Dynamic Phonological Re-Grouping (DPR): An AI-Driven Framework for Real-Time Intra-Lesson Scaffolding in Teaching at the Right Level (TaRL) Classrooms

1. Introduction

1.1 Context and Rationale

The foundational-learning crisis has intensified attention to instructional approaches that begin with what learners can actually do rather than what a grade-level curriculum assumes they should already know. Teaching at the Right Level (TaRL), developed by Pratham in India, operationalizes this principle through brief assessment, competency-based grouping, focused foundational instruction, regular monitoring, and movement between groups as learners progress. Evidence from randomized evaluations has shown that targeted instruction can produce meaningful gains when teaching is aligned with learners’ demonstrated levels rather than an over-ambitious common curriculum.¹ UNICEF’s Nigeria TaRL guidance similarly emphasizes assessment, level-based grouping, focused foundational-skills activities, monitoring, and teacher support, while the wider UNICEF TaRL platform describes the approach as starting instruction from the learner’s current level.²

The organizational strength of TaRL, however, can coexist with a diagnostic limitation. Broad literacy classifications are useful for determining whether a learner is working at a letter, word, sentence, or connected-text level, but they do not necessarily reveal the acoustic and processing bottleneck that prevents successful listening. In an English as a Foreign Language (EFL) classroom, two learners who appear to occupy the same reading level may have sharply different listening profiles. One may struggle to discriminate phonemic contrasts; another may identify the sounds but fail to locate word boundaries in connected speech; a third may recognize the words yet lose the syntactic relationship among them when speech is delivered at a natural rate.³ Connected speech makes this problem particularly visible. Fluent speech contains co-articulation, reduction, assimilation, linking, elision, stress-based variation, and changes in syllable realization.⁴ Consequently, a learner who recognizes a carefully articulated word in isolation may not recognize the same lexical item in a naturally produced utterance. Such difficulty should not automatically be interpreted as low general ability. It may reflect a temporary mismatch between the acoustic form of the input and the learner’s developing perceptual representations.⁵

1.2 Problem Statement

Conventional TaRL assessment is intentionally brief and feasible. Its value lies partly in the fact that teachers can administer it without specialized equipment and use the result to form workable groups. The same practicality means that the assessment necessarily offers a coarse representation of competence. It may reveal that a learner cannot read a word or paragraph, but it does not necessarily explain why the learner failed to understand a spoken sentence moments earlier. The issue, therefore, is not that stable TaRL groups are pedagogically misguided. Rather, a learner’s immediate support requirement can change across tasks within the same lesson. This within-learner variability is difficult to observe continuously in crowded classrooms. A teacher may infer a vocabulary problem when the actual difficulty is phonemic segmentation, or repeat an entire sentence when the key obstacle is the perception of one sound contrast or a reduced form. Continuous monitoring of response accuracy, response latency, recovery after repetition, and error patterns is demanding when one teacher is responsible for many learners. DPR addresses this specific instructional problem by distinguishing between relatively stable macro-level TaRL groups and temporary phonological micro-cohorts. Macro-groups retain the social and organizational efficiency of TaRL; micro-cohorts provide short, task-specific support when evidence indicates a shared listening bottleneck.

1.3 The AI Intervention Gap

Speech technology has created new possibilities for immediate language feedback. Research on automatic speech recognition (ASR) in language learning indicates that ASR can support pronunciation development, particularly when feedback is explicit and when learners receive opportunities for repeated practice.⁶ A meta-analysis of 15 studies reported a medium overall effect of ASR-supported pronunciation learning (g = 0.69), with larger effects for segmental than suprasegmental outcomes and stronger results when ASR was combined with peer practice or explicit feedback.⁷ A five-week study involving 117 Indonesian secondary-school learners likewise found improvements in pronunciation measures after ASR-supported practice, with more detailed phonetic feedback producing stronger gains than global feedback.⁸ A recent systematic review, however, found that CAPT research remains concentrated on adult learners, controlled production tasks, and segmental features, leaving important opportunities for classroom-based, perception-oriented applications.⁹

The existing CALL ecosystem does not always fit the conditions of TaRL classrooms. Many systems assume reliable connectivity, individual devices, headphones, cloud processing, or autonomous practice. Those assumptions can be inconsistent with low-resource, group-based settings. The design challenge is therefore not simply to insert ASR into a classroom. It is to create an AI layer that respects TaRL’s organizational logic, keeps the teacher in control, and turns speech evidence into a small number of actionable instructional decisions. DPR is proposed as that layer. It uses brief, shared listening prompts; aggregates evidence around instructional decisions; and gives teachers simple regrouping recommendations rather than requiring them to interpret complex technical dashboards.

1.4 Purpose, Objectives, and Research Questions

The purpose of this conceptual paper is to develop a theoretically grounded and empirically testable framework for using edge-oriented speech technology to support temporary phonological regrouping within TaRL lessons. Four objectives guide the framework:

i. to specify how acoustic and response evidence can be used to identify probable micro-level listening bottlenecks during instruction;

ii. to develop a conservative model for temporary intra-lesson regrouping without destabilizing ordinary TaRL routines;

iii. to specify an adaptive scaffolding process that increases acoustic complexity while gradually withdrawing assistance; and

iv. to establish an empirical research agenda for testing diagnostic validity, learning effects, equity, usability, and sustainability in low-resource EFL settings.

