Every ii-5 test scores your speaking across five constructs. The placement test gives you a quick AI read. The Baseline diagnostic deepens that into a study plan with charts. The full speaking test adds human review and a certified result. This page explains every piece of how that grading works, so the numbers feel honest.
Native Japanese isn't just correct grammar with a Japanese accent. It is built on pitch patterns, even mora timing, vowel length, and intonation shapes that don't exist the same way in English. These four features below are where most learners sound textbook, and where small adjustments make you sound natural. They feed directly into the Phonological Accuracy and Discourse Flow constructs.
Each mora is either high or low. The same string can mean different things depending on which moras are high. English speakers naturally stress for emphasis; Japanese uses fixed pitch patterns that change the word itself.
Three different words. Same hiragana. The only thing separating them is which moras are high.
Every mora gets roughly the same amount of time. A small つ (促音) isn't a consonant cluster; it's a full beat of silence. Skipping that beat changes the word.
English speakers often rush through the small っ. To a native ear, kitte said as kite is a different word, not a sloppy one.
Doubling a vowel doubles its length, and the lengthened vowel counts as its own mora. Trim the length and you change the word entirely.
Other common pairs to watch: ビル (building) vs ビール (beer), おじさん (uncle) vs おじいさん (grandfather), 主人 (husband) vs 囚人 (prisoner).
Japanese uses pitch direction at the end of a phrase to mark questions, confirmation, and affirmation. The word itself doesn't change; the rise or fall at the end does the work.
Same verb, no question particle, opposite meaning. Casual Japanese leans hard on this rise to mark questions, especially among friends.
Five constructs cover the things a friendly Japanese listener actually notices. They are scored independently, and different test levels weight them differently.
音声の正確さ
Whether Japanese sounds come out the way a native listener expects. Long vowels (長音), short consonants (促音 っ), nasal n (撥音 ん), pitch accent, and devoiced vowels all live here.
おばあさんが きょうとへ いきました。
The long vowel in obaa-san and the long vowel in Kyoto both land cleanly.
おばさんが きょとへ いきました。
Long vowels collapsed into short ones. "Obaa-san" becomes "oba-san" (a different word: aunt instead of grandmother).
Whisper transcripts are noisy on L2 Japanese, so the AI looks at clarity and consistency across the whole sample, not single-word verdicts.
語彙の柔軟さ
Vocabulary range and flexibility. Word choice variety, appropriateness, and the ability to find a word that fits the intended meaning rather than circling around it.
友達と 待ち合わせを して、駅で 30分 ぐらい 待っていました。
Specific verbs (待ち合わせ, 待つ) and a precise time expression.
友達と あう。駅で 長い 時間 いる。
Vague generic verbs (あう, いる) and approximate time. Communication works, but the listener has to fill in the details.
A larger vocabulary lets you say what you actually mean. The AI rewards precise word choices over hedging vocabulary.
文法の運用
Particles, verb conjugation, tense, aspect, and sentence structure. Errors are expected at lower levels; the AI calibrates expectations to the test level rather than against native speakers.
昨日 図書館で 本を 読んでいたら、友達が 来ました。
Past progressive (読んでいた) plus the conditional ら construction connects two events cleanly.
昨日 図書館 本 読む。友達 来る。
Particles missing, verbs in dictionary form. Listener can guess but works for it.
Self-correction (catching a particle mistake and fixing it mid-sentence) is treated as a strength, not a weakness.
談話の流れ
Coherence across multiple sentences. Use of connectives (それで, でも, だから), the ability to develop an idea over several turns, repair, and natural turn-taking signals.
最初は 難しいと 思いました。でも、毎日 練習したら、だんだん 楽しくなって きました。
Two clauses joined by でも, then a result phrase with なって きました showing change over time.
難しい。練習する。楽しい。
Three disconnected fragments. The ideas might be the same, but the listener has to construct the relationship between them.
On a reading task, this becomes about how naturally you carry the passage forward, not whether you connect ideas.
