Skip to main content
&Sageio
All notes

Sageio Type

Japanese dictation on a Mac: kanji conversion is the real test

Turning Japanese speech into kana is the easy half. Choosing the right homophone kanji — 公開, 後悔, 航海 all sound like koukai — needs the whole sentence, and matching です/ます against plain form matters the moment you dictate an email instead of a chat. Here's what separates Japanese dictation you can send from Japanese dictation you have to rewrite.

By Ming · · 5 min read

Japanese looks like a language dictation should handle well: the sounds are clear, the syllables are regular, and almost anyone can transcribe spoken Japanese into kana without much trouble. That's exactly why it's misleading. Getting from speech to kana is the easy half of the job — the half most tools solve. The hard half starts after that, when those kana have to become the specific kanji the writer meant, in a register that fits whether they're firing off a chat message or sending a client an email. A tool that stops at accurate kana has done the part that was never in doubt and skipped the part that decides whether you can send what came back. Here's what the real test actually looks like.

The koukai problem: same sound, different kanji, only context decides

Japanese resolves a small inventory of syllables into an enormous set of words, and the writing system makes you commit to which word every time. Say koukai and you might mean 公開 (to make public), 後悔 (regret), 航海 (a voyage), or 更改 (renewal) — same reading, four different meanings, four different kanji. Kikai is 機械 (machine) or 機会 (opportunity). Seika runs through 成果, 製菓, 生花, 聖歌, and more. Kana captures the sound faithfully and tells you nothing about which of these the speaker meant.

This is why the conversion step is where dictation succeeds or fails. Pick the kanji from the syllable alone and you're guessing; the only thing that disambiguates koukai is the sentence it sits in — "情報を公開する" versus "公開を後悔する." A tool that converts reading by reading, committing before it has heard the rest of the clause, will hand you a plausible-looking wrong word. And a wrong kanji in Japanese isn't a typo you skim past — it's a different word that reads as fluent nonsense. Real Japanese dictation has to hold the whole sentence and choose kanji from meaning, not from sound.

です/ます or plain form: the register has to match where you're writing

Japanese doesn't have one neutral way to say a sentence. The same thought lands as 送りました, 送った, or 送っておきました depending on who you're writing to and where. Dictate a note to a colleague in chat and plain form is right; dictate the same content into an email to a client and it has to come back in polite です/ます form, often with the surrounding phrasing that register expects. This isn't decoration — using the wrong level in Japanese business writing reads as either stiff or careless, and neither is what you meant.

That makes register a first-class part of the job, not an afterthought. A tool that transcribes whatever politeness level you happened to speak in leaves you rewriting the ending of nearly every sentence to fit the destination. The better model is to treat the rewrite as tone-aware: understand that an email and a chat message are different registers, and produce text already shaped to the one you're writing. Speaking is fast and loose about politeness; the sentence you send can't be. Closing that gap — matching the register to the context instead of echoing the speech — is where dictation stops being a transcript and starts being writing you can use.

Katakana, loanwords, and the English sitting inside the sentence

Real Japanese is full of borrowed words, and they raise their own conversion question: when sapooto should be written サポート and when it should stay the English "support," when aatikuru is an article and when it's a name. Katakana is a whole layer of the language, not an exception to it, and business Japanese leans on it constantly — スケジュール, リリース, フィードバック. On top of that, spoken Japanese in a tech or product setting mixes in bare English terms: "そのdeadlineまでにreviewを終わらせます." A writer needs the loanwords rendered as the katakana they'd actually write, the English terms preserved as English, and the Japanese around them left as clean Japanese — all assembled from one stream of speech that slid between the two.

A tool locked to a single language mangles the other half. Force everything into Japanese phonetics and an embedded "review" comes back as garbled katakana; force English and the Japanese collapses. Getting this right means recognizing, word by word, which layer each piece belongs to and writing it in that layer — the ordinary condition of working Japanese, not an edge case to bolt on later.

Why "supports Japanese" tells you almost nothing

A tool can list Japanese, transcribe a slow clean sentence into flawless kana, and still fail every part that matters: the homophone kanji that need the whole sentence, the politeness register that has to match an email versus a chat, the katakana and English woven through ordinary business speech. The feature list won't tell you which, because kana accuracy — the easy half — is what a quick demo shows. The conversion, the register, and the mixed script are what a real message exercises.

The test that actually tells you

One session settles it. Dictate the way you actually work — a sentence with a koukai-class homophone in it, phrased for the destination you'd really send it to, with the katakana and English terms you'd really use — and read what comes back. Did it pick the kanji you meant? Is it in the register that fits an email if that's what you were writing, or a chat if it wasn't? Did サポート and "review" each land in the right script? That's the difference between output you'd send and output you'd rewrite, and it's the reason we built Sageio Type around understanding the whole sentence before it commits to a single character. A language count can't answer any of it. One honest Japanese sentence can.