LingoLeap Research
TOEFL research from real practice data
Every day, thousands of learners complete TOEFL speaking and writing practice on LingoLeap and receive AI scores. We turn that anonymised, aggregated data into descriptive studies that answer questions a small sample cannot — which sounds are hardest, where sentences start to break, how first language shapes pronunciation. Every study publishes aggregates only: no recordings, transcripts, or identities.
Academic Discussion · Rubric
Where Academic Discussion answers lose points: language, not ideas
In 3,692 AI-scored answers, Language Use & Grammar was the lowest of the three rubric sub-scores, or tied for lowest, 93.8% of the time. LingoLeap's AI grader, not ETS.
93.8%
Language Use lowest or tied
Academic Discussion · Length
Does length raise your Academic Discussion score?
Only 1.3% of 3,692 answers ran under 100 words. Scores rise with length up to about 175-199 words, then level off. Observational.
1.3%
answers under 100 words
Academic Discussion · Engagement
Does naming a classmate help your Academic Discussion score?
47.3% of answers name a classmate from the prompt; they score +0.67 points higher on 0-30. A small, observational difference.
47.3%
name a classmate
Academic Discussion · AI scoring
Do templates hurt your TOEFL score? 3,692 answers measured
A template opener is associated with +0.04 points on the 0-30 scale (95% CI -0.12 to +0.20), with 80% power to detect 0.23. Language Use is the flattest of the three sub-scores. LingoLeap's AI grader, not ETS's e-rater.
+0.04
points on 0-30, 95% CI -0.12 to +0.20
Take an Interview · Speaking scoring
How TOEFL Speaking is scored: what the data shows
Across 54,247 Take an Interview answers in the reported 2.0–4.5 AI score bands, Delivery, Language Use, and Topic Development move upward together. The pattern is observational, not causal.
54,247
answers in reported bands
Take an Interview · Delivery
What separates TOEFL Speaking score bands: Delivery
Fluency, pace, and length are higher in higher AI score bands. Bucket averages peak around 140–159 wpm, while the zero-pause bucket is confounded by short and truncated responses.
140–159
highest-average wpm bucket
Take an Interview · Language Use
TOEFL Speaking Language Use: vocabulary vs sentence control
Grammar/vocabulary control and complex-sentence share are higher in higher bands, but the share of advanced (B2+) vocabulary stays low (4–7%) and roughly flat — rare words aren't what distinguishes bands here.
4–7%
advanced-vocab share, all bands
Take an Interview · First-language comparison
Do Chinese learners score differently on TOEFL Speaking interviews?
Chinese-interface and English-interface learners score almost identically across Delivery, Language Use, and Topic Development, with near-identical score distributions — no first-language gap detectable under this (weak) proxy.
≈ parity
across every dimension
Listen and Repeat · Pronunciation
Which English sounds are hardest for Chinese TOEFL learners?
Chinese- and English-interface learners score almost identically on every sound — except /θ/, the single hardest sound for Chinese-first-language learners at 63.5/100.
/θ/ 63.5
hardest sound, Chinese L1
Listen and Repeat · Speaking
What 152,218 Listen and Repeat attempts taught us
A descriptive study of sentence length, pauses, word endings, and phoneme-level pronunciation across 23,390 real TOEFL 2026 practice sessions.
152,218
scored sentence attempts
Data, citations and handouts
Every study publishes its aggregate tables as CSV under CC BY 4.0, a cite block in APA 7 and MLA 9, embeddable charts and a one-page teacher handout.
- What 152,218 TOEFL Listen and Repeat Attempts Taught Us
listen-and-repeat-sentence-length.csvlisten-and-repeat-phonemes.csvlisten-and-repeat-pauses.csvlisten-and-repeat-hardest-words.csvlisten-and-repeat-summary.csvTeacher handout (PDF)How to cite
- Which English Sounds Are Hardest for Chinese TOEFL Learners?
chinese-learners-phoneme-compare.csvchinese-learners-sentence-length.csvTeacher handout (PDF)How to cite
- How TOEFL Speaking Is Scored: What the Data Shows
toefl-speaking-by-score-band.csvTeacher handout (PDF)How to cite
- What Separates TOEFL Speaking Score Bands: Delivery
toefl-speaking-by-score-band.csvtoefl-speaking-fluency-speech-rate.csvtoefl-speaking-fluency-pauses.csvtoefl-speaking-fluency-response-length.csvTeacher handout (PDF)How to cite
- Do TOEFL Templates Hurt Your Score? 3,692 Answers Measured
templates-phrase-prevalence.csvtemplates-score-difference.csvtemplates-subscores.csvtemplates-openers.csvTeacher handout (PDF)How to cite
- Does Length Raise TOEFL Academic Discussion Scores? Data
academic-discussion-length.csvTeacher handout (PDF)How to cite
- Where TOEFL Academic Discussion Answers Lose Points: Data
academic-discussion-subscores.csvTeacher handout (PDF)How to cite
- Does Naming a Classmate Help on TOEFL Academic Discussion?
academic-discussion-classmates.csvTeacher handout (PDF)How to cite
Licence: Aggregate numbers only, under CC BY 4.0: reuse them with credit to LingoLeap and a link to this page. No learner answers, recordings or individual scores are published.
How we work
- Data comes from real, voluntary TOEFL practice on LingoLeap, anonymised and aggregated.
- Pronunciation is scored by automatic speech assessment; sentence scores are LingoLeap AI scores, not official ETS results.
- We publish aggregate statistics and charts only — never recordings, transcripts, user IDs, or individual responses.
- Studies are observational and descriptive: we report denominators and differences, not causal effects.
Want to put the research to work?
Practise TOEFL speaking on LingoLeap and get the same AI scoring and diagnostics.