Same-language captions and translated subtitles solve different problems. This guide covers listening and vocabulary in a second language (L2). It does not measure deaf or hard-of-hearing accessibility outcomes, product-task success or sales conversion.
The findings and their context
Montero Perez et al. (2013) report g = 0.988 — Listening comprehension with same-language captions. Population: Second-language learners; 15 study samples. Evidence: Table 3, p. 728; section 4.1.1, pp. 728–729; 95% CI [0.598, 1.377]. System, 41(3), 720–739.
Video transcript
When you listen in another language, a written caption can help you follow the sounds. Here the words light up as they are spoken.
In the 2013 review, fifteen study samples gave an average effect of g equals zero point nine eight eight for listening comprehension.
This is a standardised score difference. One unusually large result affected the average, so the size of the benefit varies across studies.
Montero Perez et al. (2013) report g = 0.866 — Immediate vocabulary posttest with same-language captions. Population: Second-language learners; 10 studies. Evidence: Table 4 and section 4.2.1, p. 730; 95% CI [0.578, 1.154]. System, 41(3), 720–739.
Video transcript
Hearing a word while seeing its spelling gives you two clues. The sound and the letters can point to the same object.
Ten studies in the 2013 review gave an average vocabulary effect of g equals zero point eight six six with same-language captions.
These were immediate tests. The result describes scores shortly after viewing, and the review does not establish permanent mastery of those words.
Reynolds et al. (2022) report d = 0.884 — Vocabulary learning across caption and subtitle modes. Population: Language learners; 20 studies in the meta-analysis, not all 34 scoped studies. Evidence: Table 6, pp. 10–11; random-effects column; 95% CI [0.412, 1.356]. Systems, 10(5), 133.
Video transcript
A caption can repeat the spoken language, while a subtitle can translate it. Those options give the learner different kinds of help.
Reynolds and colleagues combined twenty studies. Their random-effects estimate for vocabulary was d equals zero point eight eight four.
The review mixes subtitle formats and its results vary widely. This average cannot choose the best subtitle language for a particular viewer.
Kurokawa et al. (2024) report g = 0.56 — Incidental vocabulary learning with L2 captions. Population: Second-language learners; 89 effect sizes from 49 experiments. Evidence: Results: Overall Effectiveness of Captioning; Figure 3; 95% CI [0.38, 0.73]. Language Learning, 75(4), 939–987 (2025 issue; first published online December 10, 2024).
Video transcript
You watch a story to understand what happens. Along the way, an unfamiliar word appears in a caption and becomes easier to notice.
Kurokawa and colleagues report g equals zero point five six across forty-nine experiments on incidental vocabulary learning.
Viewers were not instructed to memorise vocabulary. This score difference compares captioned and uncaptioned groups, and some experiments also appear in earlier reviews.
Kurokawa et al. (2024) report g = 0.14 — Uncertain lesson/lecture subgroup estimate. Population: L1-targeted lesson/lecture videos for L2 learners; 7 effect sizes from 6 experiments. Evidence: Table 5, Video genre (L1-targeted only), Lesson/lecture style; 95% CI [−0.40, 0.68]; p = .606; between-genre moderator p = .520. Language Learning, 75(4), 939–987 (2025 issue; first published online December 10, 2024).
Video transcript
A smaller group of experiments used lessons or lectures. With fewer observations, the estimated effect is less precise.
For that subgroup, the review reports g equals zero point one four. The confidence interval runs from minus zero point four zero to zero point six eight.
The interval crosses zero. These results leave uncertainty about the vocabulary benefit and do not establish a difference between lectures and the other video genres.
Read the estimates without overclaiming
Five rows are not five independent replications. The first two share a review; the last is nested within the fourth. Reviews can reuse underlying studies. Do not add their counts or effects. g and d remain standardised units, never percentages.
The 2013 listening result is sensitive to an outlier: section 4.1.1 reports g = 0.76 after its removal. Reynolds reports high heterogeneity (I² = 94.783%). Models, learners and tests differ. Kurokawa excludes intentional vocabulary learning; comparing pooled averages across reviews does not establish a historical trend.
An editorial test for a multilingual demo
Offer an accurate, synchronised caption track and label translated tracks separately. Then check whether your audience understands product terms and completes a task. This is editorial design guidance, not a DemoBro outcome measured in these studies. Accessibility needs require separate evaluation.
Related evidence: visual cueing and demo-video learning.
Sources, verification and limitations
This curated guide is not an exhaustive review or a new meta-analysis. All three public full texts were checked on November 16, 2026: the KU Leuven author copy, the UCL version of record and Wiley publisher text. None is abstract-only or a preprint. Original bibliographic titles are retained.
The original paper years remain 2013 and 2022. Kurokawa uses the first online year, 2024 (December 10); the final issue, 75(4), pages 939–987, is dated 2025. No customer records, internal analytics, catalogue measurements or private experiments are used.