Parakeet TDT 0.6B v3670 MB
Kokoro 82M349 MB
Rewrite
How a take works
A take, start to finish.
Tap the hotkey to start and stop, or hold it while you speak. There is nothing to click and no window to find.
1
Press your hotkey
Ctrl+Alt+Space out of the box, or any combination you choose. The engine is held warm while the key is armed, so the first word does not wait.
2
Speak
A small listening pill shows tentative words as you talk, from a 20M-parameter streaming model on the CPU. You can see it hearing you.
3
It lands where you were
When you stop, the full take is decoded by Parakeet, pasted into the window that had focus, and your clipboard is put back the way it was. Every take is kept.
The engine
The fastest STT engine available.
Most dictation apps are a web app wrapped around a cloud API. Lite's voice engine is a compiled worker on your machine, and four small open-source models each do one job as the sound moves down the chain.
One compiled worker
The fastest path on your machine
Local mode never uploads
Sound inYour microphoneCaptured by Lite and handed straight to the engine.
TrimSilero VADFinds where speech starts and stops, so the decoder only sees words.
While you talkStreaming ZipformerLive tentative words in the listening pill, on the CPU.
When you stopParakeet TDTThe whole take becomes the final text, with punctuation and capitals.
OutPasted, and keptInto the focused window on Windows. Saved to Dictation history everywhere.
The models
Four small models, one job each.
Open weights you can look up. Small enough to stay loaded, good enough that you stop noticing them. The battery is what lives on your disk: about 716 MB before your first take, the speaking voice when you first use it, and a fifth slot if you want a rewrite pass.
Two small helpersStreaming Zipformer 20M · 44 MB · live words while you talkSilero VAD · about 2 MB · trims the silence
Rewrite, optionalYour own local model, or Automater Cloud
Turns the raw take into a clean prompt. Off until you turn it on.
Parakeet TDT 0.6B v3Speech to text · 670 MB · CC BY 4.0
The whole take becomes the final text, with punctuation and capitals. Downloads at first run with the two helpers, 716 MB together.
Kokoro 82MText to speech · 349 MB · Apache 2.0
Reads agent status and replies aloud in a voice you pick. Downloads when you first use speech.
Engine: automater-voice, on sherpa-onnx 1.13 and ONNX Runtime. First-run download for dictation is about 720 MB, resumable and checked by hash.
Compared
Unlimited dictation is a $12-a-month plan almost everywhere.
In Lite it is the default. The star scores each app on the five things that matter for dictation; the table underneath has the prices, billed yearly, with one-time licences marked.
The starFive axes scored 1 to 5 from the table; further out is better. Lite fills the whole pentagon; three competitors drawn over it, every app as a small star below.
Automater Lite · free, $0, local, three OSes, open models
Superwhisper · free local Whisper, $85, local or cloud, Mac and Windows
Wispr Flow · 2,000 free words a week, $144, cloud, Mac and Windows
Dragon Professional · no free tier, $699 once, local, Windows onlyScored from the table, not a benchmark: free to use (no cap 5, unlimited basic tier 4, capped tier 3, none 1), yearly cost ($0 is 5, under $100 is 4, $144 is 3, $699 is 1), local processing (always 5, optional 4, offline mode 2, cloud only 1), platforms (three desktop OSes 5, two 3.5, one 2), open models (named open weights 5, unknown 2, cloud model 1).
Free to useLite has no cap and no plan. Elsewhere the free tier is 1,000 to 2,000 words a week, or none.
Low yearly cost$0 against $85 to $144 a year, or $699 once. Cloud voice, if you want it, is $24.
Local processingOn your own hardware by default; the cloud only when you turn it on.
PlatformsWindows, macOS (Apple silicon) and Linux from one app.
Open modelsParakeet, Zipformer, Silero and Kokoro, named and downloadable.
LiteSuperwhisper
MacWhisper
Monologue
Willow
Aqua Voice
TypelessWispr FlowDragonThe numbersA year of unlimited dictation on each vendor's cheapest plan, billed yearly, 2026 prices. One-time licences marked.
| App | Unlimited dictation, 2026 | Monthly billing | Free tier | Where speech is processed | Platforms |
|---|---|---|---|---|---|
| Automater Lite | $0 | $0 | Everything. No word cap, no trial clock. | On your own hardware. | Windows, macOS (Apple silicon), Linux |
| Automater Cloud voice | $24 a year | $2 a month, billed yearly | Local dictation stays free and unlimited | Automater Cloud, metered, only when you turn it on | Windows, macOS (Apple silicon), Linux |
| Wispr Flow | $144 a year | $15 | 2,000 words a week | Cloud | Mac, Windows, iOS, Android |
| Willow | $144 a year | $15 | Basic model, 20 Scribe uses a week | Cloud, optional offline mode on Mac and iOS | Mac, Windows, iOS |
| Typeless | $144 a year | $30 | Limited words a week | Cloud | macOS, Windows, Linux, iOS, Android |
| Monologue | $144 a year | $15 | 1,000-word trial | Local option | Mac, iOS |
| Aqua Voice | $96 a year | $10 | 1,000 words | Cloud | Mac, Windows, iOS, Android |
| Superwhisper | $84.99 a year | $8.49 | Local Whisper models, unlimited | Local or cloud | Mac, Windows, iOS, Android |
| MacWhisper | €64 once | One-time | Free tier | Local | Mac |
| Dragon Professional v16 | About $699 once | One-time | None | Local | Windows |
Prices from each vendor's pricing page on October 22, 2026, in USD unless shown otherwise. Yearly figures are the annual-billing rate times twelve. Dragon's price is a reseller list price; Nuance publishes none. Prices change, so check the vendor before you buy. Logos identify their owners' products; the marks stay theirs.