The framework addresses two research questions:

i. How can edge-AI acoustic diagnostics systematically identify micro-level listening failures during instruction?

ii. What theoretical and algorithmic model can govern real-time, intra-lesson learner regrouping without disrupting classroom dynamics?

DPR does not claim that AI can directly infer a learner’s mental state or diagnose a listening deficit from audio alone. Its proposed evidence comes from observable responses to carefully designed tasks: forced-choice discrimination, segmentation responses, word-card reconstruction, short dictations, repetitions, or meaning checks. AI contributes rapid pattern detection; the teacher contributes interpretation, contextual knowledge, and pedagogical authority.

1.5 Contribution and Scope

The contribution of DPR is conceptual rather than empirical. The framework joins four strands that are often discussed separately: competency-based grouping in TaRL; bottom-up and connected-speech processes in listening; ASR and acoustic analysis; and human-centred, responsible AI. The paper does not present field data, an intervention effect size, or a validated diagnostic instrument. Accordingly, all claims about learning effects are framed as propositions for empirical testing. This distinction is important because recent ASR and AI-learning research demonstrates potential while also showing that performance depends on task, learner population, feedback design, model architecture, and implementation conditions.¹⁰

2. Conceptual and Theoretical Grounding

2.1 Conceptual Methodology

This paper adopts an integrative conceptual-synthesis approach. It is not presented as a systematic or scoping review and therefore does not claim PRISMA-style completeness. The framework was developed by synthesizing four bodies of scholarship:

i. TaRL and targeted foundational instruction;

ii. L2 listening, phonological processing, and connected speech;

iii. ASR/CAPT and speech-supported language learning; and

iv. responsible and inclusive AI.

The synthesis proceeded through four analytic questions. First, what learner evidence is observable quickly enough to inform instruction? Second, what evidence is sufficiently specific to distinguish a probable phonological bottleneck from a general low score? Third, what decision rules could make regrouping temporary and reversible? Fourth, what technical and ethical constraints are necessary for the framework to remain educationally useful in low-resource classrooms? The resulting architecture is therefore a design proposition: each technological function is included only when it can be connected to a teacher decision and a bounded instructional episode.

2.2 Zone of Proximal Development and Dynamic Support

Vygotsky’s¹¹ Zone of Proximal Development (ZPD) describes the space between independent performance and performance achieved with guidance. Wood et al.¹² similarly emphasize the tutor’s role in structuring assistance and gradually transferring responsibility to the learner. DPR extends this logic from a broad language level to a specific phonological task. A learner may independently recognize isolated words but require support when those words occur in connected speech. The relevant instructional zone is therefore task-sensitive and can change as acoustic demands change. DPR operationalizes this dynamic support through short cycles of probe, scaffold, practice, re-probe, and reintegration. The AI does not determine the learner’s ZPD. Instead, it shortens the interval between observable performance and a teacher-mediated response. This is consistent with the broader logic of mediated or dynamic assessment: performance after support is informative because it indicates how the learner responds to assistance, not merely whether the learner can succeed without it.

2.3 Cognitive Load Theory

Cognitive Load Theory distinguishes intrinsic, extraneous, and germane aspects of cognitive demand.¹³ Natural listening can impose substantial intrinsic demand because learners must process an acoustic signal, segment it, access lexical representations, construct syntactic relations, and integrate meaning in real time. In low-resource classrooms, extraneous demand may be added by noise, poor audio reproduction, unfamiliar accents, anxiety, or unclear task instructions. DPR responds by temporarily isolating one processing demand at a time. A diagnostic sequence may move from phonemic discrimination to word-boundary detection, lexical identification, phrase-level parsing, and finally meaning-level comprehension. Isolation is not intended to replace communicative listening. It is a short diagnostic and remedial step that allows the teacher to identify the likely bottleneck before returning the learner to richer input. The framework also treats excessive regrouping as a potential source of extraneous load; hence the use of confidence thresholds, minimum cohort sizes, transition windows, and teacher override.

2.4 Input, Comprehensibility, and Controlled Complexity

Krashen¹⁴ Input Hypothesis has been debated and refined, but the broader instructional intuition that input should be understandable while remaining developmentally challenging remains useful for designing graduated listening tasks. DPR therefore interprets the familiar idea of “i+1” conservatively: difficulty should increase in controlled dimensions rather than simultaneously increasing speech rate, vocabulary novelty, syntactic complexity, accent variability, and background noise. This design is also compatible with contemporary pronunciation scholarship that places intelligibility and comprehensibility above imitation of a single idealized native-speaker model.¹⁵ DPR is not intended to erase Nigerian or other regional English varieties. Its objective is to help learners recognize and process the forms they are likely to encounter, while ensuring that regional variation is not automatically converted into an error signal.

2.5 TaRL as the Organizational Anchor

TaRL’s central insight is that learners in the same grade do not necessarily share the same foundational learning needs. Its organizational structure depends on assessment, grouping, focused instruction, monitoring, and movement when learning changes.¹⁶ UNICEF’s Nigeria guide similarly presents TaRL as a practical model that can be adapted to local contexts while retaining its focus on current learning levels.¹⁷ DPR preserves these principles. It does not replace baseline, periodic, or teacher-mediated assessment; it adds a finer-grained layer inside an instructional episode. The distinction is therefore between macro and micro organization. A macro-group is a relatively stable TaRL grouping that supports the lesson’s overall social organization. A micro-cohort is a temporary grouping formed because several learners display the same task-specific phonological need. The distinction protects the efficiency of TaRL while allowing instruction to respond to short-term variability.

