HushQuill
Free for Windows and macOS

Your voice never leavesyour machine.

On-device dictation with AI formatting, in about 99 languages. Press a key, speak, and clean text lands in whatever app you are already in. Buy it once. Windows and Mac.

tauri 2  ·  rust  ·  whisper.cpp AVX2  ·  parakeet tdt v3  ·  8.4 MB

Download free Read the spec

Windows 10 and 11, macOS 11 and later. Free forever, no account, no telemetry.

The bar you actually see

Lands in your editor as

Why it is different

Three things the cloud apps cannot do.

Dictation is a solved problem. Dictation that does not ship every word you say to somebody else's server, that runs properly on both Windows and Mac, and that does not bill you monthly, is not.

Nothing leaves your device

Whisper and Parakeet run on your own CPU. Turn the WiFi off and dictation still works. The only cloud options are opt-in, off by default, and use a key you supply yourself.

Windows and Mac, one app

Almost every polished dictation app is Mac only. HushQuill is native on both, from one codebase, so the features land together. It runs at many times realtime on an ordinary laptop CPU with no GPU, and uses the GPU on Apple Silicon.

No subscription. Ever.

Unlimited dictation is free, permanently, with no watermark and no nag, in every language. Pro is a single payment for the accuracy and the AI layer on top. There is no monthly plan at any price, and there never will be.

What is in it

Everything below is in the free build.

Not a stripped demo. This is the whole dictation engine, uncapped, in every language. AI rewrites are metered at two a day on the free build, and snippets, history and the cloud engine are Pro. Every item here is covered properly in the user guide.

Two hotkeys

One shows and hides the floating bar, one starts and stops dictating. Both rebindable.

AI formatting

Your dictation is rewritten to suit the app you are typing into, and you can finish a take with a spoken instruction like "make that formal". Two a day free, unlimited on Pro.

Three dictation modes

Toggle, continuous (each pause types a segment and keeps listening), or push to talk.

Auto engine picking

Chooses the fastest engine for your language and your hardware. Override it whenever you like.

Silence detection

A pure Rust WebRTC VAD trims dead air and can stop the take for you after a set pause.

Types into any app

Editors, browsers, chat, terminals, Office. The overlay never steals focus from what you are writing in.

Text never gets lost

If no text box is focused the transcript parks on your clipboard, and one button re-inserts it wherever you click next.

Cleanup and dictionary

Filler removal, punctuation and capitalisation, plus your own spellings for names and jargon.

Snippets

Say a trigger phrase, get a full block of text. Signatures, addresses, boilerplate. Part of Pro; your dictionary stays free.

Voice commands

"New line", "new paragraph", "scratch that". Deliberately few, so they never collide with real speech.

Tray and autostart

Launches to the system tray at login and stays out of the way until you press the key.

Updates in the app

Signed updates install in one click. Full release notes are readable inside the app before you take them.

For the people who ask first

The honest spec.

No Electron, no bundled browser, no background service phoning home. The installer is 8.4 MB because there is not much in it beyond the app.

Speech models are downloaded on first run from within the app, so you choose what lands on disk and when.

Shell
Tauri 2, Rust backend
Local STT
whisper.cpp, AVX2 baseline
Fast engine
Parakeet TDT v3, ONNX Runtime
Voice detection
WebRTC VAD, pure Rust
Installer size
8.4 MB
Network at rest
one update check on launch
Dictation network
none, unless you enable a cloud engine
Telemetry
none
Account required
no
Windows
10 and 11, x64
macOS
11+, universal binary
Licence model
one-time, 3 devices

The maths

Cloud dictation subscriptions run about $12 to $15 a month. That is roughly $150 a year, every year, to rent something that runs perfectly well on the laptop you already own.

HushQuill Pro is $29 once. The free tier never expires and never asks for a card.

Languages

Speaks about 99 languages. Including the ones nobody else bothers with.

Your language is one pick away in Settings, or set it to auto-detect and speak whatever you like. Urdu, Hindi and other native scripts render right to left correctly, and code-switching mid-sentence does not break it.

Native script, not transliteration

Urdu comes out in Urdu script with the right punctuation, not Roman approximations.

Every language, free

Every language works on the free tier, uncapped, and there is no language pack to buy. For Urdu, Hindi and other non-European languages the free build now includes a higher-accuracy model, and Pro adds the cloud engine, which is more accurate again.

What we do not claim

We do not claim to beat Whisper on raw Urdu accuracy. We claim it is private, it is on Windows as well as Mac, it handles the script properly, and it is free to try forever.