The details
The details that make it stick.
Tap or hold
Toggle mode starts and stops on a tap. Push-to-talk listens only while the key is down. Switch in the Voice tab.
Any combination
Ctrl+Alt+Space is the default. Record your own chord, including the Windows key, and Lite keeps it.
Words while you talk
Tentative text in the listening pill, replaced by the final take the moment you stop. The pill never steals focus from where you are typing.
Your clipboard comes back
The paste goes through the clipboard and then restores whatever you had copied before. Nothing of yours is lost.
Every take kept
Dictation history stores each take with the app it went to and when. Copy any of them again, days later.
Warm when armed
While your hotkey is registered, Lite keeps the engine held in memory, so the first take of the morning is as quick as the tenth.
PRO
Local by default. Cloud when you want it.
Everything above is free and stays free. Automater Cloud voice adds the cloud for $24 a year: route speech to text and text to speech through Automater Cloud when this machine is busy or slow, on a metered allowance you can see in the app. You turn it on per device, and local mode stays the default.
Upgrade to PRO$24 / year for Cloud voice
If a cloud take fails, the local engine finishes it. PRO also adds chat and computer control from the tray, managed sessions, Fleet, built-in tools and pop-out mini-apps.
A
×
PROVoice$24 / year
Automater Cloud voice
Route voice through Automater Cloud when this PC is busy or slow.
- Route speech-to-text and text-to-speech through Automater Cloud.
- See trusted remaining balance and renewal period when available.
Allowance this year84% left
Upgrade to PRO
Use local Parakeet speech-to-text and local Kokoro text-to-speech.
Dictation history14
IN
refactor the retry queue to use exponential backoff and add a test for the jitter
IN
explain why the payments test is flaky
IN
open a pull request for the cart fix and ask for review
Automater Lite v0.1.33
Platforms
What works where, today.
The engine and the models are the same on every OS. Pasting into other apps is Windows-only for now, and this page will say so until it is not.
Windows
Windows 10 and 11, x64
- Local engine and all four models
- Global hotkey, tap or hold
- Floating listening pill with live words
- Pastes into the window that had focus
- DirectML or CUDA when the machine has them
- Dictation history
macOS
Apple silicon
- Local engine and all four models, signed and bundled
- Global hotkey opens Lite's Voice window
- Takes land in the Voice tab with Copy
- Pasting into other windows: not yet
- Floating pill: not yet
- Dictation history
Linux
x64, .deb and AppImage
- Local engine and all four models, statically built
- Global hotkey opens Lite's Voice window
- Takes land in the Voice tab with Copy
- Pasting into other windows: not yet
- Floating pill: not yet
- Dictation history
Questions
Questions people ask before they press the key.
Short answers. The long ones are in the app's Voice tab.
Is dictation really free?
Yes. No word cap, no trial clock, no account. PRO is a separate yearly upgrade that adds cloud voice and other things that act on your behalf. It never removes local dictation.
Does my audio leave this machine?
Not in local mode, which is the default. The models run in a native engine on your own hardware. Audio goes out only if you are on PRO and you turn Cloud voice on, and that choice is per device.
Which languages?
English today. The speaking voice, Kokoro, covers several languages. More dictation languages follow the model catalog; when they land, they land here.
What does it download, and what does it need?
About 720 MB the first time you turn dictation on: Parakeet, the streaming model and the voice activity model. The download resumes if it is interrupted and every file is checked by hash. It runs on an ordinary CPU; a supported adapter on Windows makes it quicker.
Can I change the hotkey, or hold it instead of tapping?
Both. Record any combination in the Voice tab and pick toggle or push-to-talk. The setting is yours; nothing resets it.
Does it work in every app?
On Windows the text is pasted into whatever window had focus when you pressed the key: editors, terminals, chat boxes, forms. On macOS and Linux, takes land in Lite's Voice tab to copy from, until pasting arrives there.
My OS already dictates. Why this?
Windows voice typing sends your audio to an online service. Lite's runs on your hardware with models you can name, keeps a history of every take, lets you pick the key and the mode, and is the same voice layer your agents use to speak back.
Does it clean up filler words or rewrite?
No rewriting. You get what you said, with punctuation and capitals. Your AI tool is on the other side of the paste; tell it what you want done with the words.
Dictation is the part of Lite that listens.
The sidekick to all your AI agents. The boss of your AIs. Free, local, no sign-up required.
Windows, macOS (Apple silicon), Linux. Free. No account.