2.6 Listening Perception and Connected Speech

Listening comprehension is not a single operation. Learners must detect acoustic contrasts, segment continuous speech, map sound sequences to lexical representations, construct syntactic relationships, and integrate information with context.¹⁸ Connected speech complicates this process because fluent production rarely resembles dictionary forms. Reduction, linking, assimilation, co-articulation, stress timing, and elision change the acoustic evidence available to the listener.¹⁹ DPR treats these phenomena as teachable patterns rather than evidence of general deficiency. Its diagnostic categories are intentionally narrow: segmentation, lexical access, and phrase/sentence-level decoding. The framework also allows an “insufficient evidence” category because an incorrect response may result from vocabulary, memory, attention, audio quality, or task misunderstanding rather than phonological processing.

2.7 ASR, CAPT, and the Perception–Production Distinction

ASR-supported language learning has demonstrated potential for immediate feedback, but most CAPT research has focused on speech production rather than incoming-speech perception.²⁰ Farrús²¹ shows the breadth of ASR applications in L2 learning, while Amrate & Tsai²² note the continued concentration of CAPT research on controlled pronunciation production. Ngo et al.²³ further show that ASR effects vary according to feedback type, target feature, treatment duration, and learner characteristics. DPR shifts the question from “Did the learner pronounce the target correctly?” to “What evidence suggests that the learner perceived and processed the target in incoming speech?” Useful evidence can include forced-choice responses, word-boundary marking, sequence reconstruction, short dictation, or a spoken repetition accompanied by a meaning check. ASR and acoustic analysis are therefore complementary rather than sufficient by themselves. This distinction also creates an important technical safeguard. Phone-level alignment or ASR transcription should not be treated as equivalent to human phonological judgment. Model output can be affected by accent, age, speech rate, microphone quality, and mismatch between training data and local learners. The proposed system should consequently report uncertainty and diagnostic suggestions rather than high-stakes scores.²⁴

Table 1. Conceptual Mapping of the DPR Framework

3. Conceptual Architecture of DPR

3.1 Architecture and instructional episode

DPR is organized as a closed instructional loop. A shared listening prompt generates learner responses; RADM estimates a probable bottleneck and confidence level; IRL determines whether temporary regrouping is justified; the selected micro-cohort completes a short scaffolded activity; ASDE reduces assistance as mastery emerges; and a re-probe returns evidence to the teacher. The cycle is designed to fit inside normal TaRL instruction rather than create a separate technology lesson.

Note. The architecture is conceptual and requires empirical validation before deployment.

3.2 Component 1: Real-Time Acoustic Diagnostic Module (RADM)

RADM collects evidence through short, carefully designed listening probes. A probe might ask learners to choose between two minimal-pair words, mark the number of words heard in a phrase, identify a missing word in a connected-speech sentence, reconstruct a short phrase from word cards, repeat a target phrase, or indicate whether two spoken forms are the same or different. A meaning check can be added when a phonological interpretation might otherwise be confounded with vocabulary or memory. Multiple probes are preferable to a single score because one task can confound listening perception with vocabulary, attention, memory, task familiarity, or anxiety. RADM should therefore distinguish observed performance from inferred cause. If a learner selects the wrong word, the system may flag a likely segmentation or lexical-access issue, but the teacher should be able to reject the interpretation. For conceptual purposes, the diagnostic state for learner ‘i’ at time ‘t’ can be represented as:

where s denotes phonemic-segmentation performance,

l lexical-access performance,

y phrase/syntactic decoding performance,

r response latency or recovery after repetition, and

q the combined audio-quality/model-confidence estimate. The vector is an internal representation rather than a learner-facing score.

Teacher-facing outputs should be deliberately simple: “Needs sound-boundary practice”; “Recognizes the sound sequence but not the word”; “Understands words; needs phrase-level practice”; or “Evidence insufficient teacher review required.” The last category is essential because abstention is preferable to a confident but unreliable classification. Confidence should fall when the recording is noisy, voices overlap, the signal is clipped or too quiet, the learner provides too little response evidence, the model encounters an unfamiliar accent or language background, or the learner appears not to have attempted the task. Low confidence should trigger a repeat probe or teacher review, not automatic regrouping.

3.3 Component 2: Intra-Lesson Re-Grouping Logic (IRL)

IRL converts diagnostic evidence into a temporary classroom decision. Its logic is intentionally conservative: learners move only when the evidence is repeated, the pattern appears instructional rather than incidental, and the movement can occur without undermining the lesson. DPR proposes three phonological sub-tiers, labelled by instructional activity instead of ability:

P1—sound/segmentation;

P2—sound-to-word/lexical access; and

P3—phrase/connected-speech processing.

These sub-tiers are not fixed developmental levels. A learner can be P1 for one phonological feature and P3 for another. The system should therefore treat tier membership as a task-state variable instead of a stable learner identity. Four controls make IRL deliberately resistant to over-reaction. First, persistence requires the same pattern across at least two short probes. Second, hysteresis makes the threshold for leaving a micro-cohort slightly higher than the threshold for entering it, reducing oscillation between groups. Third, a minimum cohort size protects the social and pedagogical value of peer interaction. Fourth, teacher override permits a contextual decision when regrouping would confuse, stigmatize, or unnecessarily interrupt instruction.

