Nihongo Practice Reading: Why Decoding Keeps You Slow, and What Builds Speed

Short answer first. Slow Japanese reading is a retrieval shortage, not a character shortage: you can already pronounce the kana and plenty of kanji, but too few words arrive whole, so every line pays a decoding cost that re-reading never removes. In numbers: a 40-line NHK News Web Easy article carrying twelve unfamiliar compounds spends about 36 extra seconds on stopping and restarting, and about six once those twelve are cards you have each retrieved five times. The practice that works is capture plus scheduled retrieval: mark the words that got away from you, confirm the strongest as cards, and let spaced repetition schedule the returns. Disclosure up front: I build Piccard, and every product number below comes from its code.

What is slow Japanese reading, actually?

It is decoding that has not yet graduated into reading: you build each word from parts (pick a reading, test it, hold it in working memory) instead of receiving it as one unit.

Reading at speed is two jobs done at once. Job one is decoding: identify the characters, choose among their possible readings, sound the result into working memory, and keep it all alive until the sentence resolves. Job two is recognition: 環境 lands on the eye and かんきょう, "environment," comes with it, whole. Every word in this paragraph arrives that way for you. Japanese reading speed is the balance between those two jobs.

Japanese makes the first job stubborn in ways most languages don't, and that is the hinge of this post. A single sentence routinely carries all three scripts (kanji stems, hiragana endings, katakana loanwords) with no spaces marking where any word ends (Japanese writing system). And kanji hold several readings each: 生 is sei in 学生, nama in 生ビール, u- in 生まれる, i- in 生きる. Some compounds ignore their parts entirely. 大人 is otona, a reading neither character contributes. Most scripts hand you a pronunciation; kanji hand you a puzzle whose solution depends on the neighbors.

Furigana, the small kana printed alongside kanji to supply the reading, is the standard accommodation, and it quietly swaps the task. The ruby does the reading for you, so you confirm rather than produce. A useful scaffold, but furigana-dense material stops testing the skill the moment you lean on it.

Fluent readers, reading research finds again and again, barely decode at all; they lift words and fixed phrases off the page as single units and spend their attention on meaning. The mirror version of this argument, for a language whose script is fully regular, is Practice Reading Korean Words: hangul makes decoding cheap within months, so speed is purely recognition. Japanese never pays off that loan. Decoding stays expensive for years, which raises the stakes on every word you move to the recognition side.

What a second pass trains

A second pass rehearses your memory of that page, and only the words travel with you to the next one.

Memory research named the relevant effect long ago: the testing effect, the finding that hauling an answer up from memory strengthens it far more than meeting that answer on the page a second time, shown cleanly in Roediger and Karpicke's 2006 experiments. A second reading asks you to produce nothing. You glance at answers sitting in plain sight, and glancing is not retrieving. What you bank is familiarity with that page, an asset worth nothing on Thursday's article.

Now the arithmetic, per word. Take a 40-line NHK News Web Easy article and suppose twelve compounds outrun you: 昨日 is yours, 経済 is not. Decoding a half-known compound costs about three seconds (choose a reading for each character, check it against context, rebuild the sentence around the pause). Retrieving a drilled compound costs about half a second. Twelve strangers at three seconds each is thirty-six seconds of stop-and-go across the article; the same twelve as familiar cards cost six. Which of the two columns those twelve words occupy is the whole of the difference. And the stalls cost more than their seconds, because with no spaces to lean on, every stall forces you to re-find a phrase boundary you had already solved.

What each format trains, and where it stalls

Every page you open is a word list in disguise, so picking material is picking vocabulary. The formats trade differently:

The JLPT orders learners from N5 up to N1 if you want an external checkpoint. It measures; it doesn't train. If N5 is your stage, the word-list view of that level has its own post: JLPT N5 Vocabulary: The List That Isn't Official.

