Über

What a thousand words gets you

The most frequent words cover an enormous share of running speech — but comprehension needs about 95 % coverage, and unassisted reading about 98 %. This course tops out at 66 % with 241 words.

So: this builds the skeleton, and the coverage number on the home screen is measured rather than promised. It is not a route to fluency and does not pretend to be one.

Method

Retrieval practice

Nothing is ever shown to be read. Every contact with a word is an attempt to recall it.

Spaced repetition

FSRS schedules each skill separately. You never see an interval — the scheduling is the app's problem, not yours.

Desirable difficulties

The same word escalates: recognise it, hear it after 1 day of stability, say it after 2, use it in a sentence after 3, produce it cold after 10.

Production out loud

Speaking cards ask you to say the word into the microphone and grade what comes back. A course that never hears you teaches reading.

Interleaving

No two consecutive exercises share a word or a skill. It feels worse and works better.

Comprehensible input

Sentences are only scheduled once every word in them is known. The build fails otherwise, so the guarantee is structural.

Gender from day one

A noun is never shown without its article, and gender carries a shape as well as a colour. Wrong articles fossilise; nothing repairs them later.

Ordering

Words are not taught in frequency order. At each step the compiler picks the word that turns the most sentences from "one word missing" into "fully readable", breaking ties by corpus frequency. Pure frequency order leaves you unable to read a single whole sentence for a very long time.

Sources

• OpenSubtitles 2018 (hermitdave/FrequencyWords) CC BY-SA 4.0 · 151,705,378 tokens, 50,000 distinct forms
• Lexicon, sentences, texts Original work · CC BY-SA 4.0
• Pronunciation audio — Wikimedia Commons German Wiktionary pronunciation project · CC BY-SA / CC0, per clip
• Archivo · JetBrains Mono SIL Open Font License

Course content is CC BY-SA 4.0 deliberately: it keeps the door open to Wiktionary as the upstream source for gender, plural and conjugation data. The application code is MIT.

Audio

Every word is a human recording from the German Wiktionary pronunciation project, via Wikimedia Commons. Each clip is loudness-normalised so a drill does not jump 20 dB between cards, and all of them are packaged with the app.

Example sentences, the texts and the letters of a word being spelled out have no recordings; nobody has read this course's invented sentences aloud. Those are synthesised once when the course is built, with Piper and a voice trained on Thorsten-Voice — the German corpus Thorsten Müller released under CC0 — and shipped as clips like the rest. They are labelled synthetic wherever they appear. Building them into the course rather than asking your browser to read them is what makes everyone hear the same voice: left to the browser, the voice changed with the device, and a machine set to another language read German with that language's accent.

Speaking cards work the other way round: they open the microphone and check what you said. That part runs through your browser's speech recognition, which on most browsers means the recording is sent to the browser vendor's service — the one thing here that leaves the device. Where the browser cannot listen at all, speaking cards are left out of the session, and Kann nicht sprechen drops them for the rest of a session without counting as a wrong answer.

Loading credits…

Your data

Progress stays on this device. There is no account, no application server and no analytics; the only thing that ever leaves is the audio of a speaking card, handled by your browser's own speech recognition. Export a backup to keep it safe or move it to another device. Importing replaces the progress on this device, so export your current progress first if you want to keep both.