Table 2. Proposed DPR Decision Matrix for Temporary Phonological Regrouping

3.4 Component 3: Adaptive Scaffolding-Decay Engine (ASDE)

ASDE governs the introduction and gradual withdrawal of support. “Decay” refers to the controlled reduction of assistance, not the deterioration of learning. Five dimensions can be manipulated: speech rate, prosodic salience, co-articulation, lexical familiarity, and signal-to-noise ratio. The framework recommends changing one major dimension at a time wherever possible. Speech rate may begin below the natural classroom rate and then move toward locally appropriate natural speech. Prosodic cues can be moderately enhanced to make stress or boundaries noticeable before being reduced. Co-articulation can begin with more carefully separated words and progressively restore natural linking and reduction. Lexical familiarity can be high at the diagnostic stage and then broaden. If noise is pedagogically relevant, it should be introduced in controlled, replicable audio rather than by deliberately making the classroom environment noisier. A practical progression is C0: isolated sound or syllable; C1: familiar word in clear speech; C2: phrase with controlled linking; and C3: short sentence in natural connected speech. Progression should occur only after performance is stable across more than one item. An 80% criterion may be used as a provisional design parameter, but it is not a universal psychological threshold and should be locally validated before field use.

Table 3. Proposed Scaffolding-Decay Sequence

3.5 Teacher-in-the-Loop Design

The teacher should see only the information required for an immediate instructional decision: the recommended tier, the suggested task, the confidence level, the remaining cycle time, and whether the learner appears to have met the exit criterion. During active teaching, the interface should avoid complex graphs, public rankings, permanent labels, deficit-oriented colour coding, unexplained pronunciation scores, and interruptions that compete with teacher talk. The teacher should be able to select “accept,” “modify,” “ignore,” or “repeat diagnostic.” These actions are not merely interface controls; they are important implementation data. If teachers repeatedly override a recommendation, the system may be detecting a systematic mismatch between its model and classroom reality. A mature DPR system should therefore learn from teacher corrections as well as learner responses, while preserving human authority over consequential decisions.

4. Pedagogical Implementation and Technical Feasibility

4.1 Classroom workflow

DPR can be embedded in a 60-90-minute TaRL lesson without replacing the broader methodology. The sequence below is intentionally lightweight. It assumes that the teacher remains responsible for the lesson objective, group management, and interpretation of learner evidence.

Table 4. Illustrative 60–90 Minute DPR-Enabled Tarl Lesson Workflow

Learners should be able to respond through low-tech means such as fingers, cards, word cards, gestures, counters, or choral repetition. A device does not need to capture every learner continuously. In a low-resource implementation, evidence may be collected from rotating representatives, small-group microphones, or structured response patterns. The purpose is instructional noticing, not continuous biometric surveillance. Micro-groups should be named after activities rather than learner status. “Sound,” “Word,” and “Phrase” are preferable to labels such as “weak,” “average,” and “strong.” The re-probe should use a parallel item rather than the exact diagnostic item so that improvement is not confused with memorization. Successful cycles should end in reintegration into the broader TaRL group; DPR is not designed to create permanent specialist groups.

4.2 Technical Infrastructure

A minimally viable DPR implementation could operate on one low-cost Android smartphone or tablet per teacher, a small portable speaker, optional low-cost microphones for rotating groups, locally stored audio and model files, a battery pack or solar charging unit where electricity is unreliable, and an offline teacher interface. A hybrid architecture is preferable to a cloud-dependent design because it allows core functions to remain available when connectivity is weak. Essential functions should include audio playback, simple response capture, local scoring, group recommendation, and offline storage. Optional functions can include detailed phoneme alignment, teacher analytics, model retraining, and periodic synchronization. Personal identifiers should be separated from performance records; synchronization should occur only through trusted connections; model updates should be verified; and raw audio should be deleted after a short retention period unless explicit research consent supports retention. Large multilingual ASR models can be computationally expensive. DPR does not necessarily require open-ended transcription. A small task-specific acoustic classifier may be more appropriate for a constrained classroom task.²⁵ Knowledge distillation, quantization, pruning, and task-specific modelling are possible engineering strategies, but the framework deliberately leaves model architecture open until empirical requirements are known.

Table 5. Minimum Viable Technical Architecture

4.3 Local Calibration and Linguistic Inclusion

A model trained primarily on dominant American or British English may misinterpret Nigerian, Ghanaian, Kenyan, or other regional English varieties. Recent research demonstrates that ASR bias can vary by accent, age, gender, language, and model architecture, and that pronunciation differences explain only part of observed disparities.²⁶ A separate audit of widely used English ASR services likewise reported performance disparities associated with accent and speaker background.²⁷ Calibration should therefore include local teachers, learners across relevant language backgrounds, multiple acceptable realizations of target words, locally meaningful vocabulary, representative classroom-noise recordings, and human-reviewed examples of correct and intelligible responses. The system should optimize instructional usefulness rather than similarity to a single prestige accent. In Nigerian classrooms, an acoustic difference may reflect legitimate Nigerian English variation, L1 transfer, ASR error, poor signal quality, or a genuine perception difficulty. These possibilities must remain distinguishable in the design. Bias auditing should be routine instead of an end-stage certification exercise. Developers should report diagnostic performance across relevant linguistic backgrounds and should involve local educators and language specialists in dataset design, threshold calibration, error analysis, and interpretation. Learners and teachers should have a meaningful route to challenge an output.

