This week the webapp went from 16 languages to 28. The number is not the interesting part. The interesting part is which languages — and why almost nobody else offers them.
The new twelve: Ukrainian, Russian, Polish, Czech, Hungarian, Dutch, Turkish, Uzbek, Georgian, Swahili, Basque, Galician, and Esperanto. A few of those you can study anywhere. Most of them you cannot. There is no major app with a serious Basque course. Galician, Georgian, Uzbek — you will find flashcard decks and a grammar PDF, and then the trail goes cold.
It is not for lack of demand — it is unit economics. A conventional language course is manufactured content: scripts written, voice actors hired, studio time booked, curriculum designed, every sentence recorded twice because the first take clipped. That cost is roughly fixed per language, so each language must clear a revenue bar to justify existing. Spanish clears it. French clears it. Basque, with maybe 750,000 speakers, does not clear anyone's bar, and so the market simply never builds it. The languages that miss the cut are not obscure — Swahili is the working language of something like 200 million people across East Africa, the language trade actually happens in. It misses the cut anyway.
For years, Mozilla's Common Voice project has been quietly collecting something extraordinary: ordinary people reading sentences aloud in their own languages, validated by other speakers, released openly. It was built for training speech-recognition machines. But a native speaker reading a real sentence is exactly what a human learner needs to train on, too.
And here is the thing the download numbers taught us: small language communities record with a zeal the big languages never need. Basque volunteers have validated more clips than Russian volunteers — a language of under a million speakers out-recording one of 145 million. Galicia, home to 2.4 million people, produced 232,000 distinct recorded sentences. And Esperanto produced the largest audio corpus we have ever shipped: 854,360 clips.
A language with no country, no army, and no government produced the largest audio corpus we have ever handled. That is what people do for a language they love.
Honestly? Selection and plumbing. We deduplicated the recordings so a thousand readings of one sentence don't masquerade as a thousand sentences, machine-glossed every unique sentence into English, staged the audio, and put our grading and spaced-repetition scheduling on top. Twelve languages took one long evening and about seventy dollars — most of it machine translation. Not because we are clever, but because the hard, human part was already done by volunteers over years.
Which is really the whole position: the recordings exist, they are open, and letting them sit unused while learners hunt for a decent Georgian course would be a waste of other people's love. We refuse to waste it.
The same product we run for Romanian and English: a native speaker says a real sentence at real speed, you type what you heard, the diff shows you exactly which letters you missed, and the scheduler brings back what you got wrong at the interval your memory needs. No cartoon owl, no streaks, no easy mode — the difficulty is the product. If you are the person who has been telling friends "there's just no good app for Basque" — there is now, and it starts at native speed.