SaunaSpeak uses learning methods that memory and language researchers have tested for decades. Here is each one: what it does in the app, what the studies found, and a link to every paper so you can read it yourself. Where the evidence is weak or mixed, we say so.
No study has tested SaunaSpeak itself. The studies test the methods, and we use the methods.
How strong is the evidence: Strong evidence.
In SaunaSpeak: A new sentence comes back after one day, then after about 3, 14 and 30 days, and the gaps keep growing after that. If you get it wrong, it comes back right away and climbs again. Sentences you find easy wait longer.
What the research found: Spreading practice over several days works better than doing the same amount in one sitting. Cepeda and colleagues reviewed 317 experiments on remembering words and facts, and spaced study led to better memory than study done all at once. A 2022 meta-analysis (a study that combines the results of many studies) looked only at second language learning and found a medium to large benefit from spacing. Longer gaps did better on tests taken weeks later, and growing gaps worked about as well as equal ones.
How strong is the evidence: Strong evidence.
In SaunaSpeak: Most cards ask you to produce the Finnish yourself. You say a sentence from memory when you see the English, or write it down when you only hear it. Checkpoints, word reviews and mistake cards work the same way: you try first, then you see the answer.
What the research found: Pulling an answer out of memory makes it stick better than reading it again. In Roediger and Karpicke (2006), students either re-read a text or took a test on it. Re-reading did a little better on a test 5 minutes later, but the tested students remembered more after 2 days and after 1 week. In a 2008 study in Science, students learned Swahili words. Those who kept being tested on words they had already got right remembered far more a week later than those who only studied those words again.
How strong is the evidence: Good evidence, from lab studies.
In SaunaSpeak: When a sentence is new, you hear it and type what you think it means before the English is shown. The new words in it are taught one at a time first, so the guess is a fair one.
What the research found: Trying to answer before you know the answer, even when the try is wrong, can help you learn the answer better than studying it for the same amount of time. Kornell, Hays and Bjork (2009) found this with word pairs and trivia questions, and Richland, Kornell and Kao (2009) with science texts. These were studies with university students, not with a language app, so take them as good support for the idea, not proof for our exact card.
How strong is the evidence: Good evidence for the parts.
In SaunaSpeak: Each sentence moves up a ladder as you learn it. First you work out what it means, then you build it from word tiles, then you say it from memory, and at the top you write it down just from hearing it. If you miss it, it drops back a step.
What the research found: People remember words better when they produce them than when they only read them. This is called the generation effect, first described by Slamecka and Graf in 1978. A large review by Dunlosky and colleagues (2013) compared ten common study methods. It rated testing yourself and spacing out your practice as the two most useful, and re-reading and highlighting as much less useful. The ladder itself is our design. The research supports its parts, producing the answer and testing yourself, not this exact order of steps.
How strong is the evidence: Good evidence, one part unclear.
In SaunaSpeak: Kuulo plays one word and you choose which of two words you heard, for example kuu (moon) or kyy (adder). You find out straight away if you were right. We picked these vowel pairs because mixing them up gives you a different word, and because of vowel harmony it changes the endings too.
What the research found: This kind of drill is called high variability phonetic training: hear a word, choose, get feedback, with the words spoken by several different voices. It started with a 1991 study that trained Japanese listeners to hear the difference between English r and l. Thomson's 2018 review of 32 studies concluded that it is an effective way to help learners hear new sounds, and often to say them better too. A large replication in 2022 found that the training helped, but it did not find a clear extra benefit from using many voices instead of one. So the listen, choose and get feedback part is well supported, and the many voices part is less certain. Kuulo uses human recordings where we have them and synthetic voices for the rest.
How strong is the evidence: Some evidence.
In SaunaSpeak: The Kuuntelu scenes are two Finns talking at normal speed. On the first listen the text is hidden and you answer a few questions. You can slow the audio down, replay one line, and open the Finnish and then the English whenever you want. On sentence cards you can also shadow: record yourself copying the native audio, then play the two one after the other.
