Technology-Mediated Pronunciation Feedback: Rapid Evidence Map

Technology-mediated pronunciation feedback can help learners notice and practise specific speech features, but the evidence is strongest for bounded training tasks, not for a universal automated “accent score.” Useful systems combine a clear target, repeated production, interpretable feedback, and an outcome that tests intelligibility or comprehensibility beyond the exact training item. The research does not justify treating transcription success as a complete measure of pronunciation.

Scope and method

This rapid evidence map was searched and accessed on 29 July 2026. It prioritised peer-reviewed reviews, meta-analyses, and empirical studies of computer-assisted pronunciation training (CAPT), automatic speech recognition (ASR), visual feedback, and peer-supported technology use. The search scope was English-language scholarship discoverable through publisher pages, PubMed, ERIC, and reference lists. It also used official framework material where a measurement distinction required a non-commercial definition.

The map included adult or adolescent second-language learners and outcomes that reported pronunciation, accentedness, comprehensibility, intelligibility, or a broader speaking measure. It excluded speech-therapy interventions, first- language literacy tools, product testimonials, unpublished vendor results, and studies whose only outcome was learner satisfaction. Conference abstracts were used only to identify questions, not as decisive efficacy evidence.

This is a rapid map rather than a preregistered systematic review. It does not claim exhaustive database coverage, duplicate independent screening, formal risk-of-bias scoring, or a pooled effect estimate.

Direct evidence in view

Evidence item	Intervention and outcome	What it supports	What it cannot establish
Amrate and Tsai, 2025 systematic review	Empirical CAPT studies with pronunciation outcomes	The field contains multiple feedback designs and generally favours focused pronunciation measures	One effect for every tool, language, learner, and speech task
Sun, 2023 mixed-method comparison	ASR plus peer correction versus teacher-led work with 61 intermediate Chinese EFL learners	A structured technology-plus-peer condition can improve selected controlled and spontaneous measures in one setting	ASR alone caused every observed difference or results transfer to other populations
Saito, 2019 measurement meta-analysis	77 pronunciation-teaching studies coded by construct, scoring method, and elicitation task	Outcome choice changes what “pronunciation improvement” means	A consumer app score is valid without its own validation
Koenecke and colleagues, 2020 audit	Commercial ASR transcription errors across matched speech corpora	Speech systems can perform differently across speaker groups	Every pronunciation engine has the same disparity or ASR error equals learner ability

The 2025 ReCALL review limited its corpus to peer-reviewed empirical CAPT studies and noted that global outcomes such as intelligibility and comprehensibility were comparatively uncommon (review and method). That matters: a learner may improve on a trained contrast or read-aloud item without yet becoming easier to understand in unprepared conversation.

Sun's 2023 study combined ASR with peer correction rather than testing a pure software condition. The reported design used two classes, read-aloud and spontaneous tasks, an IELTS speaking measure, and qualitative interviews (study record; full method). Its value is the broader outcome set. Its bundled treatment and one-site sample prevent the simple conclusion that an ASR feature will produce the same effect by itself.

Saito's measurement framework separates global from specific pronunciation, human from acoustic scoring, and controlled from spontaneous elicitation (meta-analysis and accepted manuscript). A credible product evaluation should cross these dimensions. If training is word-level, the test should still include unfamiliar words or a short spontaneous task. If the system produces a machine score, at least part of the validation should use independent human judgements tied to an explicit construct.

Four feedback mechanisms that should not be merged

Recognition feedback

A recogniser returns a transcript or indicates that an utterance was not decoded. This can reveal a mismatch worth inspecting. It does not tell the learner whether the problem was a target sound, rhythm, microphone, noise, lexical choice, or model coverage. “The recogniser understood me” is therefore an observation about one system under one condition, not proof of generally intelligible speech.

Pronunciation assessment

A dedicated assessment model compares speech features with a target and may return phoneme, word, or prosody feedback. Its validity depends on the reference population, scoring construct, audio conditions, language variety, and decision being made. The validation burden rises sharply when a score affects placement, certification, or access.

Visual or acoustic feedback

Waveforms, pitch tracks, spectrograms, and articulatory illustrations can make a feature inspectable. They are most useful when a teacher or protocol explains which difference matters. A dense display without an actionable target can add measurement without adding learning.

Human interpretation around technology

Peer or teacher review can turn a machine signal into a meaningful correction, identify an acceptable regional form, and decide when continued drilling is counterproductive. This layer makes the treatment harder to attribute to software alone, but often makes the learning design more defensible.

A minimum evaluation matrix

Do not ask only whether the average score increased. Predefine a matrix:

Dimension	Minimum evidence
Target	Exact segmental or suprasegmental feature being trained
Transfer	At least one untrained item and one less-controlled speaking task
Listener outcome	Independent comprehensibility or intelligibility judgement
Machine outcome	Versioned model score with documented failure handling
Retention	Delayed measure after the immediate practice window
Subgroups	Performance by relevant accent, first language, device, and acoustic conditions
Exposure	Completed attempts and active speaking time, not enrolment alone
Harms	False correction, frustration, exclusion, privacy, and accessibility events

This matrix connects to the oral-fluency measurement methods brief, which explains why speech rate and pauses are not substitutes for overall proficiency. The automated speech fairness map adds the subgroup and uncertainty tests required before operational use.

Implications for practice

For an individual learner, choose one observable target and use a hear–attempt–compare–retry loop. Keep a recording of the first and later untrained examples. The deliberate pronunciation practice guide provides a bounded routine, and the intelligibility goal guide helps avoid treating accent difference as an error by default.

For a school or product team, retain the raw ingredients needed to audit a claim without retaining more personal audio than necessary. Separate the learning metric from availability, engagement, and commercial conversion. Document the model and rubric version, missing-score rules, human adjudication, and the exact population for which the evidence was gathered.

Learners can use the LinguaLive speaking tools for low-stakes practice, but automated feedback should be treated as a prompt for inspection rather than an official assessment. High-stakes pronunciation decisions need qualified, independent human review.

Limitations

This map searched English-language sources quickly and may underrepresent research published in other languages or venues. Technology changes faster than many intervention studies can be replicated. “CAPT” covers substantially different systems, while studies differ in learner level, target feature, training duration, comparator, and outcome. Positive publication bias and intervention-specific tests may inflate apparent consistency.

The map does not validate LinguaLive, any named commercial tool, or a particular automated score. It does not establish that reducing accent is a desirable goal. Disability, dialect, identity, and communication context can change what fair and useful feedback looks like.

Editorial disclosure

LinguaLive develops an AI speaking-practice product and has a commercial interest in technology-mediated feedback. Vlad Podoliako is named as the article author because he is LinguaLive's founder; no academic, clinical, or assessment credential is claimed here. The source descriptions and interpretations should be independently reviewed before they support a product efficacy or institutional purchasing claim.