4.4 Teacher Preparation

Teacher preparation should focus on interpretation and instructional use rather than programming. Training should cover common phonological listening bottlenecks, the meaning of P1–P3 recommendations, recognition of model uncertainty, acceptance/modification/rejection of recommendations, peer-mediated activities, privacy and learner dignity, and recording of implementation difficulties. Teachers should also learn when not to use DPR. If equipment is unreliable, the class is emotionally unsettled, or the day’s objective is primarily creative reading or discussion, conventional TaRL instruction may be the more appropriate choice.

5. Discussion

5.1 Theoretical Contribution

DPR makes three linked theoretical contributions. First, it extends TaRL from level-based organization toward task-responsive instructional ecology. Stable macro-groups remain the backbone of the lesson, but temporary phonological cohorts allow the teacher to respond to a specific listening need without redefining the learner’s overall level. Second, DPR connects acoustic phonetics to sociocultural pedagogy: acoustic evidence becomes useful only when interpreted through teacher mediation, peer interaction, gestures, word cards, and shared meaning-making. Third, DPR reframes AI adaptation as decision support rather than automated personalization. This position is consistent with human-centred AI guidance that emphasizes agency, inclusion, transparency, and institutional responsibility.²⁸ The framework also makes a methodological contribution by treating the instructional episode; not the individual score as the central unit of adaptation. A score detached from task, acoustic conditions, and support can be misleading. An episode, by contrast, records the prompt, response, probable bottleneck, scaffold, re-probe, and teacher decision. This makes the framework more compatible with formative assessment and classroom inquiry.

5.2 Pedagogical Value and Friction

The potential pedagogical value of DPR lies in timing. TaRL already gives teachers a mechanism for matching instruction to demonstrated need; DPR proposes a mechanism for making one aspect of that matching more immediate. The expected benefit is therefore not necessarily a new category of pedagogy but a shorter distance between noticing and responding. This is particularly relevant to listening, where the obstacle may emerge during a few seconds of connected speech and disappear when the input is slowed or segmented. The principal pedagogical risk is fragmentation. If every small difference triggers a new group, classroom routines become unstable and learners may feel continuously ranked. DPR therefore prioritizes shared bottlenecks, short cycles, activity-based group names, and teacher discretion. A cohort of four learners who share a segmentation difficulty may be pedagogically useful; four separate one-learner groups usually are not. A second risk is teacher overload. An AI system that generates too many alerts becomes another classroom-management problem. The interface should therefore be sparse, with a limited number of actionable recommendations and an option to mute or postpone the system. The teacher’s workload should be treated as a primary outcome in future trials, not as an implementation footnote.

5.3 Algorithmic Bias and Linguistic Justice

ASR systems are not neutral listeners. Koenecke et al.²⁹ documented racial disparities in automated speech recognition, while subsequent work has shown that bias can also be associated with regional and non-native accents and can vary across model architectures.³⁰ In an EFL classroom, this creates a fundamental interpretive problem: a model may be less accurate because the learner’s speech differs from the model’s training distribution, not because the learner has failed to learn. DPR should therefore distinguish at least five possibilities: an intelligibility-relevant perception difficulty; a legitimate regional English feature; L1-related phonetic transfer; an ASR recognition failure; and insufficient audio quality. This distinction is central to linguistic justice. The system should not reward approximation to an American or British accent when the instructional goal is successful discrimination, recognition, processing, and comprehension. Calibration and auditing must be designed around the target learning objective rather than around the prestige of the model’s training data.

5.4 Privacy, Child Protection, and Responsible AI

Student audio can contain voice characteristics, linguistic background, identity information, and evidence about educational performance. UNESCO’s human-centred guidance stresses data protection, human agency, inclusion, equity, and institutional validation of AI tools in education.³¹ UNICEF’s child-centred AI guidance emphasizes safety, privacy, fairness, transparency, inclusion, and the best interests of children.³² These principles should be translated into concrete DPR safeguards.

i. collect only the audio necessary for the immediate instructional task;

ii. process audio locally whenever feasible;

iii. avoid retaining raw recordings by default;

iv. separate learner identity from diagnostic records;

v. obtain appropriate parental/guardian consent where required and explain participation in age-appropriate language;

vi. prohibit use of DPR outputs for promotion, punishment, permanent labelling, or high-stakes ranking;

vii. encrypt retained data and define a clear deletion schedule; and

viii. provide a non-digital alternative for learners who cannot or do not wish to participate.

Privacy is therefore not merely a technical property. Teachers, school leaders, parents, and learners should understand what DPR can infer, what it cannot infer, how decisions are made, who can access the records, and when data will be deleted.

5.5 Limitations of the Framework

The framework has several limitations. First, it is conceptual and has not demonstrated causal effects on listening comprehension, reading achievement, or participation. Second, the three phonological sub-tiers simplify processes that are overlapping in real learners; segmentation and lexical access may fail simultaneously. Third, the framework assumes that teachers have sufficient time and support to interpret recommendations, which may not be true in all TaRL implementations. Fourth, edge AI does not remove technical constraints. Low-cost devices may have limited memory, battery life, microphone quality, or processing capacity. Fifth, thresholds cannot be assumed to transfer across languages, age groups, regions, or device types. Sixth, a narrow focus on measurable acoustic features could unintentionally reduce listening to a technical decoding exercise. Listening is also strategic, social, contextual, and meaning-oriented.³³ DPR should therefore occupy a limited and purposeful part of a wider listening pedagogy. Finally, the framework does not prescribe a single AI architecture. The most technically sophisticated model may not be the most educationally appropriate. The relevant criterion is whether the system produces reliable, interpretable, equitable, and actionable evidence under real classroom constraints.