場面への調整
Register choice (formal vs casual), politeness level, contextual appropriateness, and whether you interpret what the question is actually asking. A polite response to a casual question is just as off as the reverse.
先生、すみません。少し お時間 よろしいですか。
Formal address, hedged request, polite verb form. The listener (a teacher) is treated correctly.
先生、ちょっと いい？
The grammar is fine, but the register is wrong for the listener. Native ears notice immediately.
On higher-level tests, situational tuning carries more weight. On lower-level tests, it is observed but not scored heavily.
We use AI for what it is genuinely good at: fast, consistent grounding on every transcript and audio signal. The model is anchored to specific evidence, not vibes. It does not invent claims. Anywhere the AI is uncertain, that uncertainty is recorded and surfaced to the human grader (on the full test) or shown to you directly (on the Baseline diagnostic).
OpenAI Whisper turns your audio into a Japanese transcript with word-by-word timestamps. This is the foundation for everything that follows. Whisper is biased toward Japanese vocabulary for the test prompts so short clips do not get mis-decoded as English filler.
Speaking pace (words per minute), pause patterns, pacing consistency, and ASR clarity are computed from the Whisper timestamps. Japanese filler words like えーと and あの are counted and density-normalized so a 30-second response with two fillers is not lumped with a 5-second response that has two.
The placement test sends transcripts and signals (and the actual image, when an image is part of the prompt) to GPT-4o. The full speaking test sends the same evidence plus per-question tier weights to Claude Opus 4.7. Both score every construct 1 to 5 on every gradable question.
The AI produces warm, specific feedback in English with Japanese sounds and particles referenced inline. Hedging language like "try to" or "consider" is banned in favor of direct, actionable next steps.
The AI never decides your certificate. On the full speaking test, the AI produces a pre-grade. A certified human grader then reviews everything, edits what they disagree with, and submits the final result. Without that human step, no certificate is issued.
On the paid full speaking test, your final score is reviewed and signed off by a qualified human grader. The AI pre-grades to keep things consistent and to save the grader from typing every word of feedback from scratch. The judgment, the override, and the certificate are theirs.
Graders are working Japanese teachers and licensed examiners. They have heard thousands of L2 learners at every level and know what an Ii3 speaker sounds like compared to an Ii4.
Every grader uses the same construct definitions, tier weights, and scoring scale. The grader portal seeds each question with the AI pre-grade so two different graders on the same submission converge to similar scores.
The AI is advisory. The human grader can override any AI score, rewrite any feedback, and is the only one who can issue a passing certificate. Your certificate has their judgment behind it, not just a model.
The moment you submit a full test, the AI pipeline runs. Whisper transcribes every answer, signals are extracted, and Claude Opus 4.7 produces per-question 1 to 5 scores plus written feedback. This happens whether or not you stay on the page.
A certified grader opens your submission. They see the AI pre-grade as a starting point, listen to every audio response, edit any scores they disagree with, and finalize the per-construct feedback. Tier weights are applied automatically (primary x3, secondary x2, observed not scored).
Once the human grader submits, you receive a notification email and your results page unlocks. If you pass, your certificate is ready to view and share. The result is final and reflects both AI grounding and human judgment.
Every audio response is evaluated by how it sounds to a native listener. What you say matters, but how you say it matters more. No matter your level, the ii-5 model judges your Japanese the way a real native ear would.
Not every question measures all five constructs. A reading task is built to test pronunciation and discourse flow; a casual conversation prompt is built to test grammar and situational appropriateness. We score each question only on the constructs it was designed to test, and we weight those scores by how central each construct is to that prompt.
Across the full test, every construct shows up on many questions, weighted differently each time. The Inversion Measurement Model combines those weighted scores so that your Phonological Accuracy result comes from the questions that actually depended on pronunciation, your Discourse Flow result comes from the questions that actually depended on sustained speech, and so on.
The placement test is free and runs in under five minutes. Take it once to see roughly where you sit, then choose the depth you want.