Chosen material outlasts assigned material

The words live there: the things you were going to read or watch anyway contain exactly the words you personally keep meeting, and a syllabus cannot know them. That fork is worked through in The 1000 Most Common Japanese Words: every good list hands you the same few thousand, and past them sits a remainder made of exactly the words your own material keeps using. Lists settle the first part. Nobody can hand you the remainder, and speed lives there.

The document route, Japanese direction, works like this: paste an article, a chapter, or a raw vocabulary list, and what returns is a ranked field — every candidate tagged high, medium or low by learning value, strongest first, top tier already ticked, and a confirm button that will enroll no more than 40. That number is written into the server code. List- and chapter-sized pastes return a single batch of 25 or fewer; only text past 120,000 characters splits (four chunks, approaching 100). The propose cache holds ten minutes, so an identical re-paste returns the identical batch — depth arrives when you paste the next slice of the document, never before.

The ranking exists to aim your review minutes. Words you plainly already retrieve sit near the bottom, unticked, spared from re-certifying what a frequency list would have handed you anyway.

From watching to reading: the caption route

Open a YouTube video carrying a Japanese caption track and the extension unrolls its transcript one line at a time. Click the line you stumbled on (the line is the unit, and nothing anywhere converts the whole video) and it returns that line's candidates, each one a word with its gloss and a sentence showing it at work. The confirm cap here is 20 per round, a smaller constant than the document route's 40. A video with no caption track simply has no transcript to unroll.

A caption line is reading practice in miniature: one complete sentence, arriving at the speed people actually talk, about something you picked, with the scene still attached. The click doubles as diagnosis (this sentence beat me) and cure (the card carries the sentence and its timestamp). Come review time, a caption card restarts the video at its own line, so the word comes back inside its original context rather than on a bare flashcard. The full mechanics of subtitle capture are laid out in How to Save Words from YouTube Subtitles (and Actually Remember Them).

For text rather than video, the flow-preserving move is the double-click. Click twice on any Japanese word in whatever you're reading, whether that's a forum thread, a wiki, or the news, and a small dictionary window opens on the JMdict entry; making it a card is one click further. You leave the line for a breath at most. Capture happens mid-read; retrieval happens later, on schedule.

Uploaded audio and video follow the same shape: the file is transcribed into lines you pick from. Point-and-pick by design. Nothing converts a whole recording on its own.

The schedule, and a week run on it

Untended memories sink on a timeline first plotted in the 1880s, the forgetting curve. The repair is spaced repetition: each item comes back shortly before it would fade, and the wait lengthens with every answer you get right. Piccard schedules on FSRS-6 and leaves nothing to configure: the retention target is fixed at 0.9, and grading means choosing among again, hard, good, or easy. Review is the retrieval half of the loop. Every card in today's queue is a word some page once made you stop for.

The week I would run:

1. One chosen thing a day: an Easy article or a captioned episode, with the double-click dictionary as the only permitted interruption.
2. Keep only the words that beat you. Twelve real strangers are worth more than a bloated list; the tier ranking exists to enforce exactly that taste.
3. Once a week, feed the week's collected text to the document route; confirm the ticked top tier. Forty per confirmation is the ceiling, which at any sustainable pace is more than a week's worth.
4. Answer every card the queue serves up each day, and when a caption card brings its video moment, read the line on screen out loud, then rate.

The short version

Slow Japanese reading taxes every line roughly three seconds per undrilled compound, and kana fluency never repealed the tax. Three scripts in one line, no spaces, multi-reading kanji: decoding stays expensive for the long haul, which is why each word promoted to the recognition side keeps paying out.

Practice, then, means retrieving. Keep reading whatever you were going to read regardless; catch whatever slipped past you (double-click on text, caption-click on video); feed the week's text to the document route and confirm the strong tier. From there the scheduler owns the calendar. The timing is the training.

The free tier is 150 energy (about 150 cards — a dozen captures a week keeps the tier alive for three months), then prepaid top-ups at $5 / $10 / $20 — pricing here. No subscription, and topped-up energy never expires. The caption route starts with a video you already have open: Piccard on the Chrome Web Store.

All guides: Piccard Blog.