How Katsuyo is made

Katsuyo is a free Japanese conjugation practice app and reference. It drills the conjugations of every verb and adjective on the JLPT vocabulary lists, from N5 to N1, and publishes the full conjugation table of each one at playkatsuyo.com/words. It runs in the browser and on iPhone and iPad, has no advertising and no tracking, and needs no account.

Who makes it

Katsuyo is built and maintained by one independent developer, as a one-person project. There is no company behind it and no content team: the word list, the conjugation engine, the quiz design, the English prompts and the audio pipeline described below are all one codebase, and the same code produces the app and these pages.

Corrections are welcome. If a form, a reading or an English meaning on any page looks wrong, email feedback@keshiki.app with the page's address, or use the feedback form in the app.

Where the words come from

The Japanese Language Proficiency Test publishes no official vocabulary list. Katsuyo uses the community-compiled JLPT lists that most learners study from, matched entry by entry to JMdict, the open Japanese–English dictionary maintained by the Electronic Dictionary Research and Development Group. Every reading and every English meaning on this site is JMdict's, and a word's JLPT level is the level its list assigns it.

Only conjugable words are kept: verbs, い-adjectives and な-adjectives. Each is assigned a conjugation class — ichidan, one of the godan endings, one of the irregulars, or an adjective class — from JMdict's part-of-speech tags, which is what the engine conjugates from.

How the conjugations are produced

One conjugation engine generates every form in the app and on every reference page: 21 forms for a verb, 11 for an adjective, across the plain, polite and ている paradigms. There are no hand-typed tables. The engine is covered by automated tests for each class and each irregular, and every form quoted in a page's prose — the sentence under each headword, the pattern card — is checked against the engine when the site is built, so a page cannot describe a form its own table contradicts.

Where a verb has a form that is grammatically valid but that Japanese does not actually use, the table still shows it, dimmed, so the paradigm stays complete. How that is decided is the next section.

Which forms the game quizzes, and why

A conjugation the game asks for has to be one that people say. Each verb-and-form pair is counted in two corpora: the Japanese OpenSubtitles corpus, for speech, and the CC-100 web corpus, for writing. A pair is quizzed only when it clears an evidence floor in at least one of them. That is why, for example, the imperative of ある is in the table but never in a quiz.

The English clause each quizzed form is asked by — "Hand it over!" for 寄越せ — is written per form, not per word, because a verb's forms do not share one sense. The clauses are authored and reviewed by hand, and the same clause appears beside that form on the word's reference page.

Every wrong answer in a quiz is a misspelling: a form built the wrong way — the wrong sound change, the wrong vowel row, a る that should have dropped — and checked against every form of every word in the list, and against JMdict, so that it is not a real word. The game never shows a real form as a wrong answer, so nothing it marks wrong is something a learner should unlearn.

The N-level tag beside a form on a reference page is Katsuyo's own: the JLPT level at which the game starts quizzing that form. It is a pedagogical ordering, not a claim about the test.

How the audio is made

The pronunciation of each conjugated form is synthesised, not recorded: the clips are generated with Google's Gemini text-to-speech model, one clip per form. Synthesis can misread a kanji — the same characters often have more than one reading — so every clip is checked before it is published. A transcription model listens to the clip and writes down what it heard in kana; if that does not match the intended reading, a second forced-choice check decides between the intended reading, what was heard, or neither; clips that fail are regenerated from a spelling that fixes the reading, and the ones still in doubt are listened to by a person. A reference page only offers a play button for a clip that passed.

Privacy

No ad networks, no web fonts, no cookies. The site keeps a count of page views with GoatCounter, which stores no identifier and cannot follow anyone from one visit to the next; the iPhone and iPad app carries nothing of the sort. What the app does send, and only with an optional account, is described in the privacy policy.

Credits and licences

This site uses the JMdict/EDICT and KANJIDIC dictionary files. These files are the property of the Electronic Dictionary Research and Development Group, and are used in conformance with the Group's licence (CC BY-SA 4.0). Corpus counts use the OpenSubtitles Japanese corpus from OPUS and the CC-100 Japanese corpus. Japanese text is set in Noto Sans JP under the SIL Open Font License 1.1.

Last updated 11 September 2026.