What 152,218 Listen and Repeat attempts taught us

What we wanted to understand

We kept the questions deliberately narrow. The data can show recurring patterns in LingoLeap practice; it cannot, by itself, prove why they happen.

• What changes when a sentence gets longer?
• Which sounds and word endings repeatedly receive low pronunciation-assessment results?
• What happens to the platform score as a learner pauses more often?

The setup

Pattern 1: Longer sentences lose more of the prompt

This was the clearest pattern in the dataset. Reproduction falls from 78.8% on the shortest sentences to 50.0% on the longest, and the score falls with it.

What surprised us is what did not change. Average pronunciation accuracy stays in a narrow range, from 84.8 to 84.5, while reproduction changes by 28.8 percentage points. In other words, learners do not appear to pronounce every surviving word dramatically worse; they simply bring back less of the sentence.

Pattern 2: A few sounds are much harder than the rest

/θ/ — the “th” in thin — is the lowest-scoring sound at 61.6/100; 37.6% of its 17,575 measured occurrences fall below the provider’s 60-point threshold.

The gap is large enough to be useful, but it is easy to over-interpret. These are automatic-assessment outputs, not phonetician judgements. Microsoft defines AccuracyScore as acoustic similarity to a native pronunciation model and treats word scores below 60 as mispronunciations [3].

Pattern 3: The word survives, but the ending disappears

Word-final -s is dropped on 11.4% of the 237,757 words that carry it.

This is the kind of error that is easy to miss by ear. The main word is still there, so the repetition feels correct, but a small grammatical signal has vanished. Even when the ending survives, the closing sibilant averages only 74.3/100 — audible but weakly formed. Endings are acoustically brief and can be difficult to preserve under time pressure. The present data do not identify whether memory load, articulation, or recognition error caused an individual omission. The loss is meaning-bearing: dropping the -s turns closes into close and weekends into weekend; the same happens to past-tense -ed and to numbers and times.

Pattern 4: Every additional pause comes with a lower score

Score falls monotonically with each added pause — from 3.68/5 for the cleanest deliveries to 2.22/5 at five or more breaks.

The nearly step-by-step decline looks compelling, but it is not a pause penalty estimate. Lower-proficiency or less-complete attempts may both pause more and score lower.

Pattern 5: A small set of words repeatedly scores low

Content words carrying the hardest sounds and clusters score lowest — for example turned (41/100) and loaner (44/100).

What might connect these patterns

The findings describe two observable dimensions: reproduction and pause behaviour on one side, and automatic pronunciation assessment on the other. Working-memory research makes processing load a plausible interpretation of the length pattern [1], while language-assessment research cautions that automated speech scores require construct-level validation [4]. These mechanisms remain hypotheses here; the dataset establishes patterns, not their causes.

For learners, the descriptive results still identify concrete diagnostic targets: compare short and long prompts, inspect whether the sentence ending survives, and review low-scoring phonemes. The companion method guide turns those observations into a practice protocol, but its recommendations should not be read as an experimentally estimated treatment effect.

From aggregate patterns to real sentences

Every figure above is an average over thousands of attempts. Each individual Listen and Repeat sentence also has its own page — the real target sentence, how it splits into chunks, which specific words are dropped, and where speakers pause. The concrete case behind the aggregate:

Closing thoughts

Dataset note

Notes and sources

1. [1] Baddeley, A., Gathercole, S., & Papagno, C. (1998). The phonological loop as a language learning device. Psychological Review, 105(1), 158–173.
2. [2] Saito, K. (2019). Effects of second language pronunciation teaching revisited. Language Learning, 69(3), 652–708.
3. [3] Microsoft. (2026). Use pronunciation assessment. Microsoft Learn.
4. [4] Xi, X. (2023). Examining the validity of automated speech scoring systems. Journal of the Acoustical Society of Japan, 79(3), 170–176.

Questions we hear from learners