6. Future Empirical Agenda

The framework is intended to generate a programme of empirical research instead of a single efficacy trial. The research sequence should begin with feasibility and diagnostic validity before progressing to learning effects, equity, and sustainability.

Table 6. Proposed Four-Phase Empirical Validation Programme

6.1 Measurement Priorities

Diagnostic validity should be established before claims about instructional effectiveness. Learner performance on DPR probes should be compared with trained teacher ratings, human-coded perception tasks, phonological-awareness measures, delayed listening tests, reading and vocabulary measures, and audio-quality indicators. Agreement statistics, sensitivity, specificity, false-positive rates, and abstention rates should be reported alongside overall accuracy. This is especially important because a system that produces frequent false regrouping decisions could harm classroom efficiency even if its average classification accuracy appears acceptable. A subsequent cluster-randomized or stepped-wedge study could compare three conditions: standard TaRL; TaRL plus non-adaptive audio practice; and TaRL plus DPR. Primary outcomes should include listening comprehension, phonemic segmentation, connected-speech recognition, reading fluency, and learner participation. Secondary outcomes should include teacher workload, instructional time, learner confidence, regrouping frequency, and implementation fidelity. Because DPR is a classroom technology, process outcomes are necessary for interpreting whether any observed learning effect is attributable to the algorithm, the additional practice time, or the broader implementation package.³⁴

6.2 Equity and Robustness Testing

Equity testing should examine diagnostic performance across relevant linguistic backgrounds, accents, ages, genders, prior exposure to English, and disability-related needs where ethically and methodologically appropriate. Researchers should also test performance under different noise conditions and across different device types. A system should not be scaled simply because it performs well on a narrow benchmark. Robustness should be established in the environments in which the system is intended to operate. Sustainability studies should document device costs, battery and connectivity requirements, teacher training time, maintenance needs, model-update procedures, data governance, and the feasibility of local technical support. A useful DPR system is therefore not necessarily the most technically advanced one. It is the system teachers can trust, learners can use without stigma, and schools can sustain after external project support ends.

7. Conclusion

Dynamic Phonological Re-Grouping has been proposed as an AI-assisted extension of Teaching at the Right Level for EFL classrooms in which broad competency groups may conceal moment-to-moment listening barriers. The framework preserves TaRL’s central principles: assessment, level-sensitive instruction, teacher support, and movement based on demonstrated learning while adding a temporary layer of phonological diagnosis inside the lesson. The framework consists of three linked components. RADM identifies probable difficulties in segmentation, lexical access, and phrase-level decoding through brief, teacher-mediated probes. IRL translates repeated evidence into temporary micro-cohorts while protecting classroom stability through confidence thresholds, persistence, hysteresis, minimum cohort size, and teacher override. ASDE adjusts acoustic and linguistic complexity and gradually withdraws support as performance stabilizes. Together, the components define a closed instructional episode rather than a permanent learner classification. The central proposition is intentionally restrained: AI should not replace the teacher or convert TaRL into individualized software tutoring. Its educational value, if future evidence supports it, would lie in reducing the delay between noticing a specific learning bottleneck and responding to it. For low-resource EFL classrooms, that principle places pedagogical fit, local calibration, linguistic justice, privacy, and sustainability above technical novelty. DPR should therefore be treated as a testable framework whose value must be demonstrated through transparent, context-sensitive empirical research.

Endnotes

1. Curtiss Wyss, M., Qargha, G.O., Arenge, G., Mukoyi, T., Elliott, M., Matsheng, M. & Clune, K. (2023). Adapting, innovating, and scaling foundational learning: Four lessons from scaling Teaching at the Right Level in Botswana. Washington, D.C.: Center for Universal Education at the Brookings Institution.; Banerjee, A.V., Banerji, R., Berry, J., Duflo, E., Kannan, H., Mukerji, S., Shotland, M., & Walton, M. (2016). Mainstreaming an Effective Intervention: Evidence from Randomized Evaluations of 'Teaching at the Right Level' in India. MIT Department of Economics Working Paper No. 16-08, Available at SSRN: https://ssrn.com/abstract=2846971 or http://dx.doi.org/10.2139/ssrn.2846971

2. UNICEF. (2024). Teaching at the right level (TaRL) developed by Indian NGO Pratham, enables children to acquire foundational reading and arithmetic skills quickly. https://www.unicef.org/flnhub/teaching-right-level-tarl

3. Field, J. (2008). Listening in the language classroom. Cambridge University Press.; Goh, C. (2000). A Cognitive Perspective on Language Learners’ Listening Comprehension Problems. System, 28, 55-75. http://dx.doi.org/10.1016/S0346-251X(99)00060-3; Vandergrift, L., & Goh, C. C. M. (2012). Teaching and Learning Second Language Listening: Metacognition in Action. Routledge.

