A field with a name: second language acquisition
Second language acquisition (SLA) is the academic field that studies how people actually learn a language beyond their first one. Researchers have been testing questions like how much exposure is enough, when review works best, and why some study methods stick while others don't since the 1970s. The four ideas below come from that field, not from a marketing team. Each one has a name, a date, and a paper behind it.
1. Comprehensible input
Linguist Stephen Krashen proposed the comprehensible input hypothesis in the late 1970s and early 1980s, as one piece of a broader theory of how people acquire a second language. His shorthand for it is i+1: "i" is a learner's current level of competence, and "+1" is the next small step just past it. Input that's mostly understandable, with a bit of unfamiliar language mixed in, is what actually moves someone forward, in his account. Drilling grammar rules in isolation teaches you to talk about the language, not to use it without stopping to think.
Krashen's model has five linked hypotheses, not just this one. Another piece, the affective filter hypothesis, holds that stress and anxiety block acquisition even when the input itself is right. That's part of why Lira leans on real texts you actually chose to read instead of graded exercises: lower pressure, and input that lands closer to i+1 because you picked something you can mostly already follow. The input hypothesis is also one of the more argued-over parts of SLA theory: "i" is hard to pin down precisely, and the hypothesis resists direct testing. What has held up is the general shape of the idea: material a bit past what you already know beats material that's too easy or too hard, which is still what reading-based methods are built around.
Krashen, S. (1985). The Input Hypothesis: Issues and Implications.
Er fühlte eine tiefe ardent longingSehnsucht nach der Heimat.
He felt a deep longing for his homeland.
2. How much vocabulary you need
Linguist Paul Nation quantified the lexical coverage needed to read without constant decoding effort: roughly 90-95% known words for reading to stay productive, 98% to feel nearly effortless. Below that threshold, an unknown word forces a conscious pause to guess or look it up, which interrupts the automatic word recognition that fluent reading actually runs on. Enough of those interruptions in one passage, and working memory fills up with decoding instead of meaning, which is what most learners experience as frustration rather than progress.
That threshold is essentially a way to measure Krashen's i+1 in practice: a text sitting in the 90-95% coverage range is, by definition, mostly understandable with a manageable amount of new language mixed in. Lira computes this coverage for real, per text and per user: a difficulty badge on every book shows whether it lands in that range, and a pre-reading banner highlights the most frequent words worth learning first, so you spend that effort on the words that pay off soonest.
Nation, P. (2006). How large a vocabulary is needed for reading and listening? Canadian Modern Language Review, 63(1), 59–82.
3. Reviewing at the right time
Every word you translate while reading gets queued into a review schedule managed by FSRS, an open-source algorithm that calculates the optimal moment for the next review, right before the word is likely forgotten. That moment comes from published research on memory, not a fixed interval picked in advance.
Spaced repetition itself isn't new: Hermann Ebbinghaus plotted the first forgetting curve back in 1885, and Piotr Woźniak's SM-2 algorithm from 1987 is still the default in tools like Anki. SM-2 applies roughly the same multiplier to every card. FSRS, published in 2022, replaced that with a model that learns separate memory parameters per word from a person's own review history, which is what gets it closer to the actual moment a specific word is about to slip rather than a one-size-fits-all schedule.
Ye, J., Su, J., & Cao, Y. (2022). A Stochastic Shortest Path Algorithm for Optimizing Spaced Repetition Scheduling. Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 4381–4390.
How FSRS works in Lira →4. Active recall over re-reading
Lira's crossword puzzles aren't just a side game: reconstructing a word from a clue requires active recall, more demanding than simple recognition. Cognitive psychology research shows active recall anchors memory more durably than passive re-reading.
The core finding goes back further than 2006: researchers were measuring a testing effect as early as 1909, and Roediger and Karpicke's study is a well-known modern replication, not the origin. Their experiment had students either restudy a passage or take a practice test on it. A week later, the group that had been tested remembered more, even though both groups had spent the same amount of time with the material.
Roediger, H. L., & Karpicke, J. D. (2006). Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention. Psychological Science, 17(3), 249–255.
None of these four mechanisms is unique to Lira. What the app does is chain them with no friction: you read a text you chose, translate a word with a tap, it joins an FSRS review queue, and you work it actively through crossword puzzles.
How this compares
Lira uses FSRS, not SM-2
Anki still ships SM-2 by default. FSRS models each word's own forgetting curve instead of a fixed interval, which published research puts at roughly 40% fewer reviews for the same retention.
Lira vs Anki →Lira calculates real coverage, not a level guess
Duolingo sorts learners into a level once, at onboarding, and mostly leaves it there. Lira recomputes lexical coverage per text and per user, every time: the difficulty badge and priority-words banner above are what that calculation looks like on screen, not a label picked once and forgotten.
Lira vs Duolingo →Lira groups word forms by root, honestly counted
"had", "hasn't", and "has" resolve to one vocabulary entry in Lira instead of inflating the count as three separate words. That inflated counting is a known complaint about how some reading apps report progress.
Lira vs LingQ →