement from the beginning with a 73% retention rate compared to 70% for F-D2QLN. With 79% at iteration 2, 80% at iteration 4, and 82% at iteration 9, DMBO-BiGRU consistently performs better than F-D2QLN, which lags behind with 75%, 74%, and 75%, respectively. With DMBO-BiGRU achieving 93% at iteration 12 and 96% at iteration 13 compared to F-D2QLN’s 84% and 94%, an evident difference manifests in subsequent iterations, highlighting the model’s higher ability to sustain learner interest. This suggests the longer-term retention and enhanced language learning through the integration of VR-facilitated learning and BiGRU’s ability to recognize long-term patterns in sequential data.

Table 4.

Retention rate versus number of iterations.

No of iterations	Retention rate (%)
No of iterations	F-D2QLN²⁵	DMBO-BiGRU [proposed]
1	70	73
2	75	79
3	69	75
4	74	80
5	55	66
6	58	70
7	61	72
8	70	74
9	75	82
10	79	87
11	80	89
12	84	93
13	94	96

Discussion

English language learning is greatly impacted by VR environments that offer immersive, dynamic, and engaging educational opportunities. By practicing real-life communication in simulated environments, VR helps students improve their vocabulary, pronunciation, and fluency. The existing methods are F-D2QLN²⁵ and ANN.²⁴ The F-D2QLN’s high computational complexity is a drawback when it comes to examining how VR environments affect English language acquisition. Real-time adaption in immersive VR environments is difficult due to the increased processing cost caused by the fuzzy logic integration. Furthermore, D2QLN and other deep reinforcement learning models need a lot of training data and adjustment, which can make them less applicable to a wide range of language learners.²⁵ It can be difficult to obtain the vast volumes of high-quality labeled data that ANNs need for efficient training in instructional virtual reality environments. It’s also challenging to understand how certain VR components affect language learning results because of their black-box nature.²⁴ The DMBO-BiGRU improves sequence learning efficiency, captures bidirectional contextual dependencies, and dynamically adjusts model parameters to enhance the analysis of how virtual reality settings affect learning English. While BiGRU efficiently stores past and future contextual information, DMBO guarantees quicker convergence and better feature selection, which improves pronunciation recognition, sentiment analysis, and engagement monitoring in VR-based language learning.

Conclusion

VR has become an essential tool in education, especially in language learning, in the quickly changing digital world. It investigates how immersive VR technologies affect learning English, emphasizing the improvement of communication skills through engaging and dynamic learning environments. Noise filtering and normalization were used as preprocessing techniques to assure the quality of the dataset, which was gathered via student interactions in virtual reality environments. Fluency, language use, and contextual appropriateness are among the important variables that are extracted using t-SNE. The DMBO-BiGRU model evaluates language complexity, communication patterns, and learning development through learner interactions in VR environments. The suggested method is compared to the existing method in terms of accuracy (97.2%), VR score (97.52%), training gain (98.49%), detection score (92.6%), and the retention rate at iteration 13 (96%). VR-based English language instruction provides immersive and interactive learning; its drawbacks include expensive implementation costs, the requirement for specialist technology, and the possibility of motion sickness in users. For future enhancement of AI-driven adaptive learning in VR, increasing accessibility with affordable solutions, and incorporating real-time feedback mechanisms should be the primary targets.

Footnotes

ORCID iD

Yaping Lou

Funding

The author received no financial support for the research, authorship, and/or publication of this article.

Conflicting interests

The author declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

The authors declare that the data supporting the findings of this study are available within the article. The raw/derived data supporting the findings of this study are available from the corresponding author at request.*

References

AlGerafi

Zhou

Oubibi

, et al. Unlocking the potential: a comprehensive evaluation of augmented reality and virtual reality in education. Electronics 2023; 12(18): 3953.

Ahmad

Alam

Rahmat

, et al. Academic and administrative role of artificial intelligence in education. Sustainability 2022; 14(3): 1101.

