Dictation for Windows

Talk. It types.
Offline.

Hold a key anywhere in Windows, say your sentence, let go. The text is cleaned up and lands at your cursor, in whatever app you were already in. Download a free speech model once and every word is transcribed on your own PC: no API key, no connection, no quota, no per-month fee.

Download for Windows 14-day free trial · Windows 10/11 · signed installer

If Windows SmartScreen asks about this new release, click More info → Run anyway. The installer is code-signed and verified.

Hold, speak, release. Your voice never leaves the machine.

Dictate into

Anywhere Windows lets you type · no per-app plugin · no account

Three seconds, start to finish

One key. Any app.
No window to open.

Dictation is a global shortcut, not an app you switch to. Hold it while you talk and let go when you are done. Tap it once instead and it latches, so you can speak at length with your hands free: Enter inserts what you said, Escape throws it away.

Hold

Press and hold Alt+Shift+D wherever your cursor already is. Rebind it to anything you like in Settings.

Speak

Your microphone is read only while the key is down, and not a moment before or after.

Release

The transcript is tidied up and inserted at your caret. The app you were in does not need to know anything about KeptInFlow.

Alt+Shift+D

Dictate anywhere

Email clients, browsers, IDEs, chat apps, terminals: anywhere Windows lets you type. The text is pasted at your cursor, so there is no per-app integration to install or to break.

clean up

Sentences, not a raw stream

Grammar, punctuation, and filler words are fixed on the way in, and spoken commands like "new paragraph" are obeyed. Switch cleanup off and the raw transcript pastes instead, which is faster.

Alt+Shift+A

Speak a request, not just words

Hold a different key and say what you want done ("make this more formal"). Selected text is the target, and the reply arrives in a review card at your caret: Enter pastes it, C copies it, Escape throws it away.

Alt+Shift+Space

Talk to the command palette

Say what you want and the palette opens with your words already in the box and the best match picked out. Nothing runs until you press Enter, so a misheard word is two keystrokes to fix.

Dictation shows up wherever speech is the fastest input: a mic button on every AI note's prompt box, and the captions panel of a screen recording. Same engine, same models, same key.

Free models, downloaded once

The speech model lives
on your machine.

Take the model KeptInFlow recommends for your PC, or pick your own from Whisper and NVIDIA Parakeet. It downloads once, and dictation is free forever after that.

ModelDownloadWhat it is good for
Whisper Tiny 32 MB Fastest, runs on any machine. Rough accuracy.
Whisper Base 60 MB Fast, decent accuracy. Any language, or an English-only build.
Whisper Small 190 MB Good accuracy. Any language, or an English-only build.
Whisper Large v3 Turbo 574 MB Top Whisper quality, any language. Slow without a big CPU.
Parakeet v3 669 MB Best accuracy and fast. English plus 24 European languages.
no key

Nothing to sign up for

On-device dictation needs no API key, no account, and no subscription. Download a model and it works, including on a plane or behind a firewall that blocks everything.

auto

Language is auto by default

Auto-detect handles most speech. Pin a language if you want to, and Settings greys out any model that does not cover the one you picked.

latency

About as fast as the cloud

Ten seconds of speech is typed out in under two, which is what a cloud round trip costs anyway. Plain x64 CPU, no graphics card needed.

cloud too

A cloud engine, if you want one

Prefer a hosted transcriber? Point dictation at Groq, OpenAI, or Gemini with your own key instead. Left on Auto, it uses your local model whenever one is downloaded. On-device stays unlimited either way.

Built for the skeptical

Your voice stays here.

Your microphone opens when you hold the key, and not otherwise. Here is the whole of it.

The microphone opens when you hold the key

KeptInFlow reads your microphone while the dictation key is down, and not otherwise. There is no wake word, no always-listening mode, and no background recording of any kind for this feature.

With the on-device model, your audio never leaves the PC

The audio is transcribed by the model on your disk. No audio is uploaded, and there is no KeptInFlow server in the path. You can dictate all day with the network off.

No account, no telemetry, no voiceprint

Nothing to sign up for. Your voice is never turned into a voiceprint, and nothing you dictate is sold, shared, or used to train anything.

Cleanup is the one step that can call a model

Tidying the transcript into proper sentences is an AI step, so the finished text (never the audio) goes to whichever provider you configured. It is on by default. Turn cleanup off and the raw on-device transcript pastes with nothing leaving the machine at all; offline, cleanup skips itself.

Every feature in the app that reads anything is listed on one page: what it reads, whether it leaves your PC, and whether it is on the day you install. See what it watches

Looking at Wispr Flow and the rest

Dictation you buy,
not dictation you rent.

Good dictation apps exist. Almost all of them are a subscription, and almost all of them are only a dictation app. Those are the two things worth comparing.