4. Bayaro, M, Olorunfemi, P.A, Michael, C.I, Adeleke, H. A. (2026). Beyond Native Norms: Decolonizing AI Speech Recognition for Accent Equity in Nigerian ESL Classrooms. Center of Artificial Intelligence,1(2), 111-121. https://doi.org/10.65591/CAI-206-2026; Olorunfemi, P.A, Nnaji, C.M & Michael, C.I. (2026). An Investigation into AI-Assisted Aspect-Based and Integrated English Teaching Approaches Using ADLL and CPCT Theories. Center of Artificial Intelligence. 1(2). 84-95. Doi: 10.65591/CAI-147-2026; Boulhrir, T., Ghreir, H., Hamash, M., & Robert, M. (2026). Artificial Intelligence in Education: Mapping Adaptive Learning and Learning Analytics in K–12 Online, Virtual, and Distance Learning. The International Review of Research in Open and Distributed Learning, 27(2), 122–148. https://doi.org/10.19173/irrodl.v27i2.9370

5. Henderson, A. (2014). RICHARD CAULDWELL , Phonology for listening: Teaching the stream of speech . Birmingham: Speech in Action, 2013. Pp. 332. ISBN-13: 978-0954344726. Journal of the International Phonetic Association, 44, 180 - 182. Doi: 10.1017/S0025100313000352; Cutler, A. (2012). Native listening: Language experience and the recognition of spoken words. MIT Press

6. Caines, A., Benedetto, L., Taslimipoor, S., Davis, C., Gao, Y., Andersen, Ø., Yuan, Z., Elliott, M., Moore, R., Bryant, C., Rei, M., Yannakoudakis, H., Mullooly, A., Nicholls, D., & Buttery, P. (2023). On the application of large language models for language teaching and assessment technology. arXiv. https://arxiv.org/abs/2307.08393; Wang, F., Cheung, A. C. K., & Chai, C. S. (2024). Language learning development in human-AI interaction: A thematic review of the research landscape. System, 125, Article 103424. DOI: http://dx.doi.org/10.1016/j.system.2024.103424; Boulhrir, T., Ghreir, H., Hamash, M., & Robert, M. (2026). Artificial Intelligence in Education: Mapping Adaptive Learning and Learning Analytics in K–12 Online, Virtual, and Distance Learning. The International Review of Research in Open and Distributed Learning, 27(2), 122–148. https://doi.org/10.19173/irrodl.v27i2.9370

7. Ngo, T.-N. N., Chen, H.-J. H., & Lai, K.-W. K. (2024). The effectiveness of automatic speech recognition in ESL/EFL pronunciation: A meta-analysis. ReCALL, 36(1), 4–21. https://doi.org/10.1017/S0958344023000113

8. Bashori, M., van Hout, R., Strik, H., & Cucchiarini, C. (2024). I can speak: Improving English pronunciation through automatic speech recognition-based language learning systems. Innovation in Language Learning and Teaching, 18(5), 443–461. https://doi.org/10.1080/17501229.2024.2315101

9. Amrate, M., & Tsai, P.-H. (2025). Computer-assisted pronunciation training: A systematic review. ReCALL, 37(1), 22–42. https://doi.org/10.1017/S0958344024000181

10. Farrús, M. (2023). Automatic speech recognition in L2 learning: A review based on PRISMA methodology. Languages, 8(4), 242. https://doi.org/10.3390/languages8040242; Amrate, M., & Tsai, P.-H. (2025). Computer-assisted pronunciation training: A systematic review. ReCALL, 37(1), 22–42. https://doi.org/10.1017/S0958344024000181

11. Vygotsky, L. S. (1978). Mind in society: The development of higher psychological processes. Harvard University Press.

12. Wood, D., Bruner, J. S. & Ross, G. (1976). The role of tutoring in problem solving. Journal of Child Psychology and Psychiatry, 17, 89-100.

13. Sweller, J. (1988) Cognitive Load during Problem Solving: Effects on Learning. Cognitive Science, 12, 257-285. https://doi.org/10.1207/s15516709cog1202_4; Sweller, J., Ayres, P. and Kalyuga, S. (2011) Cognitive Load Theory. Springer, New York. https://doi.org/10.1007/978-1-4419-8126-4

14. Krashen, S.D. (1985) The Input Hypothesis: Issues and Implications. Longman, London.

15. Derwing, T. M., & Munro, M. J. (2015). Pronunciation Fundamentals: Evidence-Based Perspectives for L2 Teaching and Research (Vol. 42). John Benjamins Publishing Company. https://doi.org/10.1075/lllt.42; Bayaro, M, Olorunfemi, P.A, Michael, C.I, Adeleke, H. A. (2026). Beyond Native Norms: Decolonizing AI Speech Recognition for Accent Equity in Nigerian ESL Classrooms. Center of Artificial Intelligence,1(2), 111-121. https://doi.org/10.65591/CAI-206-2026

16. Banerjee, A.V., Banerji, R., Berry, J., Duflo, E., Kannan, H., Mukerji, S., Shotland, M., & Walton, M. (2016). Mainstreaming an Effective Intervention: Evidence from Randomized Evaluations of 'Teaching at the Right Level' in India. MIT Department of Economics Working Paper No. 16-08, Available at SSRN: https://ssrn.com/abstract=2846971 or http://dx.doi.org/10.2139/ssrn.2846971; Abdul Latif Jameel Poverty Action Lab (J-PAL). (2018). Teaching at the Right Level to improve learning (Evidence to Policy Case Study). https://www.povertyactionlab.org/case-study/teaching-right-level-improve-learning

17. UNICEF. (2024). Teaching at the right level (TaRL) developed by Indian NGO Pratham, enables children to acquire foundational reading and arithmetic skills quickly. https://www.unicef.org/flnhub/teaching-right-level-tarl

18. Field, J. (2008). Listening in the language classroom. Cambridge University Press.; Goh, C. (2000). A Cognitive Perspective on Language Learners’ Listening Comprehension Problems. System, 28, 55-75. http://dx.doi.org/10.1016/S0346-251X(99)00060-3; Vandergrift, L., & Goh, C. C. M. (2012). Teaching and Learning Second Language Listening: Metacognition in Action. Routledge.

