How we count: the method behind the word lists and the ranking

Every number on this site comes from the same pipeline, and this page describes it so you can judge what it is worth. In short: real Japanese subtitles, the first three episodes, dictionary forms, JLPT lists that are themselves unofficial, and translations shown only when we are confident they match.

42 series
46,826 subtitle lines
134,908 word tokens
Human-translated lines: 84% on average

Source material

For each series we use the Japanese subtitles of episodes 1, 2 and 3, the same files Animei plays in its player. Subtitles are fan or official transcriptions of the Japanese audio, not the English script, so the words are the ones characters actually say. We skip a series when fewer than two of its first three episodes have usable subtitles (that is why Evangelion is missing). Opening and ending songs, sound-effect captions in brackets and speaker labels are removed.

Tokenisation

Japanese has no spaces, so the text is segmented with a morphological analyser (MeCab with a modern dictionary) and then re-joined into learnable units: a verb with its conjugation becomes one word in its dictionary form, so 分かった, 分かんない and 分かる all count as 分かる. We keep nouns, verbs, adjectives, adverbs and a short list of fixed interjections (ありがとう, ただいま), and drop particles, auxiliaries, numbers and the generic verbs する, なる, ある, いる from the word lists (they are still counted in the coverage numbers, since a beginner knows them).

Character names and world-specific proper nouns are removed by hand for each series. Single kanji that almost always appear glued to another kanji (影 inside 影分身) are treated as fragments, not words. Katakana tokens are kept only if a dictionary knows them.

Word lists

A series page shows the 60 most frequent remaining words that appear at least twice and have a dictionary entry. The meaning comes from the dictionary used in the Animei app, limited to the first sense for rare words so that 錬金術 does not come out as "a way of making money". For a few dozen very common words where the first dictionary sense is not the one a learner needs, the gloss is set by hand (俺, お前, やつ). Words that mean something specific in a show (魔王 in Frieren, 巨人 in Attack on Titan) carry a note.

JLPT levels

The JLPT organisers publish no official vocabulary lists. We use two unofficial sources: the JLPT tags attached to dictionary entries and a public N5–N1 list of about 7,900 words. When they disagree we take the easier level, on the theory that a word a beginner course teaches is a beginner word. A handful of levels are corrected by hand where both sources are clearly off for anime (ありがとう is treated as N5). Words on no list are marked "show and story words".

The difficulty numbers

N5–N4 speech is the share of content-word occurrences in the first three episodes that are on the N5 or N4 lists. It is computed over every occurrence, not over the 60-word list, and it includes the generic verbs above. Names and series-specific proper nouns are excluded from both numerator and denominator; words with no level count in the denominator only. On any JLPT list is the same calculation for N5 through N1 together. Distinct words per 1,000 is the number of different dictionary forms per thousand content-word occurrences. Average line length is the mean number of characters per subtitle line after cleaning.

An earlier version of the series pages used the share of N5–N4 words within the 60-word list. That number is still shown, but it describes the list, not the show, and it is no longer used for ranking.

Example lines and translations

Each word gets one short line (5 to 30 characters) from the subtitles, preferring lines whose English translation visibly contains the word's meaning. Translations come from human English subtitles where available (84% of lines on average) and are aligned by timestamp, which sometimes attaches a neighbouring line's translation. We therefore show a translation only when it contains the meaning of the word itself and of at least one other word in the line; otherwise the Japanese is shown alone. This is deliberately strict, so many lines have no translation rather than a wrong one. Episode numbers and timestamps are kept for every example.

What the numbers cannot tell you

Three episodes are a starting point, not the whole series. Vocabulary shifts when the plot moves to a new setting.
Nothing here measures speed, mumbling, accents, shouting or background noise, all of which matter as much as vocabulary.
Grammar difficulty is only hinted at by line length. A short line can still be a hard one.
JLPT levels are estimates built on unofficial lists.
The counts are per subtitle line: a word said twice in one line counts once.

Updates and corrections

Pages are regenerated when the data or the method changes, and the date at the top of each page is the last time its content actually changed. If you spot a wrong meaning, a misassigned level or a name that slipped through, tell us through the support page and we will fix the source, not just the page.