A dictation subscription, next to this

The left column is the shape of the category rather than any one product. Prices move, so check them yourself.

A dictation subscriptionKeptInFlow
What it costs Wispr Flow Pro is $144 a year, every year (annual billing, checked July 2026) $59 once at the founding price, then $79, then $99
If you stop paying Dictation stops The app is yours. You keep the last version you got
Where your speech is transcribed Depends on the vendor, and it is the question to ask before you subscribe On this PC, with a free model you download once
How much you can dictate Whatever the plan allows Unlimited on-device, with no quota to run out of
With the network off Depends on the vendor Works. The model is on your disk
What else is in the box Dictation Meetings, OCR, capture, clipboard, text expansion, and system-wide AI, in the same purchase

The honest trade: a hosted transcriber can throw a much larger model at your audio than your laptop can. On short, ordinary speech that gap is small, and you can still send a run to a cloud engine with your own key when you want the bigger model.

The same one-time license

Dictation is one part of it.

You are not buying a dictation app with extras bolted on. Dictation is one tool in a Windows app that also records your meetings, reads your screen, and puts AI inside every app you type in.

No subscription required

Pay once. Own it.

Try every feature free for 14 days. No card to start.

App forever. Updates for 1 year. Keep the last version you got if you don't renew. This is not a stealth subscription.

Free trial

14 days

  • Every feature, no limits
  • On-device dictation models, free to download
  • 1,000 free AI actions included to try it without a key
  • No card up front. Nothing to cancel.
Download and start

License

$99 $59 one-time

Founding price for the first 100 · 1 year of updates included

  • Unlimited on-device dictation, no quota to run out of
  • Own it forever, no subscription
  • 1 year of updates included
  • Renew updates later for $29/yr, or don't, and keep what you have
  • 30-day refund, no questions asked
  • $229 lifetime-updates option
Buy KeptInFlow, $59

$59 for the first 100 buyers. Then $79 for the next 200. Then $99. Each step ends when it sells out, not on a timer.

Want every future update instead of one year? Lifetime updates, $229. Both are a single payment, and both come with the 30-day refund.

Don't want to deal with an API key? There is an optional plan that handles the AI for you: $12/month, no key to manage, 5,000 fast actions and 1,200 minutes of cloud dictation a month. It unlocks the full app while it runs, so it works on its own or on top of a license. On-device dictation and transcription stay unlimited either way.

Before you ask

Questions, answered.

Does dictation really work offline?

Yes, once you download a speech model. Turning your speech into text runs on your CPU, with no API key, no account, and no connection. One honest caveat: the cleanup pass that fixes grammar and filler words is an AI step, so with no network it is skipped and you get the raw transcript rather than an error. Switch cleanup off and offline dictation is the whole feature.

Is there a limit on how much I can dictate?

Not with the on-device model. There is no monthly quota, no minute counter, and nothing metered, because the transcription is happening on hardware you already own. The optional $12/month AI Included plan has 1,200 minutes of cloud dictation for people who would rather not manage an API key, and on-device dictation stays unlimited whether or not you take it.

Which apps does it work in?

Anywhere Windows lets you type: Outlook, Word, Slack, browsers, IDEs, terminals, chat apps. KeptInFlow pastes the finished text at your cursor, so there is no per-app integration to install or to break when the app updates. If you switch windows mid-sentence it leaves the text on your clipboard instead of guessing.

Which model should I pick?

Settings reads how much memory your PC has and marks one as recommended, so the usual answer is "the one already at the top". Start there and try a bigger one if you want more accuracy. Models are one click to delete if you change your mind.

How is this different from Windows voice typing?

Voice typing is built into Windows and free. KeptInFlow adds the parts around it: a choice of speech models you control, an AI cleanup pass that fixes grammar and filler words and obeys spoken commands, a spoken-request key that edits your selected text instead of transcribing it, and voice into the command palette. The rest of the app comes with it.

Can I speak an instruction instead of dictating words?

Yes. Alt+Shift+A records a request rather than text: hold it, say "make this shorter and friendlier", and your selected text is what it works on. The reply lands in a review card at your caret, so Enter pastes it and Escape discards it. Nothing is written into your document until you say so.

Do I need an API key at all?

Not for dictation itself, once the on-device model is installed. The AI cleanup pass does call a model, and the trial includes 1,000 free AI actions with no key of your own, unlocked by the email you enter in setup, so you can see the difference. For everyday use you add your own key from OpenAI, Anthropic, Google, xAI, or OpenRouter, and you pay your provider directly.

Windows only?

Yes, Windows 10 and 11. That is where the deepest integration is possible today. macOS is on the roadmap but not dated; the license will cover both when it arrives.

Say it once. Keep the app.

Download for Windows 14-day free trial · no account · 2-minute setup

If Windows SmartScreen asks about this new release, click More info → Run anyway. The installer is code-signed and verified.