You talk three times faster than you type. Scribe closes the gap, privately.
Hold a key, speak, release. Polished text lands at your cursor in any app on Windows 11. No required cloud. No account. No subscription. No audio ever leaves your PC.
Install the trusted, Microsoft-signed release and let the Store keep it updated automatically.
Download Scribe AI free from the Microsoft Store
⚡ ~¼-second response · 🏎️ transcribes ~30× faster than realtime · 🔒 On-device speech · 💸 $0 forever
If Scribe earns a place in your workflow, click Star at the top of this GitHub repository. It helps other people find private, offline dictation.
Scribe is a lightweight tray app that turns your voice into text anywhere on Windows: your editor, browser, chat, terminal, notes, email. Dictation apps usually make you choose. The accurate ones ship your voice to someone else's server and charge monthly for the privilege, and the private ones type like it's 2009. Scribe refuses the trade: a state-of-the-art speech model (NVIDIA Parakeet TDT 0.6b v3, the same family topping open ASR leaderboards) runs entirely on your CPU, decoding a sentence in the time it takes to lift your finger off the key. Measured on a desktop CPU: ~223 ms typical decode, real-time factor ~0.03×.
- 🔒 Private by architecture, not by promise. Audio is captured and transcribed locally, then discarded unless you explicitly enable local audio history.
- ⚡ Two keys, your choice. Hold Right Ctrl (or any key), talk, release. Add an optional second hotkey when you want dictation that always skips AI cleanup. Prefer hands-free? Toggle mode ends the dictation by itself when you stop talking.
- 🌍 Speaks your language. The bundled model transcribes about 25 European languages out of the box, no setup: dictate in English, German, Spanish, French, Italian and more, and it just works.
- 🧠 It understands how people actually talk. Say "send it Wednesday… I mean Thursday" and, with AI cleanup on, only Thursday survives. Repeat yourself and it writes the point once.
- 🔢 Numbers, dates and acronyms come out written, not spoken. "Twenty three licenses at three thirty p m on july third" becomes 23 licenses at 3:30 PM on July 3, the way an editor would write it, applied automatically.
- 🎭 Different apps, different voices. Per-app profiles give Outlook polished prose, Slack a casual tone, and your terminal one terse line, automatically, based on where your cursor is.
- ⌨️ Terminal-smart. Line breaks become spaces in terminals so a long dictation arrives as one message instead of firing Enter mid-thought. Built by someone who dictates into CLIs all day.
- 📖 Your vocabulary, your snippets. A dictionary locks in your jargon (
azure→Azure,dot net→.NET), imports/exports as CSV to share with your team, and even suggests terms from your own dictation history. Nine curated opt-in libraries cover Azure, Microsoft 365, GitHub, modern developer tools, .NET and C#, data engineering, data science, machine learning, and more. Say a trigger phrase and a whole saved template types itself. - 🧹 AI polish on your terms. Grammar and structure cleaned by an on-device model (fully offline), your Azure deployment, or any OpenAI-compatible server you already run (Ollama, LM Studio, OpenRouter…). Your models, your keys, your costs. Flip it on or off right from the tray.
- 📊 Performance you can verify. A built-in diagnostics panel computes latency percentiles from your own dictations, on your own disk. We don't ask you to take the speed claims on faith.
- 📈 Usage without surveillance. Track local dictation totals, speech time, active days, top apps, a trend chart and recurring terminology, and add uncovered terms to your dictionary with one click. AI insight is a separate explicit action and sends only aggregate totals and dictionary term labels to the provider you configured.
- 🪶 Stays out of the way. A tray app with a small glass recording pill you can place on any corner or edge of your screen, and a Windows 11-style settings app when you want to tune it.
Everything lives in a clean, Windows 11-style settings window: pick your microphone, speech model, and push-to-talk keys (hold or toggle), then browse focused sections for dictation behaviour, the overlay, AI cleanup, your dictionary, snippets, per-app profiles, the playground, history, usage and diagnostics.
Click a spot on the mini screen and preview the real pill at that position before you save. Optional silence auto-stop ends a toggle dictation when you go quiet.
Voice snippets expand a spoken trigger, like "insert my standup update", into a saved, multi-line template. Text-expander speed, no keyboard required.
Profiles adapt dictation to the app you're speaking into: the AI writing style and line-break behaviour switch automatically based on the focused window. First matching profile wins; everything else uses your global settings.
Turn on AI cleanup to have a language model fix punctuation, capitalization, sentence structure, spoken self-corrections and repeated points before the text is inserted. The default provider runs fully offline through Foundry Local. If the model isn't ready, dictation just continues with the raw transcript.
Point Scribe at a model you've already deployed in Microsoft Foundry: it signs in with your
existing az login, discovers your deployments, and lists them in a browsable dropdown (type
to filter) so you pick a model instead of remembering deployment names. Live in more than one
tenant? Choose service principal sign-in instead and Scribe authenticates as exactly the
identity you name, every time (setup guide). Or aim it at any
OpenAI-compatible endpoint: Ollama or LM Studio on localhost, vLLM on your homelab, OpenRouter,
or api.openai.com with your own key. Only the transcribed text is ever sent (never audio), and
only to the endpoint you configure. And when you want the raw transcript, toggle AI cleanup
straight from the tray menu with no settings trip required.
The dictionary replaces spoken words and phrases with the spelling you actually want, and feeds the AI cleanup a glossary of your preferred vocabulary. Build it in seconds: import a CSV your team shares, grab the self-documenting template, or let Learn from history spot the acronyms and product names you keep saying and add them for you.
