It only counts if you say it
The main path is to say it out loud, not typing and not picking from four choices. However well you drill, not speaking isn't speaking.
Offline · practise out loud on your own material
Just drop it in — PDF, video, audio. Sentences split, readings marked, shadowed line by line; then the words you save get drilled until you can say something of your own. None of it leaves this computer.
For people who can read but can't speak yet (roughly N4 and up) · free version stays free · Pro ¥1,980 a year (Japanese yen, tax included) · your stuff stays on your computer
This isn't a video. It's the desktop app's own interface, running in this page. The readings, the pitch, the meanings and the audio are all real.
About 300 KB · audio loads when you need it
The demo is a scaled-down desktop window, so it's small on a phone. Turn your phone sideways, or use a computer.
⚠️ The only thing faked in this page is the scoring — actually listening to you needs the model on your computer, which a web page can't do. So the two outcomes here (passed / didn't catch that) are set up in advance to match what the real app does. Nothing is made up.
The main path is to say it out loud, not typing and not picking from four choices. However well you drill, not speaking isn't speaking.
Parsing, audio, scoring and transcription all run on your own CPU — there's no cloud bill ticking by the hour, so you can use it as much as you like, and your sentences and recordings stay on this computer by default.
Double-tap Ctrl to grab a word and the sentence it came from too — so you practise what you actually read, not a random line from an example bank.
Any sentence you pick can be makes a model reading on your machine, as often as you like. Because it knows which line you're saying, it can line it up beat by beat — long sounds, small tsu, high and low. You can see which beat was off.
Being good at drills isn't being able to talk. manabiu pushes every pattern all the way to using it naturally in a real conversation — skip that and you get someone great at drills and lost in conversation.
Drag in a video or audio file and it's your material. No subtitles? Let it listen through once and it writes a draft on your machine — nothing counts until you've checked it. None of it leaves your computer, and there's no limit on length.
All of it works offline. There's no server behind it, and your sentences go to nobody.
Anything your computer can judge reliably never goes online. Only the last step — real conversation — goes to a cloud AI, and even then it shows you, never scores you.
Hear a pattern used for real, then say it after
On your computer · reliableSwap and change, until you don't have to think
On your computer · right or wrong is clearUse the pattern to say something true about you
Mostly on your computerusing it naturally in a real conversation
Cloud · shows you, doesn't judge youWords you save on your computer show up on the phone version. No app to install — just open it in a browser, and it works with no signal. The words that are due, their sentences and the model audio are all there.
The phone version is for review only, not speaking practice — scoring needs the local model, which a phone can't run, so that part stays on your computer.
Generated straight from the git history.
Use it first. Pay if it's worth it. One payment, no auto-renewal.
¥0forever
Use it to look things up, and collect words while you're at it.
The half where you actually speak.
Prices are in Japanese yen, tax included. Card (Apple Pay works), Alipay and WeChat Pay.
Tools like this are usually a monthly subscription, and they cap transcription by the hour. manabiu runs on your own computer with no cloud costs, so it can be a one-time purchase, with no limit on length.
When a monthly or yearly plan ends you go back to the free version, and not one word or day of progress is touched.
The forever price is an early-bird one; it goes back to ¥7,800 once pitch is scored right/wrong. 7-day refund, no questions asked.
Windows 10/11 · 64-bit · about 460 MB of models and dictionaries download the first time you open it. macOS is in progress, Linux after that.
The installer isn't ready yet. Want it as soon as it is? Email [email protected] and I'll tell you when it's out.
Want to pay in RMB, or just talk first?
Xiaohongshu / Xianyu search for manabiu日语工具 to find a reseller — pay in RMB and they'll send you the key straight away.
Email [email protected] — ask anything before you buy. Didn't get your key, or changed machines more than three times? Same address.
You don't have to buy anything to write. Tell me where you're stuck — chances are I've been stuck there too.
If you bought the forever version, all of this comes free.
Looking words up, speaking practice, your word list and your stats are all offline — by default not one byte leaves this computer. Only AI chat goes online, and it needs your own key: leave it empty and that one feature is off, nothing else changes. Phone sync is off by default. Turn it on and it syncs only the words you saved, their sentences and your progress — and one button wipes it all, cloud copy included.
Windows now; macOS in progress, Linux after that. On Linux, double-tap Ctrl to grab a word works under X11 but not under Wayland.
About 460 MB: speech scoring, the voice, and two dictionaries. It downloads when you first open the app, and after that everything is offline. If one piece is missing, only that piece is unavailable and the app still opens. The transcription model is another 130 MB or so, downloaded only when you use it.
That pair is free and about as good as it gets at looking up and remembering, and I'm not rebuilding it. manabiu is the half after that: pushing what you remember until you can say it. If looking up and remembering is all you want, use them instead.
There's a web version for review (works offline) — go through what's due on your way to work. Speaking practice stays on the desktop — scoring needs the local model, and a phone can't run it.
Because we can't do it accurately, and a judgement you can't trust is worse than none — the same rule as everywhere else here. For now you get the high-low of each beat to compare against.