19. Henderson, A. (2014). RICHARD CAULDWELL , Phonology for listening: Teaching the stream of speech . Birmingham: Speech in Action, 2013. Pp. 332. ISBN-13: 978-0954344726. Journal of the International Phonetic Association, 44, 180 - 182. Doi: 10.1017/S0025100313000352; Cutler, A. (2012). Native listening: Language experience and the recognition of spoken words. MIT Press

20. Wang, F., Cheung, A. C. K., & Chai, C. S. (2024). Language learning development in human-AI interaction: A thematic review of the research landscape. System, 125, Article 103424. DOI: http://dx.doi.org/10.1016/j.system.2024.103424

21. Farrús, M. (2023). Automatic speech recognition in L2 learning: A review based on PRISMA methodology. Languages, 8(4), 242. https://doi.org/10.3390/languages8040242

22. Amrate, M., & Tsai, P.-H. (2025). Computer-assisted pronunciation training: A systematic review. ReCALL, 37(1), 22–42. https://doi.org/10.1017/S0958344024000181

23. Ngo, T.-N. N., Chen, H.-J. H., & Lai, K.-W. K. (2024). The effectiveness of automatic speech recognition in ESL/EFL pronunciation: A meta-analysis. ReCALL, 36(1), 4–21. https://doi.org/10.1017/S0958344023000113

24. Feng, S., Halpern, B. M., Kudina, O., & Scharenborg, O. (2024). Towards inclusive automatic speech recognition. Computer Speech & Language, 84, 101567. https://doi.org/10.1016/j.csl.2023.101567

25. Xu, H., Gan, W., Qi, Z., Wu, J., & Yu, P. S. (2024). Large language models for education: A survey. IEEE Transactions on Learning Technologies, 17, 340–355. ArXiv, abs/2405.13001

26. Feng, S., Halpern, B. M., Kudina, O., & Scharenborg, O. (2024). Towards inclusive automatic speech recognition. Computer Speech & Language, 84, 101567. https://doi.org/10.1016/j.csl.2023.101567; Koenecke, A., Nam, A., Lake, E., Nudell, J., Quartey, M., Mengesha, Z., Toups, C., Rickford, J. R., Jurafsky, D., & Goel, S. (2020). Racial disparities in automated speech recognition. Proceedings of the National Academy of Sciences, 117(14), 7684–7689. https://doi.org/10.1073/pnas.1915768117

27. DiChristofano, A., Shuster, H., Chandra, S., & Patwari, N. (2023). Performance disparities between accents in automatic speech recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, 37(13), 16200–16201. https://doi.org/10.1609/aaai.v37i13.26960

28. Miao, F., & Holmes, W. (2023). Guidance on Generative AI in Education and Research (pp. 14-17). UNESCO.; UNICEF. (2024). Teaching at the right level (TaRL) developed by Indian NGO Pratham, enables children to acquire foundational reading and arithmetic skills quickly. https://www.unicef.org/flnhub/teaching-right-level-tarl

29. Koenecke, A., Nam, A., Lake, E., Nudell, J., Quartey, M., Mengesha, Z., Toups, C., Rickford, J. R., Jurafsky, D., & Goel, S. (2020). Racial disparities in automated speech recognition. Proceedings of the National Academy of Sciences, 117(14), 7684–7689. https://doi.org/10.1073/pnas.1915768117

30. DiChristofano, A., Shuster, H., Chandra, S., & Patwari, N. (2023). Performance disparities between accents in automatic speech recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, 37(13), 16200–16201. https://doi.org/10.1609/aaai.v37i13.26960; Feng, S., Halpern, B. M., Kudina, O., & Scharenborg, O. (2024). Towards inclusive automatic speech recognition. Computer Speech & Language, 84, 101567. https://doi.org/10.1016/j.csl.2023.101567

31. Miao, F., & Holmes, W. (2023). Guidance on Generative AI in Education and Research (pp. 14-17). UNESCO.

32. UNICEF. (2024). Teaching at the right level (TaRL) developed by Indian NGO Pratham, enables children to acquire foundational reading and arithmetic skills quickly. https://www.unicef.org/flnhub/teaching-right-level-tarl

33. Graham, S. (2017). Research into practice: Listening strategies in an instructed classroom setting. Language Teaching, 50(1), 107–119. DOI: 10.1017/S0261444816000306; Vandergrift, L., & Goh, C. C. M. (2012). Teaching and Learning Second Language Listening: Metacognition in Action. Routledge.

34. Caines, A., Benedetto, L., Taslimipoor, S., Davis, C., Gao, Y., Andersen, Ø., Yuan, Z., Elliott, M., Moore, R., Bryant, C., Rei, M., Yannakoudakis, H., Mullooly, A., Nicholls, D., & Buttery, P. (2023). On the application of large language models for language teaching and assessment technology. arXiv. https://arxiv.org/abs/2307.08393