Research
What we build on, what we measure, and what we have not measured yet.
YNova is a product, not a paper. But every rule in it has a reason, and every claim we make in public should have a study behind it. This page lists the reasons, the studies, and the gaps, in that order.
The pedagogy
The first language is a scaffold, not the destination.
NOVA explains in the learner's own language, at a level the learner sets. When and how that support should fade is one of the studies below. The position itself comes from here.
- 1
García, O. & Li Wei (2014). Translanguaging: Language, Bilingualism and Education. Palgrave Macmillan.
Why the first language is a resource in learning a second, not an obstacle.
- 2
García, O., Johnson, S. I. & Seltzer, K. (2017). The Translanguaging Classroom. Caslon.
Stance, design and shift: the three moves NOVA's bilingual scaffolding is built around.
- 3
Cenoz, J. & Gorter, D. (2020). Teaching English through pedagogical translanguaging. System, 92.
Pedagogical translanguaging in English teaching specifically.
- 4
Cummins, J. (2007). Rethinking monolingual instructional strategies in multilingual classrooms. Canadian Journal of Applied Linguistics, 10(2).
The case against English-only instruction.
- 5
Tian, L. & Macaro, E. (2012). Comparing the effect of teacher codeswitching with English-only explanations on the vocabulary acquisition of Chinese university students. Language Teaching Research, 16(3).
Evidence that a first-language gloss helps vocabulary learning.
- 6
Zhao, T. & Macaro, E. (2016). What works better for the learning of concrete and abstract words: teachers' L1 use or L2-only explanations? International Journal of Applied Linguistics, 26(1).
When a first-language explanation helps most.
We wrote about this in plain language: Why YNOVA lets learners think beyond English. The technical follow-up is a two-part series: Introducing Vāc-Jiàn and the evidence and engineering behind it.
Practice and feedback
Why the loop is shaped the way it is.
Test first, prompt a self-correction before correcting, bring things back when they are about to be forgotten. Each of those is a finding before it is a feature.
- 7
Wang & Zhang (2026). Meta-analysis of AI-supported language learning across 40 studies. Language Teaching.
The pooled effect for speaking is the largest of the skills studied (g = 1.36). We treat it as encouraging, not as proof about NOVA.
- 8
VanLehn, K. (2011). The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational Psychologist, 46(4).
What tutoring systems can and cannot match in a human tutor.
- 9
Kulik, J. A. & Fletcher, J. D. (2016). Effectiveness of intelligent tutoring systems: a meta-analytic review. Review of Educational Research, 86(1).
Baseline expectations for tutoring software.
- 10
Lyster, R. & Saito, K. (2010). Oral feedback in classroom SLA: a meta-analysis. Studies in Second Language Acquisition, 32(2).
Why NOVA prompts a self-correction before it gives the correction.
- 11
Li, S. (2010). The effectiveness of corrective feedback in SLA: a meta-analysis. Language Learning, 60(2).
How much, and what kind of, corrective feedback helps.
- 12
Kim, S. K. & Webb, S. (2022). The effects of spaced practice on second language learning: a meta-analysis. Language Learning, 72(1).
The spaced review that brings back what you got wrong.
- 13
Adesope, O. O., Trevisan, D. A. & Sundararajan, N. (2017). Rethinking the use of tests: a meta-analysis of practice testing. Review of Educational Research, 87(3).
Why the training loop tests before it teaches.
Speech assessment
Pronunciation is an AI estimate, and we say so.
We evaluated open-source pronunciation models in August 2026 and chose one to run alongside the tutor. Its band is shown as an AI estimate, separately from the other criteria, because we have not yet validated it on our own learners.
- 14
Chen, Y.-W. et al. (2024). MultiPA: a multi-task speech pronunciation assessment model for open response scenarios. Interspeech 2024.
The open-source model behind NOVA's AI pronunciation estimate. The published correlations with human raters (about 0.73 overall on the speechocean762 set) are the authors' results, not ours.
How each criterion is scored, and which signals are measured rather than estimated, is on the scoring page.
Studies we owe
Planned, designed, not yet run.
These are the studies that would let us make stronger claims. We list them so you can hold us to them. Results will be published here whichever way they go.
Calibration of NOVA's estimates against teacher marks
Not startedPairs of NOVA estimates and independent teacher marks on the same recordings. Today the count of such pairs is zero. Until this study reports, we do not describe the estimates as accurate, only as labelled.
Examiner-rated reference set
PlannedTwenty to fifty sessions rated by a qualified IELTS examiner, used as the fixed reference every scoring change is checked against before it ships.
Pronunciation validity study
PlannedThree to five hundred adult attempts, each with at least two independent qualified pronunciation ratings, to test whether the AI estimate tracks human judgement on our learners rather than on a benchmark set.
Does bilingual scaffolding help, for whom, and when should it fade?
DesignedA within-subject crossover with fifteen to twenty learners over six weeks. Directional only; a pilot-scale A/B test would be powered near ten percent, so we will not run one and call it evidence.
Collaborate
If you study language assessment or bilingual pedagogy, we want to hear from you.
We can offer a live product used by real learners and teachers, consented and anonymised data under a written agreement, and the engineering time to instrument a study properly. We ask for rigour and for the right to publish the result.
- Co-design one of the studies above, or propose one.
- Rate a reference set as a qualified IELTS examiner.
- Replicate our bilingual-hint findings in your own classroom.
contact@ynova.ai
See the rules in the product.
Every principle above is visible in a single Part 1 session: the labelled estimate, the hint in your language, the correction that waits for you first.