What the research found: Hiding the text on the first listen is our choice, not something a study proved. We do it because a real conversation gives you no text, so that is the skill to practise. Reading along does help, though. A 2013 meta-analysis found that video with captions in the language being learned helped learners understand more and learn more words than video without captions. That is why the text is always one tap away. Shadowing means repeating speech out loud right after you hear it. Hamada (2016) found that it helped Japanese learners of English hear sounds better, and that the lower-level learners also improved their listening test scores. That was one study with 43 students, so it is a small piece of evidence.
How strong is the evidence: Good evidence, from classrooms.
In SaunaSpeak: Each Situation is a task with a goal, like ordering coffee or meeting your new neighbor. You talk with an AI character in spoken Finnish until the task is done.
What the research found: Task-based language teaching builds lessons around real tasks instead of grammar lists. A 2019 meta-analysis of 52 studies found a strong positive effect on learning. A later re-analysis of that research, with stricter rules for which studies count, found a smaller effect that was still positive. Both looked at courses with human teachers, not AI characters. They support the idea of learning by doing tasks, not our app in particular.
How strong is the evidence: Good evidence, from human teachers.
In SaunaSpeak: When you write something in Sauna Chat or a Situation, you get your own sentence back, fixed. Each correction is saved as a review card, so later you produce the right form from memory instead of just reading it once. You can turn this off.
What the research found: Li (2010) combined 33 studies and found that correcting learners' mistakes had a medium effect on learning, and that the effect was still there on later tests. In those studies, teachers and researchers gave the corrections. An AI can also be wrong, which we say openly on the About page.
How strong is the evidence: Some evidence, weaker alone.
In SaunaSpeak: You say a sentence out loud and get a score for the whole sentence and for each word, from Microsoft's Azure Speech service.
What the research found: Ngo, Chen and Lai combined 15 studies of speech recognition tools for learners of English and found a medium benefit for pronunciation overall. The benefit was bigger when the tool clearly showed what was wrong, and for single sounds more than for rhythm and intonation. It was smaller when people practised alone, and most practice in an app is done alone. So treat the score as a guide, not a grade.
How strong is the evidence: Good evidence, with caveats.
In SaunaSpeak: The free decoder takes a line of spoken Finnish, shows the written form, and names the rule behind each change, for example that the question ending -ko becomes -ks, so onko becomes onks. Every form has its own page in the list of spoken forms.
What the research found: Norris and Ortega (2000) combined studies of language teaching and found that explicit teaching, where the rule is pointed out, led to bigger gains than teaching where learners had to find the pattern alone. They also noted that many of the tests in those studies favoured explicit knowledge, so the real difference may be smaller. Schmidt (1990) argued that learners have to notice a form before they can learn it. That is a theory: often cited, and still debated.
How strong is the evidence: Good evidence, outside language learning.
In SaunaSpeak: When you sign up, we ask when you will practise: morning, lunch or evening. Your reminder email and app reminder come at that time. If your streak is at risk you get at most one reminder a day, and never one about a streak that is already lost.
What the research found: Deciding exactly when you will do something is called an implementation intention. Gollwitzer and Sheeran (2006) combined 94 tests and found a medium to large effect on people reaching their goals. Lally and colleagues (2010) followed 96 people building a new daily habit. It took between 18 and 254 days for the habit to become automatic, and missing a single day did not really set it back. That is one reason SaunaSpeak does not send guilt messages about a missed day.
Spoken Finnish first. No study tells you which kind of Finnish to learn first. We teach puhekieli, the spoken Finnish you hear from Finns every day, and show the written form next to it. That is a judgement about what learners need, not a research result. Why spoken Finnish first.
Mixing reviews in with new sentences. A session spreads your due reviews between the new sentences. We do this to keep spacing working and the session varied, but we do not claim the mixing helps by itself. A 2019 meta-analysis found that mixing topics helps most when the things being learned are hard to tell apart, and it advised caution for learning words.
If you know a study we missed, or think we read one wrong, write to mail@saunaspeak.com. We will fix this page.