Villegas-Ch

García-Ortiz

Sánchez-Viteri

. Educational advances in the metaverse: boosting learning through virtual and augmented reality and artificial intelligence. IEEE Access 2024; 12: 59093–59112. DOI: 10.1109/ACCESS.2024.3393776.

Wei

. Artificial intelligence in language instruction: impact on English learning achievement, L2 motivation, and self-regulated learning. Front Psychol 2023; 14: 1261955.

Zhang

Sun

Deng

. Design and application of intelligent classroom for English language and literature based on artificial intelligence technology. Appl Artif Intell 2023; 37(1): 2216051.

Zhang

. Secure preschool education using machine learning and metaverse technologies. Appl Artif Intell 2023; 37(1): 2222496.

Shafiee Rad

Roohani

. Fostering L2 learners’ pronunciation and motivation via affordances of artificial intelligence. Comput Sch 2024; 12: 1–22.

Qassrawi

ElMashharawi

Itmeizeh

, et al. AI-powered applications for improving EFL students’ speaking proficiency in higher education. Assessment 2024; 8: 9.

Tolba

Elarif

Taha

, et al. Interactive augmented reality system for learning phonetics using artificial intelligence. IEEE Access 2024; 12: 78219–78231.

10.

Dimitriadou

Lanitis

. A critical evaluation, challenges, and future perspectives of using artificial intelligence and emerging technologies in smart classrooms. Smart Learn Environ 2023; 10(1): 12.

11.

Lee

Hwang

. Technology-enhanced education through VR-making and metaverse-linking to foster teacher readiness and sustainable learning. Sustainability 2022; 14(8): 4786.

12.

Jawarneh

Alshar′

e M

Dewi

, et al. The impact of virtual reality technology on Jordan’s learning environment and medical informatics among physicians. International Journal of Computer Games Technology 2023; 2023(1): 1678226.

13.

Shi

Sitthiworachart

Hong

. Supporting project-based learning for students’ oral English skill and engagement with immersive virtual reality. Educ Inf Technol 2024; 29(11): 14127–14150.

14.

Tai

Chen

Todd

. The impact of a virtual reality app on adolescent EFL learners’ vocabulary learning. Comput Assist Lang Learn 2022; 35(4): 892–917.

15.

. Interactive virtual reality environments for English language learning: immersive technologies to enhance language acquisition and communication skills. J Comput Methods Sci Eng 2025; 10: 14727978251318803.

16.

Chen

Hsu

Lin

, et al. Robot-assisted language learning: integrating artificial intelligence and virtual reality into English tour guide practice. Educ Sci 2022; 12(7): 437.

17.

Divekar*

Drozdal*

Chabot*

Zhou

Chen

Zhu

Hendler

Braasch

. Foreign language acquisition via artificial intelligence and extended reality: design and evaluation. Comput Assist Lang Learn 2022; 35(9): 2332–2360.

18.

Qiao

Zhao

. Artificial intelligence-based language learning: illuminating the impact on speaking skills and self-regulation in Chinese EFL context. Front Psychol 2023; 14: 1255594.

19.

Lee

. Problem-based gaming via an augmented reality mobile game and a printed game in foreign language education. Educ Inf Technol 2022; 27(1): 743–771.

20.

Smutny

. Learning with virtual reality: a market analysis of educational and training applications. Interact Learn Environ 2023; 31(10): 6133–6146.

21.

Ironsi

. Investigating the use of virtual reality to improve speaking skills: insights from students and teachers. Smart Learn Environ 2023; 10(1): 53.

22.

Xue

Wang

. The interior environment design for entrepreneurship education under the virtual reality and artificial intelligence-based learning environment. Front Psychol 2022; 13: 944060.

23.

Hui

Guan

. Research on learning concentration recognition with multi-modal features in virtual reality environments. Sustainability 2023; 15(15): 11606.

24.

Luo

. Practice of artificial intelligence and virtual reality technology in college English dialogue scene simulation. Wireless Commun Mobile Comput 2022; 2022(1): 4922675–4922679.

25.

Sun

. Enhancing higher education English learning through virtual reality and game-based approaches using the fuzzy deep model. Comput Aided Des Appl 2023; 20: 231–248.