or Ctrl + Win + Space
01 · Hold
Hold Ctrl + Win
Press it wherever you can type. Prefer hands-free? Tap Ctrl + Win + Space and talk until you're done.
Hold Ctrl + Win and talk the way you think. Saydraft tidies your words on your PC and types them wherever your cursor is.
Windows 10 & 11 · Unsigned portable · Offline after setupUnsigned Windows preview · Offline after setup
Fig. 1 — hold, speak, let go
Hold to talk · release to type
um, let's meet at two… actually, make that three. just a quick coffee?
↓ Cleaned on this PC
Let's meet at 3.Just a quick coffee?
Illustrative example · no microphone
audio uploads. Recognition and cleanup both run on your PC.
Free and MIT-licensed. No subscription, no word quota.
highlighted languages, plus 25 European via Parakeet v3.
is enough. No graphics card required to dictate.
No new app to learn. Saydraft sits in your tray and works in any editable Windows app — mail, docs, chat and your code editor.
or Ctrl + Win + Space
01 · Hold
Press it wherever you can type. Prefer hands-free? Tap Ctrl + Win + Space and talk until you're done.
so, uh, I was thinking —
actually, let's push the
launch to Friday
02 · Talk
Pause, restart, change your mind. Speech recognition runs on your processor — no graphics card needed.
Let's push the launch to Friday.
03 · Let go
Saydraft tidies the transcript and types it into the app you were using. Shift + Alt + Z pastes it again.
After recognition, Saydraft removes hesitations, applies your corrections and formats lists — locally. Raw mode keeps every word.
Um, let's meet at two… actually, make that three. Just a quick coffee?
Clean — what you get
Let's meet at 3.
Just a quick coffee?
Tone per app, your own vocabulary, short spoken cues — the details that make dictation sound like you, kept on your PC.
Styles
Choose tone and cleanup intensity for each app group, and preview the result before you save.
“cooper netties”→ Kubernetes
Dictionary
Add names, protected terms and replacements. Automatic learning is off by default.
“my signature”
Best,Snippets
A short spoken cue expands into your saved text — your exact sign-off, intact.
we should probably maybe ship it
“make it more direct”
We should ship it.
Command mode · Labs
Select text, hold Ctrl + Win + Alt and say the change. Experimental, off by default.
Move stand-up to ten tomor
Live preview · Labs
An optional English preview model shows partial words as you speak. Experimental.
Right at your fingertips
Coming from another dictation app? Change any shortcut in Settings.
Speech recognition and cleanup run locally. Installed models keep working with Wi-Fi off — and you can check that yourself.
01 · Input
Your microphone
02 · This PC
Parakeet v3 — speech to text
Cleanup model — llama.cpp
Your dictionary & snippets
03 · Output
Your cursor, any app
Wi-Fi off still dictating
Fig. 2 — where your words go
How Saydraft compares with the dictation tools people most often switch from.
| Feature | Saydraft | Wispr Flow | Typeless | Windows voice typing |
|---|---|---|---|---|
| Dictation connection | Offline with compatible models installed Project privacy guide | Internet required for transcription Flow documentation | Internet required for transcription Typeless FAQ | Standard Win+H uses online recognition Microsoft documentation |
| Windows availability | Windows 10/11 development source Project setup status | Windows app Flow platforms | Windows app Typeless platforms | Built into Windows Microsoft feature guide |
| Free use | Free, open source; no subscription quota Project license and limits | Limited weekly words; paid unlimited plan Flow plans | Limited weekly words; paid unlimited plan Typeless plans | Included with Windows Microsoft feature guide |
Checked 2026-11-14 against each product’s documentation. Features and plans change. Not affiliated with Wispr Flow, Typeless or Microsoft.
Parakeet v3 covers 25 European languages. One shared “All languages” model adds the rest. Install once, then dictate offline.
Coverage and measured accuracy are separate. Review results for your accent, names and recording conditions.
07 — Questions
Still curious? Read the docs or open an issue on GitHub — the whole project is in the open.
Speech recognition and text cleanup run locally. Saydraft has no cloud transcription, telemetry, or account. Model downloads happen only when you start them during setup. Optional update checks are off by default and request public release information from GitHub without your audio or text.
Yes, once: choose the recommended models, review their sizes and licenses, and start the setup download. You can pause and resume. Later launches reuse the installed models offline.
Highlighted choices are Arabic, Bengali, Brazilian Portuguese, Chinese (Mandarin), Filipino / Tagalog, French, German, Hindi, Indonesian, Japanese, Korean, Malay, Persian (Farsi), Spanish, Tamil, Telugu, Thai, Urdu, Vietnamese. Parakeet v3 covers 25 European languages; the other highlighted choices use the optional shared All languages model. There are no separate language packs. Choosing a language starts no download.
Speech recognition can run on the CPU. Hardware and model choice affect speed; optional AI cleanup requires more resources. The performance check helps choose a suitable setup, including rules-only cleanup.
Saydraft inserts text at the cursor in editable Windows apps. Password fields are excluded. Some protected or elevated apps need extra care; start with an ordinary text editor.
Choose a tone and cleanup intensity for each app group, add names to your dictionary, and create spoken snippets. Raw mode preserves your transcript without prose cleanup.
Saydraft is free and open source under the MIT license. Speech and cleanup models have their own licenses, shown before download and listed in the project credits.
No. The example is a canned simulation. This website does not request microphone access, upload audio, or use tracking analytics.
No. Saydraft has no subscription word quota. Each dictation has a configurable recording cap, 20 minutes by default, with a warning at 19 minutes. That duration cap is separate from the words you can dictate across sessions.
Meeting notes are an experimental Pro hardware-profile feature with explicit model setup, device selection and recording consent. Full live/frozen workflows, a 30-minute meeting and broader speech, speaker and summary accuracy remain open acceptance work. Ordinary Scratchpad notes are separate.
Make room for your next thought
Unsigned preview · Windows 10 & 11 · Setup and checksums ↗