Shaberu

Talk, the way you already talk.

The filler gone, the punctuation in, your own names spelled the way you spell them — straight into whatever you were already typing in.

Anywhere there is a cursor

  • Terminal
  • Gmail
  • Slack
  • Notion
  • ChatGPT
  • LINE
  • Messages

What the recogniser heard

um so can you push the fix to cloud flare tonight uh and bump the post gres version and remember to update the docs

What Shaberu sent

Can you push the fix to Cloudflare tonight, bump the Postgres version, and remember to update the docs.

Recognition Filler and punctuation Your vocabulary Polish

Tap to start, tap to send

There is no hold-to-talk. Tap right Shift and a capsule appears at the bottom of the screen: cancel on the left, the live waveform and a running clock in the middle, send on the right. Tap again to send, Esc to throw it away.

On iPhone it is a keyboard

Switch to Shaberu with the globe key, talk, and the finished text is typed into the field you were in. The same account and the same settings as the Mac.

Your names and your jargon

Add them once and they reach the recogniser before it listens, rather than being swapped in afterwards by a rule that cannot hear anything. Both devices read the same list.

This one is real: type a word in, take one out. It is yours until you close the tab.

  • Shaberu
  • Cloudflare
  • Postgres
  • Anthropic
  • Kubernetes
  • appkit

The last fifty, on your own machine

Text that landed in the wrong window can be copied back out, and anything you would rather not keep can be deleted on the spot. Plain text, on your Mac — there is no copy on our server for us to hand back.

Emoji are spoken, not hunted for

"Heart emoji" is ❤️ and "thumbs up emoji" is 👍 — around a hundred names, in English and in Chinese. The name on its own is left alone: "a change of heart" is the far more common sentence, and it stays one.

Say what you want done with it, mid-sentence

Ask for a new line, a new paragraph or a list while you are still talking and it is carried out rather than typed. The instruction never reaches the page, and the sentence around it does. It is the difference between dictation that transcribes you and dictation that takes notes for you.

make that a list first confirm the budget then set the timeline then find the people

  • Confirm the budget
  • Set the timeline
  • Find the people

Built for the way people actually talk

Nobody speaks in finished sentences. Real speech has false starts in it, and filler, and the same word said twice while you decide what the next one is — and it is full of names no recogniser has heard of, the ones your team and your work are actually made of. Most dictation hands all of that back to you as one long unpunctuated transcript and leaves the tidying up as your problem. Shaberu does the tidying: the recognition, your own vocabulary, the punctuation and a final polish all run on our server before the text comes back, which is why it keeps getting better without you updating anything. The recording is transcribed and dropped in the same request, and never written to disk on our side.

The recording is not kept anywhere

Your audio goes to our server, is transcribed, converted and polished, and the text comes back. It exists only in memory for the length of that one request. It is never written to disk, and there is no storage bucket it could be written to.

  • The transcript is not kept either. The finished text goes back to the app and is gone from the server with it — there is no copy here to switch off, and none to delete. If your app asks us to hold an answer so a failed retry cannot lose it, what we hold for five minutes is sealed with a key only your device has.
  • Recognition is performed by OpenAI and polishing by Anthropic, as processors on our behalf. Neither trains on data submitted through their APIs.
  • Deleting your account removes the vocabulary, the settings and the device list immediately and irreversibly.