AI formatting

Speak once. It lands written for wherever you are.

HushQuill sees which app has focus and rewrites the same dictation to suit it. Pick a destination and watch what changes. Two a day are free; Pro removes the meter and lets you redefine any of these styles in your own words.

What you said fillers struck out

um so the login thing is broken again, when you use a really long email address it just fails silently, no error, nothing, and uh I think it is the same bug we saw last month, we should probably fix it before Friday

What lands in the app style: email

You can also just say it. Finish a dictation with make that formal or turn that into bullets and it obeys, whatever app you are in. Formatting runs locally through Ollama by default, so the text never leaves your machine. The free tier includes 2 rewrites a day, which is enough to see whether you want it.

Pricing

Free is the mechanics. Pro is the intelligence.

If it is rules and speed, it is free. If an AI model does work on your words, it is paid. That is the entire tier line.

Free

$0 forever

No account, no trial clock, no expiry.

  • Unlimited dictation, no caps or watermark
  • All three dictation modes, both hotkeys, the bar
  • Parakeet fast engine and Whisper tiny plus base
  • About 99 languages with auto detection
  • Higher-accuracy model for Urdu, Hindi and other non-European languages
  • Cleanup, custom dictionary, voice commands
  • 2 AI rewrites a day, so you can taste the paid part
Windows macOS

Pro

$29 once, first 500

$39 after that. One payment, 3 devices, free updates, no subscription at any price.

  • Everything in Free
  • Unlimited AI formatting, no daily meter
  • App-aware rewriting: email reads like email, chat stays short, code stays terse
  • Spoken instructions: end with "make that formal" and it does
  • Cloud engine: the most accurate transcription HushQuill offers, using your own free API key
  • Dictation history: your last 500 takes, searchable, kept on your device
  • Snippets: say a trigger, get a saved block of text
  • Define your own rewrite styles in your own words
  • Formatting runs locally through Ollama by default
  • Every future update included, no renewal
Buy Pro

Buying Pro sends you a licence key by email. Paste it into Settings and Pro unlocks on up to 3 machines. The free build already does unlimited dictation, so you can start before you decide.

Free, no account, no sign-up. One small installer, and the speech models download from inside the app on first run so you pick what goes on disk. These links always serve the current build.

Both platforms warn you on first launch for now. The builds are not certificate-signed yet. On Windows, SmartScreen says the publisher is unknown: click "More info", then "Run anyway". On macOS, Gatekeeper blocks it because the app carries our own certificate rather than an Apple-issued one: open System Settings, go to Privacy and Security, and click "Open Anyway" next to the HushQuill message. If macOS instead says the app is damaged, that is the same block with a different message: open Terminal, run xattr -cr /Applications/HushQuill.app, then open it normally. Signing is planned on both platforms and the warnings go away when it lands. Updates the app installs afterwards are cryptographically signed and verified either way.

Questions

The things people actually ask.

