Voice dictation for Windows. Hold a hotkey, speak, let go — your words are typed wherever your cursor is: an email, a chat, a document, a code editor, a browser.
- Fast and accurate. Your voice goes to a speech model you choose — Groq (free), ElevenLabs (paid, the most accurate) or one that runs on your own PC — and the text lands a moment after you let go.
- AI polish. Each dictation is tidied up the way Wispr Flow does it: punctuation, grammar, false starts, the "um"s — without adding words you didn't say.
- Free with one free key. No subscription and no account with us: you bring your own key from Groq, whose free tier covers everyday dictation.
- Private. No sign-up, no tracking. Your history stays on your PC.
- Learns your words. Fix a word once and Scribe learns it; names you say often are learned too. No list to keep up with.
- Download Scribe.zip.
- Extract it to
C:\. Right-click the ZIP → Extract All… → change the folder toC:\→ Extract. You get aC:\Scribefolder. (Any folder you'll keep works — just not one inside OneDrive, which would sync all of Scribe's files.) - Double-click
setup.batinC:\Scribe.- Windows may double-check a file from the internet: click Run — or, if it says
"Windows protected your PC", More info → Run anyway. (
setup.batis a plain text script — open it in Notepad to see what it does.) - It installs everything Scribe needs — including its own copy of Python if your PC doesn't have one — then starts Scribe. The first time takes a few minutes.
- Windows may double-check a file from the internet: click Run — or, if it says
"Windows protected your PC", More info → Run anyway. (
- The welcome opens. Pick Groq (free), click Open Groq's API keys page, sign in, click Create API Key, copy it and paste it into Scribe. It's checked on the spot.
That's it. Hold Ctrl + Win, say something, let go.
Scribe starts when you sign in to Windows and waits quietly in the system tray (bottom right, near the clock). Double-click its icon — or open Scribe from the Start menu — to see the dashboard.
| Do this | To |
|---|---|
| Hold Ctrl + Win, speak, release | Dictate. The text appears at your cursor. |
| Ctrl + Win + Z | Undo the last dictation |
| Say only "scratch that" | Undo the last dictation, by voice |
| Say only "fix that" | Polish the last dictation again |
| Say only "new line" / "new paragraph" | Insert a line break |
| Double-click the tray icon | Open the dashboard |
| Right-click the tray icon → Quit | Quit |
While you talk, a small capsule at the bottom of your screen shows Scribe listening — a wave flowing into a text cursor, as lively as your voice. When you let go it settles into a line while Scribe types, then folds away.
Voice commands work when they are the whole dictation — tap the hotkey, say just the command, release. You can change the hotkey in Settings; if you use Windows' virtual-desktop shortcuts (Ctrl + Win + ←/→), pick Ctrl + Alt instead.
A few things Scribe takes care of for you:
- Your text lands where you started. Click into another window while Scribe is transcribing and the text still goes to the window you were dictating into. If that window is gone, the text waits on your clipboard and a notification says so.
- Back-to-back dictations arrive in order, even if you start the next one early.
- Windows shortcuts still work. Holding the hotkey and pressing any other key cancels the dictation instead of recording it.
- Your clipboard stays yours. Scribe borrows it for a moment to paste, then puts back what you had — and dictations never show up in your clipboard history (Win + V).
- Long dictations can run up to 5 minutes.
It follows Windows' light or dark mode, or pick one in Settings → Appearance.
- Home — what Scribe is doing right now, then your latest dictation: the text, the app it went to, your pace, how fast the text appeared, and whether AI polish changed it (Show what you said shows your own words). Below: today, the last 30 days and your recent dictations — click one to copy it.
- Insights — a quiet report on how you dictate: speaking pace, how fast text appears, how often polish helps, the hours and apps you dictate in, the words Scribe learned for you, and the words you use most. Built only from your own history, on your PC.
- Dictionary — the words Scribe learned for you (remove any with one click), plus anything you'd like to add by hand.
- Settings — everything Scribe can do, including your keys, AI polish and updates.
| Groq (recommended) | ElevenLabs | On this PC | |
|---|---|---|---|
| Cost | Free tier (covers everyday dictation) | Paid plan (Starter is enough) | Free |
| Speed | About a second after you let go | Under a second — it listens while you talk | A few seconds, depending on your PC |
| Accuracy | Very good | The best | Good |
| Your audio | Sent to Groq | Sent to ElevenLabs | Never leaves your PC |
| AI polish | Included (same key) | Add a free Groq key | Not available |
Pick one in the welcome, or later in Settings → Cloud. Every key is checked with its service before it's saved, and kept in Windows Credential Manager — never in a file, never shown in full.
- Groq — get a key at console.groq.com/keys (sign in with Google, Create API Key). Silence is trimmed out before anything is sent, and a recording with no speech is never sent at all.
- ElevenLabs — get a key at elevenlabs.io → Developers → API Keys, with Speech to Text access (and optionally User → Read, so Settings can show how much of the month is left). Add a Groq key too and Groq becomes the backup whenever ElevenLabs can't answer.
- Offline backup — with a cloud service, Scribe also keeps a speech model on your PC (about 480 MB, downloaded in the background) for when the internet is down. Switch it off in Settings if you'd rather not.
AI polish (on by default with Groq) cleans each dictation up — grammar, punctuation, false starts — keeping your meaning, tone and details, and it never adds words you didn't say: an answer that does is thrown away and your own words are typed instead. Choose Light (punctuation, capitals and the "um"s only — every other word stays) or Full (also smooths rambling) in Settings → Cloud → Polish style. Takes of three words or fewer are left alone, and your Dictionary always has the last word.