Why Shaberu

ShaberuTypelessWispr Flow
Chinese the way Taiwan writes itTraditional characters, and Taiwanese word choice. A video is 影片 and stays 影片; it does not come back as 視頻.Traditional Chinese is recognised well, but the wording it writes is often the Mainland one — ask for 影片 and you get 視頻 — and Simplified characters do turn up.Traditional and Simplified are two separate settings and it honours the one you pick. The wording underneath is not tuned for Taiwan either way.
Chinese and English in one sentenceSay 「這個 deadline 我先 hold 住」 and it comes back written that way, with the spacing between the two scripts sorted out.Handles it too. No need to pick a language first.Their own documentation is explicit: Flow picks one language per session, and with Chinese and English together it may write the English in Chinese characters or the other way round.
Your own names and jargonThe vocabulary reaches the recogniser before it listens, rather than being swapped in afterwards by a rule that cannot hear anything. Both devices read the same list.A personal dictionary, and it learns from the corrections you make.Respellings you type get added to your dictionary automatically.
An instruction said mid-sentenceAsk for a new line or a list while you are still talking and it is carried out rather than typed. Emoji are spoken by name. On the free plan too.Select any text and rewrite, translate or ask about it out loud — a wider set of things than we do.Command Mode is a Pro feature.
Where the recording goesIt exists in memory for one request and is gone with the answer. Never written to disk, and there is no bucket here it could be written to. The transcript is not kept either; your history is the last fifty, on your own machine.They say the same thing: zero cloud retention, never trained on, history kept on the device. This row is a draw.Audio, transcripts and history are stored on their servers, and whether they are kept — and whether they train models — are settings you manage. Against that, they hold SOC 2, ISO 27001 and HIPAA paperwork and we hold none of it.
The free plan5,000 words of finished text a week, every feature switched on, no card and no expiry.8,000 words a week, which is more than ours. The catch is standard accuracy and standard queueing at busy times.2,000 words a week on desktop and 1,000 on iPhone, after a 14-day Pro trial.
What it costsUS$7.99 a month, or US$59.99 a year — US$5.00 a month.US$30 a month. The US$12 a month everyone quotes is the annual plan: US$144 up front.US$15 a month, or US$12 a month billed annually: US$144 a year.
DevicesMac (macOS 14+) and an iPhone keyboard, on one account. There is no Windows build and no Android one. This row is theirs.macOS, Windows, iOS and Android.macOS, Windows, iOS and Android, plus a meeting notetaker on the Mac.
OfflineNo. Recognition and the polish both run on the server. What that buys is a word it gets wrong today being fixed today, without you updating anything.Also needs a connection.Also needs one — their own documentation says transcription always happens in the cloud.

Typeless and Wispr Flow prices, limits and features were read off their own pricing pages and help centres in August 2026; if they change them, this table will not know until we go and look again. The row about Chinese wording is what we got saying the same sentence into all three.

Plans

The free column is deliberate

Shaberu is a side project, built by one person in their spare time, so that the way Taiwan speaks does not quietly disappear with time. The free allowance is generous on purpose, and the price is about as low as tools like this go. If it earns a place in your day, paying for it is what keeps it going.

Free

US$0

  • 5,000 words of finished text a week, resetting every Monday
  • gpt-4o-mini-transcribe, polished by Claude Haiku 4.5
  • Your vocabulary, both devices and every feature, all of it
  • No card, and no expiry

Upgrade to Pro →

FAQ

Fair questions

Does it work offline?

No. Recognition, the wording and the polish all run on the server, and there is no degraded path on a plane. That is the trade the design makes on purpose: a word it gets wrong is fixed by a deploy the same day, rather than by an app update everyone has to install.

Is my audio stored?

No, and neither is the text. The audio lives in memory for one request and is gone when the dictation comes back, and so does the text — unless the app asks us to hold its answer for five minutes so a dropped connection cannot lose the words with it, and that copy is sealed with a key only your device has. Your history is the last fifty on the device itself — the server has no list to keep.

Which apps does it work in?

All of them. Shaberu types into whatever is in front of you, so Gmail, Slack, Notion, Messages, the terminal, ChatGPT and LINE are the same case. If there is a cursor, there is somewhere for the text to land.

Why does it need Accessibility permission?

It is what lets Shaberu see a shortcut key pressed in another app and put text into that app. macOS offers no other way. You grant it once, under System Settings, Privacy and Security, Accessibility.

How is the free allowance counted?

In words of finished text rather than minutes of audio, so a pause, a false start, or a recording your microphone never picked up costs nothing. A script written without spaces between its words is counted per character instead, at the same rate. A minute of ordinary speech is about 150 words, which makes 5,000 words roughly half an hour a week.

What does the finished text look like?

Like something you would have typed yourself: punctuated, capitalised, the filler gone, and every name off your vocabulary spelled the way you spell it. There is no setting for it and there does not need to be.

Does it run on an Intel Mac?

Yes. Nothing runs on your machine, so there is no Apple Silicon requirement. macOS 14 or later is the whole of it, plus a network connection.

What about the iPhone?

It is a dictation keyboard with the same account, the same vocabulary and the same pipeline, reached with the globe key instead of a modifier key. It is in open beta on TestFlight while the store listing is being prepared, and the button above is the invitation.

Say it, and it is written.

5,000 words a week, free, no card. Decide after you have used it for a week.