Why we're building lingi, and how it's different

lingi is an app for Polish speakers learning Spanish from Spain. It differs from the popular apps in four ways: it schedules reviews with the FSRS algorithm, asks you to recall each answer before showing it, teaches the most common words first, and explains everything in Polish or English.

Language apps are very good at getting you to open them every day. That matters. We wanted something else on top of it. We wanted every minute in the app to go on what research says makes words stick, and we wanted Polish speakers to learn Spanish without routing everything through English.

That's lingi. Here's the research we lean on, what we took from it, and where the limits are.

Memory fades on a curve, so reviews should follow it

In 1885 Hermann Ebbinghaus memorised lists of nonsense syllables and measured how much effort he saved when he relearned them. After a day about a third of the saving was left, after a month about a fifth [1]. People often get this wrong. He measured savings on relearning, not how much he could recall. Murre and Dros repeated the experiment in 2015 and got a very similar curve [2].

The more useful finding is what happens when you spread reviews out. Cepeda and colleagues pooled 839 assessments from 317 experiments. Spaced practice beat massed practice, and the best gap between sessions grew with how long you need to remember [3]. In a later study with more than 1,350 people, the best gap produced 64% better recall than reviewing on the same day [4].

The catch is that spacing feels worse than it is. In Kornell's flashcard study, spacing worked better for 90% of participants, yet after the first session 72% of them believed cramming had worked better [5].

What we do with it. lingi schedules reviews with FSRS. For each card it estimates how well you remember the word from your ratings, and brings it back around the point where the chance you'd recall it drops to 90%. Prefer fewer reviews? Set it to 85%. FSRS is open source and has a public benchmark built on about 727 million reviews from 10,000 Anki users. There, FSRS-6 predicts recall better than older models, including the one Duolingo published in 2016 (log loss 0.346 against 0.469, lower is better) [6][7]. We don't know what Duolingo runs today, so that comparison only covers the published model. For now we use the default FSRS parameters rather than fitting them to each learner's history.

Recalling teaches more than rereading

In Roediger and Karpicke's study, students read a passage and then either read it again or tried to recall it. Five minutes later, rereading came out ahead. A week later it flipped: 56% against 42% [8].

Their 2008 study with 40 Swahili-English word pairs is even starker. When students kept recalling words they'd already got right once, they remembered about 80% a week later. When those words dropped out of testing, they kept 36% and 33%. More study after the first successful recall had no measurable effect. And everyone, whatever the method, predicted they'd remember about half [9].

In lingi. Every lingi card asks you for the answer before it shows it. On Say cards you say the word out loud before you reveal it. If you'd rather type, turn on typed answers and the app tells you when an accent or an article is missing. Then you rate how it went, and that rating decides when the card returns.

The words you'll actually hear, first

Words aren't equally useful. Paul Nation worked out that to follow a text without help you need to know about 98% of its words. For English that means 8,000 to 9,000 word families for reading and 6,000 to 7,000 for listening [10]. Those numbers are for English, but the lesson carries over: order matters, because common words pay off straight away.

How lingi handles it. The lingi course runs 95 lessons from A1 to C1, about 7,900 words and phrases in total. Inside each lesson the most used words come first. The order follows estimated spoken frequency, checked against the Instituto Cervantes level inventories (PCIC). We don't have our own corpus, so it's an estimate, not a measurement. If you already know some Spanish, a placement test of at most 30 words, about 2 minutes, suggests where to start. Some of its words are made up. That's Meara and Buxton's method for catching people who tick "I know it" too generously [11].

Words in sentences, phrases as a whole

A one-word translation is only the start of knowing a word. Stuart Webb found that ten encounters with a word in context can produce sizeable gains, but even ten may not be enough for full knowledge [12]. Much of language is ready-made. Erman and Warren estimated that around 55% of the English texts they studied was prefabricated [13]. In a small study by Boers and colleagues, learners who used more of these set phrases got higher ratings for their speaking [14].

In the app. Every word has an example sentence with audio. Gap cards show the base form as a hint, as in Mi amigo ____ (ser) polaco, and the line under the sentence names the form: es = ser, 3rd person singular. From B1 on, more and more cards are whole phrases like darse cuenta de or a pesar de.

Listening: lots of input you understand, but not only that

Stephen Krashen argued that we acquire a language by understanding messages slightly above our level, and that speaking emerges on its own [15]. The idea shaped a lot of teaching, and it has serious critics. Barry McLaughlin argued it can't be tested, because it never defines what comprehensible input is [16]. Merrill Swain, drawing on Canadian French immersion students, argued that understanding isn't enough and learners also need to speak and write [17]. We take both sides: plenty of listening, and practice producing the language too.

Our take. Nearly every card plays Castilian Spanish, with a slower version too. The voices are synthetic, not recorded actors. If you study with the sound off, lingi keeps track of the words you learned without audio and brings them back for a listen later. We're also testing a listening lab with an audio series about Ola from Kraków moving to Madrid.

Explanations in Polish

Translating into your first language is often treated as a shortcut to avoid. The research disagrees. In Laufer and Shmueli's study, words glossed in the learners' first language were always remembered better than words glossed in the target language [18]. A meta-analysis by Yanagisawa, Webb and Uchihara, covering 42 studies and 3,802 learners, found the same: first-language glosses beat second-language glosses, at every proficiency level [19]. Lee and Macaro found that links to the first language helped both children and adults, children more so [20].

What changes for you. Translations, hints and grammar notes in lingi are in Polish. If you'd rather have English, switch whenever you like and keep your progress.

Short, daily, and no prizes for showing up

In Baddeley and Longman's classic study, postal workers learning to type made the fastest progress with one hour a day and the slowest with two two-hour sessions [21]. In fairness, that was an hour, not five minutes, and that group needed more weeks on the calendar. The study says spreading practice out pays off. It doesn't say any dose is enough.

Deci, Koestner and Ryan's meta-analysis of 128 studies found that rewards for simply completing a task reduced people's later wish to do it for its own sake (d = −0.36), while positive feedback increased it [22]. These were lab studies, not apps, but the direction is clear.

What we do with it. You pick a daily goal of 5, 10, 15 or 20 minutes. We count a streak of study days, and you get one freeze a week for a bad day. There are no points, badges or leaderboards. After a session you see numbers you can check: words known, minutes, retention and the review forecast. Miss a few days and the backlog is spread over the next few instead of landing all at once.

How this compares with the big apps

The big apps use research too, and Duolingo publishes a lot of its own. In 2016 it described a review model that lifted daily engagement by 12% [7]. A 2021 study, mostly by Duolingo staff, found that people who finished the beginning Spanish or French course read at ACTFL Intermediate Low and listened at Novice High [23]. An independent review of 35 papers on Duolingo notes that most focus on app design and use non-probability samples [24].

Gamification helps a lot of people show up every day, and we don't think it's bad. What we change is what happens in those few minutes: an open review algorithm, recall before the answer, common words first, and explanations in your own language. We don't have learner outcome data of our own yet. We'll publish it when we do.

What lingi won't do

Flashcards build vocabulary. They won't make you fluent unless you also listen and speak a lot, and the app says so next to every level estimate. We have prototypes of an AI voice tutor and pronunciation drills, and they'll go in only when they actually help. We don't store audio of tutor conversations.

If you'd like to try it, join the waitlist. We'll email you once, when lingi launches.