Does anything I say leave my computer?
No, not on the default settings. Speech is transcribed by Whisper or Parakeet running locally on your CPU, and the AI formatting layer runs locally through Ollama. Two features can reach the internet and both are opt-in and off until you switch them on: the Pro cloud speech engine, and a cloud provider as an alternative for AI formatting. Both need an API key you supply yourself, from whichever provider you choose. The app also checks once on launch whether a newer version exists, which sends no personal data and no audio.
Does it work with no internet at all?
Yes. Once the speech model has been downloaded, dictation works with the network off entirely. That is the point of the app. The one exception is the optional Pro cloud engine, which is off unless you turn it on and add your own key.
Pro has a cloud engine. Does my audio still stay on my machine?
By default, yes, on every tier. Every engine HushQuill uses out of the box runs on your own machine and works offline. The cloud engine is opt-in: you have to choose it and paste your own API key, and only then is audio sent, to the provider you picked. It exists because the most accurate speech models are too slow to run on a normal laptop, and for Urdu and Hindi that gap is large enough that some people want the choice. If you never enable it, nothing about HushQuill touches the network except checking for updates.
Why do Windows and macOS warn me on first launch?
Because the builds are not certificate-signed yet. On Windows an OV code-signing certificate costs a few hundred dollars a year and is budgeted but not bought, so SmartScreen shows "unknown publisher": click "More info" then "Run anyway". On macOS the equivalent is an Apple Developer ID plus notarisation, which is planned but not yet done, so Gatekeeper blocks the app until you allow it: open System Settings, go to Privacy and Security, and click "Open Anyway". If macOS says the app is damaged instead, run xattr -cr /Applications/HushQuill.app in Terminal and open it again. Neither affects the app itself, and the updates it installs afterwards are cryptographically signed and verified regardless.
Is on-device dictation as accurate as the cloud ones?
Close, and it depends which model you run. The engines here are the same families the cloud services use, run locally instead of on someone else's server: Parakeet for English and 24 other European languages, Whisper for the rest. What we ship are the small and base Whisper models, not the largest one, so a cloud service running a large model will beat us on hard audio, heavy accents and noisy rooms. For ordinary dictation into an ordinary microphone the difference is small, and you can point the app at Groq's hosted large model with your own key if you want that ceiling for a specific job.
Will it understand my accent?
Usually, because the models were trained on a wide spread of speakers rather than on one. There is no per-user voice training here, and none is needed: nothing is tuned to you, so nothing has to be re-tuned when someone else uses your machine. Where accents genuinely bite is proper nouns and jargon, and the custom dictionary fixes those directly by mapping what the engine hears to what you meant.
Do I have to say "comma" and "full stop"?
No, and you deliberately cannot. The engines punctuate from your speech rhythm, so you just talk. Spoken punctuation was built once and then removed, because mapping ordinary words like "period" and "colon" to symbols corrupted normal dictation the moment anyone used those words as words. The only spoken commands are the ones that cannot collide with real speech: "new line", "new paragraph" and "scratch that".
Can I teach it names, jargon and spellings?
Yes, two ways, both free. The dictionary maps what the engine hears to what you meant, one pair per line, so colleagues' names and industry terms come out right every time. Snippets go the other way: say a short trigger and get a full block of text, for signatures, addresses and boilerplate. Neither sends anything anywhere, and neither requires the AI layer.
Which apps does it type into?
Any of them. Text is inserted the way a paste works, so editors, browsers, chat clients, terminals, Office and IDEs all just take it, and the floating bar never steals focus from what you are writing in. Two limits, both from the operating system rather than from us: Windows does not allow any normal app to type into a window running as administrator, and some apps ignore programmatic insertion. In every one of those cases the text goes to your clipboard and the bar tells you so, rather than vanishing.
Do I need a GPU, and what does it cost my machine?
No GPU. Everything is tuned for ordinary laptop CPUs with no graphics acceleration, and the fast engine runs at many times realtime on a plain office laptop. On Apple Silicon, Whisper uses the GPU automatically. The installer is 8.4 MB; the speech model is the real weight and it is downloaded from inside the app on first run, so you choose what lands on disk: 74 MB for Whisper tiny, 141 MB for base, 190 MB for the higher-accuracy model, about 640 MB for Parakeet. Memory tracks whichever model you loaded. Nothing runs between dictations, so it sits in the tray costing you nothing until you press the key.
Is there a free trial?
There is no trial, and there never will be, because there is no clock to run out. The free tier is permanent and does unlimited dictation. The only metered thing is AI rewriting, at 2 a day, and that resets daily rather than expiring.
Is there a subscription?
No, and there will not be one at any price. Pro is a single payment of $29 for the first 500 licences and $39 after that, valid on 3 devices, with free updates. Since the app costs nothing to run once you own it, charging rent for it would be indefensible.
Which languages does it handle?
Roughly 99, with automatic detection available. The fast local engine covers 25 European languages including English. Everything else, including Urdu, Hindi and Arabic, goes through the multilingual Whisper models in native script. Every language works on the free tier, uncapped. A free higher-accuracy option roughly halves the mistakes on Urdu and Hindi, and Pro adds a cloud engine that is better again.
How is this different from the popular cloud dictation apps?
Two differences that matter. Theirs send your audio to a server to be transcribed, so they need a working connection and you are trusting a privacy policy; this one transcribes on your own machine and works in aeroplane mode. And theirs charge monthly, forever, for something that costs nothing to run once it is on your laptop; this one is free for unlimited dictation and $29 once for the AI layer. On model size the gap has closed: Pro can use a cloud engine with your own free API key, so you can have the large hosted models when you want them and stay entirely offline when you do not.
Is the Mac version the same as the Windows one?
Yes. Same features, one codebase, built universal for both Intel and Apple Silicon. Whisper uses the GPU on Apple Silicon, so transcription is quick there. The two platforms release independently, so one can sit a version ahead of the other for a short while. Both download buttons above always hand you the current build for that platform.
What happens to my dictated text if I click the wrong place?
Nothing is lost. If no text box has focus when you stop speaking, the transcript is parked on your clipboard and the bar tells you so, with a button that types it into whatever you click next. The same button also repeats your last dictation on demand.