FastPen is fast, private dictation for Windows. Press a hotkey, speak, and your words appear in any app — email, Word, browser, chat.
Download for Windows — free 🏪 Microsoft StoreSpeech recognition runs on your computer. No cloud, no account, no subscription — your voice never leaves your machine.
Whisper turbo: the most accurate, great for mixed-language dictation too. Parakeet: the fastest, 25 languages with automatic detection. Both download in one click, right from the app.
Say “enter” — and Enter is pressed. “New line”, “tab” and “backspace” work too: dictate and send entirely hands-free. Since v0.3, commands are recognized together with their speech context — far more reliably.
An optional local AI (Ollama) polishes the transcript — with whole-sentence context. Light stays faithful to your words; Compose rewrites more boldly: it resolves self-corrections, formats lists, reconstructs garbled words. This runs on your machine too.
Text appears the moment you pause. Six seconds of speech is transcribed in about 0.2 seconds on a modern CPU.
FastPen types straight into whatever app has focus. It's your CPU — no word counts, no monthly quotas, no premium tier.
Press Ctrl+Alt+Space
Speak naturally — pause whenever you like
Your words are typed into the active window
Microphone distance matters more than any setting. A boom headset is the best choice: the distance stays fixed, so the quality stays consistent. We test FastPen with a Poly Blackwire 3210 USB headset.
Whisper turbo delivers the most accurate text, with the recognition language pinned — and handles mixed-language dictation too. Parakeet is the fastest: ideal for slower machines and voice commands.
Say voice commands on their own: sentence — short pause — “enter”. The command arrives as its own segment and is recognized reliably.
With a headset mic, set the VAD threshold to 0.30–0.40 in Settings (in our measurements 0.30 also catches short, snappy commands reliably); with a desktop microphone the default 0.50 works well.
FastPen contains no analytics, no telemetry and makes no network requests other than downloading the speech model from GitHub on first start. What you say is processed in memory, typed into your app, and forgotten.
And private doesn't mean compromised: in our own informal test on identical reference audio, FastPen's local recognition tied the built-in cloud dictation of Windows on meaningful errors.
FastPen is under active development. A few things on the roadmap:
Behavior tailored per window — different settings for chat vs. a text editor.
How much you dictated and how much typing time you saved — measured.
The raw transcript appears in the overlay instantly, and the polished version slides into place — no waiting feeling.
FastPen is free and built by one developer. If it saves you time, you can support development with a coffee. 💛
☕ Support me on Ko-fiFastPen is completely free. Donations are entirely voluntary and don't unlock any features, licenses or extra support — just a thank-you that helps fund future development.
Yes. FastPen is free to download and use, with no feature gates. If it becomes popular, a paid version with extra features may follow — dictation will stay free.
Yes. FastPen is free for business and commercial use too — including AI cleanup and every other feature, with no restrictions. See LICENSE.txt in the package for details.
Same app, two channels. The Store build installs signed by Microsoft (no SmartScreen warning) and updates automatically. The portable ZIP lives in a single folder, no installer — and usually gets new versions a few days earlier, since Store updates go through Microsoft's review. One difference: the Store build can't install Ollama (used for AI cleanup) by itself — there you install Ollama manually from ollama.com (the ZIP build handles that on its own).
No. Recognition runs entirely on your CPU using the open Parakeet v3 model. FastPen works offline once the model is downloaded.
The Parakeet model covers 25 European languages including English, German, French, Spanish, Italian, Polish, Hungarian, Czech, Romanian and Ukrainian, with automatic detection. The Whisper turbo model (downloadable from the app) adds even more languages, and the language can be pinned.
Whisper turbo delivers the most accurate text — we recommend it for prose and for mixed-language dictation alike, with the recognition language pinned. Parakeet (the default) is many times faster and recognizes voice commands most reliably — pick it on slower machines. Both models are downloaded by the app when you select them (~1 GB and ~465 MB).
Said on their own, with a short pause before and after: “enter” → Enter; “new line” → Shift+Enter; “tab” → Tab; “backspace” → deletes the last word (along with trailing punctuation). Since v0.3 the Parakeet model recognizes commands together with the preceding speech context, making them notably reliable mid-dictation.
Primarily on machines with an NVIDIA GPU — Ollama uses it automatically and corrections are near-instant. It works on CPU too, but slowly (several seconds per sentence depending on your hardware); if you try it on CPU, start with the smaller gemma3:4b model. For dictating to AI agents we recommend leaving it off — agents understand raw text just fine.
Light stays faithful to your words: it punctuates and drops fillers but never rephrases — best for notes, where what you actually said matters. Compose works more boldly: it collapses self-corrections (“on Monday — no wait, Tuesday”), reconstructs garbled words from context, and formats enumerations as lists — best for emails you’ll proofread before sending anyway. Switch anytime in Settings.
That's one of the main use cases: press the hotkey (push-to-talk works great), speak your prompt, then say “enter” — and the message is sent hands-free. For mixed-language prompts use Whisper turbo. Agents understand the raw transcript just fine, so you can leave AI cleanup off.
Windows 10 (version 1803 or newer) or Windows 11, 64-bit. Any reasonably modern CPU works — no GPU needed. About 700 MB of disk space with the Parakeet model (about 1.7 GB with Whisper turbo as well). A separate build for older 32-bit Windows is also available — use the Parakeet model there.
Not recommended for dictation. When the earbuds' microphone is active, Bluetooth switches to the hands-free profile with heavily compressed, phone-call-quality audio — recognition accuracy drops noticeably. On Windows 10 there is no way around this (the higher-quality LE Audio is only available on Windows 11 with compatible earbuds). Use a wired or USB boom headset instead. If you do use Bluetooth: untick “Enable audio enhancements” in the microphone's advanced settings.
Not currently — but you don't have to take our privacy claims on faith. FastPen makes no network requests except downloading the model on first run, and that's verifiable: watch it with a firewall, or simply block it from the internet entirely. It keeps working offline.
FastPen is a new app from a small developer and isn't code-signed yet, so SmartScreen may show a warning. Click “More info” → “Run anyway”. Code signing is planned. To skip this entirely, install the Store build — it arrives signed by Microsoft, with no warning.