macOS · Linux · runs on your own computer · $19 once
Coii VoiceInput turns speech into text in the window you were already working in: a sentence back in about half a second, punctuated, in Chinese or English or both. The model runs on your own computer. The recording never leaves it, and is never kept. There is no account.
30 days free, no card · then $19 once · macOS 13+ and Linux x86_64
Coii VoiceInput — History
Ship the release on Friday, and tell the team.
SlackRecommended0.5sVerified
明天下午三点开产品评审会，记得带电脑。
NotesRecommended0.5sVerified
Move the retro to Thursday so the whole team can be there.
MailRecommended0.6sUnverified
Add a note about the flaky search test before I forget it.
ChromeRecommended0.7sCopied
nothing kept but the text
nothing sent · 0 requests
Nothing is uploaded
Audio and text stay here
No account
Nothing to sign up for
Works with Wi-Fi off
The model is on your disk
$19 once
30 days free first
How it works
No window to switch to, no Send button, no server to wait for. Coii VoiceInput listens for a key you already have.
The pill appears over whatever you were doing and the level meter moves, so you know it is hearing you. Where the text will go is decided the moment you press the key, not later.
Let go, and a row goes into your history before the engine is asked a single question. Nothing downstream — a slow model, a crashed engine, a paste that never lands — can cost you the sentence. The audio itself is discarded once transcribed.
Straight into the field you left, through the accessibility API where that works and a synthetic paste where it does not, with your clipboard put back afterwards. About half a second for a sentence with the default engine.
Getting the text into your app
Typing into somebody else’s application is the least reliable thing this app does. So it is tried in three ways, and each one is honest about what it can prove.
“Sent, but this app cannot tell whether it landed” is the ordinary case in a browser or an Electron editor. It gets its own outcome and its own line in the pill rather than being reported as success — and your clipboard is handed back either way, because a dictation never keeps it.
The ladder
Accessibility API
Written into the focused field directly. Read back — so this one can say Verified.
Synthetic ⌘V
Your clipboard borrowed for one keystroke and given back — whatever the paste did.
Your history, and it says so
Nothing typed, and nothing taken: the clipboard is left exactly as you had it. The sentence is in your history, and the pill and a notification say so.
And a step before all three: a password field, a moved focus or a missing permission is refused up front, with the text put somewhere safe rather than typed somewhere wrong.
Features
A first word clipped off, a paste that lands in the wrong window, a name spelled four different ways — the awkward cases shaped this.
The row is in the database before the engine is asked anything, and long before another application is typed into. A crashed engine leaves the dictation in the history rather than nowhere, and so does a refused paste — one click from the clipboard, rather than dropped onto it. The recording itself is never kept: the text is the thing worth keeping.
Swapping engines is a click, and a new one arrives as a file rather than an app update. Local speech recognition moves every few months; the app does not have to be rewritten to change its mind.
The default engine detects the language itself, handles both mixed inside a single utterance, and punctuates the Chinese. Many more languages are one engine away.
A substring finds a sentence in any script, with no word boundaries to get wrong. Type two letters and it still matches rather than going blank.
The pill is a panel that cannot take focus. If it could, the field you were typing in would stop being focused the instant you started talking, and the app would be typing into itself.
Repeated words are collapsed and your own vocabulary is applied, so a name the model has never seen still comes out spelled right. Both versions are kept: what the model said and what you got. The defaults fix artefacts of transcription and leave style alone. Filler removal is off.
Engines
Several models ship with the app, all of them local. They differ in size, speed and languages, and switching is a click, not a reinstall.
Recommended
Chinese and English, mixed in one sentence, with punctuation. What the first run installs.
Light
A fraction of the size, for an older Mac or a slow connection. Words on screen inside a minute.
Most languages
Dozens of languages in one model, and the path for a model you bring yourself.
A model your Mac cannot run still appears in the list, greyed, with the reason. Sizes are shown before you download.
What it sends
“Private” is easy to write on a page, and a dictation app is the worst place to take it on trust. So here is the complete list instead: every connection the app can make, and the one row that says what it never does.
01
When you open the licence panel
A HEAD request to cloudflare.com (or www.apple.com if that fails), read for its Date header only, so a wound-back clock cannot buy free trial days. It carries no identifier. Offline is not an error.
02
When you install an engine
The model weights are fetched once from a public model host, which sees your IP address and which file you asked for. After that the files are used from your disk and the host is never contacted again.
03
When you activate
Your licence key goes to Lemon Squeezy, plus a device label made from your Mac’s name. At most three calls per device: check, activate, deactivate. Not your audio, not your transcripts, not how often you dictate.
04
Everything else
Usage statistics are optional and off by default. If you opt in, PostHog receives feature events, app version and coarse performance data, never your content. You can turn this off in Settings. No automatic crash reporting. Once a day it asks the download server whether a newer version exists — nothing about you goes with the question, nothing installs unless you press it, and a switch in Settings stops it. Engines that run as a separate process talk to the app over the local loopback only.
Dictating with the network off does exactly what dictating with the network on does.
Read the privacy policyPutting text into somebody else’s window is the part of this app that is different on every platform. On macOS it is the accessibility API followed by a synthetic ⌘V. It confirms what it can, says “unverified” when the application will not answer, and points at your history when it fails — never at a clipboard it took from you.
On Linux it is a Ctrl+V. Under Wayland the desktop asks once whether the app may send it, and remembers the answer; the dictation key needs a desktop with the global-shortcuts portal, which on Ubuntu means 25.04 or later — on 24.04, use the Xorg session. A terminal pastes with Ctrl+Shift+V instead, so there the sentence waits in your history. Windows builds, but is not released.
Pricing
Thirty days with everything unlocked, no card and no account. Then $19 USD, once.
Personal licence
$19USD, once
One payment. No subscription, nothing renews.
You can buy from inside the app at any point in the trial.
Checkout by Lemon Squeezy, our merchant of record. Tax handled, invoice included.
Lost your key? Lemon Squeezy’s order lookup mails it to the address that paid. No email to us required.
Questions
Thirty days, every engine, nothing to sign up for. Dictate one message tonight and see.