Scribe learns your names and jargon by itself, so there's nothing to keep up with:
- Fix a word once. If Scribe types "cal she" and you correct it to "Kalshi" right after, Scribe notices and learns "Kalshi" — a small notice tells you — and your voice model listens for it from then on. When what Scribe typed isn't a real word ("Koushi"), Scribe also swaps it for your spelling every time; it never swaps everyday words, so "their" can't turn into "Theo". This works in browsers (Gmail, ChatGPT, Slack on the web…), Office and most chat apps; a few apps don't share their text (the VS Code editor, games, remote desktops), and there Scribe simply doesn't learn.
- Say it often. A name or acronym you say at least five times (Webull, SCHD) is added to the words Scribe listens for — quietly. Everyday words and likely mishearings are left out.
The Dictionary page shows what Scribe learned and where each word came from. Remove one with ✕ and it's gone for good — Scribe won't learn it again. You can still add words or fixes by hand ("when Scribe types… write this instead"). Switch learning off in Settings → Dictation → Learn my words automatically. Changes apply from your next dictation.
Once a day Scribe checks whether a newer version is out and tells you (with On this
PC chosen, that check starts switched off — turn it on in Settings → About). To
update, open
Settings → About → Update now — Scribe closes, updates itself and opens again in
about a minute. (Or double-click update.bat in Scribe's folder.) Your history,
settings and keys are never touched by an update.
- Scribe has no account, no analytics and no tracking. Nothing is sent to the author of Scribe, ever.
- Your audio goes only to the voice model you chose, only while you dictate. With On this PC, nothing leaves your computer.
- With AI polish on, the text of each dictation is sent to Groq to be tidied.
- Once a day Scribe asks GitHub for the latest version number (switch it off in Settings → About; with On this PC it starts switched off).
- To learn from your fixes, Scribe reads the text box it just typed into — only that box, only for a minute, only on your PC — and saves nothing but the corrected word. Password boxes are never read. Switch it off in Settings → Dictation.
- Your keys are stored by Windows (Credential Manager), encrypted with your sign-in.
- With a cloud voice model, the words in your Dictionary (including the ones Scribe learned) go along with each dictation as spelling hints, and to Groq with AI polish.
- Your history, settings and Dictionary are stored only on your PC, in
%APPDATA%\Scribe.
Everything Scribe stores is in %APPDATA%\Scribe (right-click the tray icon →
Open data folder):
| File | What |
|---|---|
config.json |
Your settings (your API keys are not here) |
vocabulary.json |
Your Dictionary — including the words Scribe learned, and how |
dictation_log.jsonl |
Your dictation history, for the dashboard |
cloud_usage.jsonl |
Cloud usage, for the usage meters in Settings |
error_log.txt |
Errors and notices, if any |
Settings and Dictionary are saved crash-safely, with a last-good backup. If a file is ever damaged, Scribe keeps it, restores the backup and tells you.
Double-click uninstall.bat in Scribe's folder. It removes Scribe from sign-in, the
Start menu and the taskbar, then asks before deleting your history, your saved keys and
the downloaded speech models. Close its window and delete the Scribe folder to finish.
Scribe tells you when something goes wrong — with a short Windows notification that
says what it did about it — and writes the details to error_log.txt.
- "Windows protected your PC" (or a security warning) when running setup.bat — click More info → Run anyway (or Run). Scribe isn't signed with a paid certificate, so Windows doesn't recognize it.
- setup.bat says Scribe is still inside the ZIP — right-click the ZIP → Extract
All… →
C:\→ Extract, then runsetup.batinC:\Scribe. - setup.bat says the folder is inside OneDrive — move the Scribe folder to
C:\Scribeand runsetup.batthere. (OneDrive would copy thousands of Scribe's files to the cloud, and can lock them while Scribe installs or updates.) - setup.bat asks to install Microsoft's Visual C++ runtime — allow it: the speech libraries need it, and a freshly installed Windows may not have it yet.
- The install fails deep inside "onnxruntime" — the folder's path is too long. Move
the Scribe folder somewhere short, such as
C:\Scribe, and runsetup.batagain. - The welcome or the dashboard doesn't open — usually the Microsoft Edge WebView2 Runtime is missing (it comes with Windows 11). Scribe offers to open its download page; without it, Scribe sets itself up with the defaults and still works from the tray.
- Nothing happens when I hold the hotkey — look for a Scribe notification, check
error_log.txt, or quit Scribe from the tray and start it again from the Start menu. - Text doesn't appear in one particular program — programs running as administrator don't accept typing from normal apps (a Windows rule). Run that program normally, or paste: the text is also left on your clipboard.
- It records from the wrong microphone — Scribe follows your Windows default microphone, including when you plug one in. To always use one mic, pick it in Settings.
- Something else — open an issue and describe what happened. Please don't paste your API key or private dictations.
git clone https://github.com/Dimas-100/scribe.git C:\scribe
C:\scribe\setup.batvenv\Scripts\python app.py --show runs Scribe with a console that shows what it's
doing. There's no build step. The tests in tests/ are standalone scripts run from the
project root (venv\Scripts\python tests\storage_test.py); the safe ones use a temporary
data folder and mock the microphone, model and network. A clone updates with git pull
(or update.bat). CLAUDE.md explains the architecture and threading
model; ROADMAP.md lists ideas for what's next.
Requirements: Windows 10 or 11 (64-bit; ARM laptops run it through Windows' x64 emulation), a microphone, the WebView2 Runtime (built into Windows 11), Microsoft's Visual C++ runtime (setup.bat installs it if missing), about 0.5–2 GB of disk for a local speech model, and a 64-bit Python 3.11–3.13 (setup.bat fetches one if needed).
PolyForm Noncommercial 1.0.0 — free to use, change and share for any noncommercial purpose. Commercial use needs the author's permission.