The Diagnostics section computes latency percentiles from your own dictation history. Nothing is collected; it's your data on your disk. On a typical desktop CPU, Parakeet decodes at a real-time factor around 0.03×, which is ~30× faster than the audio itself.
History keeps your recent dictations reviewable and copyable, with per-entry audio if you opt in. Delete one entry or clear everything; it never leaves your PC either way.
The Usage section summarizes retained history across 7, 30 or 90 days, or all retained history. Every metric uses the same selected period. It shows totals, active days, speech time, top apps, a trend chart and recurring technical terms, and any recurring term your dictionary doesn't cover yet gets an Add button that locks in its spelling on the spot. Opening or refreshing Usage stays fully local. The optional AI insight button sends only aggregate totals and term labels already in your dictionary. Terms mined from your dictations but not yet in your dictionary stay on your machine, and it never sends transcript text, audio, application names or timestamps.
You'll need: Windows 11 (x64). That's it. The speech model is bundled, so there's nothing else to install.
- Open Scribe AI in the Microsoft Store.
- Select Install. Microsoft signs, delivers and updates the app through the Store.
- Launch Scribe AI. It appears in your system tray.
Prefer a portable build, or need the original standalone installer? Advanced users can still visit
the GitHub Releases page. Download
Scribe-win-x64-Portable.zip to run it from any folder, or
Scribe-win-x64-Setup.exe for the independently updated Velopack installation.
Windows security prompt for GitHub downloads: Direct GitHub releases are intentionally unsigned, so Windows may show an "Unknown publisher" or SmartScreen warning. The Microsoft Store version is signed by Microsoft and is the recommended installation.
Then hold Right Ctrl, say a sentence, and let go. The text lands wherever your cursor is. Right-click the tray icon for settings, one-click vocabulary learning, copying any of your last five dictations, and pausing or quitting. If a dictation ever fails to insert, Scribe notifies you and keeps the text ready to copy from the tray. Review history and local Usage from Settings.
- Hold your push-to-talk key. The glass pill shows it's listening, with a live level meter.
- Speak naturally. Voice-activity detection trims the silence around your words.
- Release. Scribe transcribes on your CPU, optionally polishes with AI (using the profile for the app you're in), applies your dictionary and snippets, and types the result into the focused app.
Everything is configurable from the tray: microphone, hotkey (hold or toggle), silence auto-stop, the pill and where it appears, voice-activity detection, line-break handling, per-app profiles, snippets, post-processing, start-with-Windows, and how text is inserted.
Dictation core
| Feature | What it does |
|---|---|
| Push-to-talk | Separate AI-capable and optional dictation-only hotkeys, each hold or toggle on any key or two-key chord, with a capture UI that pauses dictation while you rebind |
| Silence auto-stop | Toggle-mode dictation ends itself when you go quiet |
| On-device speech recognition | Bundled NVIDIA Parakeet TDT 0.6b v3 handles ~25 European languages automatically; optional verified Moonshine Base and Tiny downloads provide fast English-only alternatives |
| Recording pill | A glass WinUI 3 overlay with a live level meter, placeable on any of 9 screen anchors with an on-screen preview |
| Smart text injection | Unicode or clipboard insertion with automatic fallback, and terminal-aware line-break flattening so newlines never fire Enter |
| Hotkey self-healing | Detects and repairs stuck modifiers and silently removed keyboard hooks, so push-to-talk keeps working across long sessions |
Text quality
| Feature | What it does |
|---|---|
| Personal dictionary | Spoken-form to replacement rules with whole-word matching, CSV import/export, and history-mined suggestions |
| Dictionary libraries | Nine curated opt-in libraries, including dedicated .NET and C#, data engineering, and data science and machine learning packs, plus your own custom CSV libraries |
| Voice snippets | A spoken trigger phrase expands into a saved multi-line template |
| Per-app profiles | Writing style and line-break behavior switch automatically based on the focused app |
| AI cleanup | Optional polish through Foundry Local (fully offline), Microsoft Foundry (your az login or a service principal), or any OpenAI-compatible endpoint; benchmark-validated prompts, your dictionary as a glossary, and raw-transcript fallback if the model misbehaves |
| Playground | Captures normal push-to-talk dictation, then shows raw recognition, final highlighted dictionary, library, and snippet replacements, and per-step timing |
Insight and recovery
| Feature | What it does |
|---|---|
| History | Retained dictations with optional audio, replayable and deletable, all local |
| Usage insights | Totals, speech time, active days, top apps, a trend chart, and recurring terminology with one-click add to dictionary |
| Dictation recovery | Your last five dictations stay copyable from the tray, and a failed insertion notifies you instead of losing text |
| Diagnostics | P50/P95 decode latency and real-time factor computed from your own history |
| AI usage insight | Opt-in, explicit, and aggregate-only: sends totals and dictionary term labels, never transcripts, audio, app names, or timestamps |
App
| Feature | What it does |
|---|---|
| Tray quick actions | Pause, AI cleanup on/off, learn from history, copy recent dictations, reopen the welcome tour |
| Auto-updates | Microsoft Store installs are signed and updated by Microsoft; standalone GitHub installs use Velopack delta updates |
| Offline by architecture | The dictation path needs no network, sends no telemetry, and keeps every stat on your disk |
Numbers below come from the checked-in benchmark reports, reproducible with the commands in each document. Speech decode runs on CPU; your Diagnostics page shows the same percentiles for your own hardware.
| Path | Measurement |
|---|---|
| Speech decode (typical desktop CPU) | ~223 ms typical, real-time factor ~0.03× (about 30× faster than the audio itself) |
| 10-second audio aggregation | 69 µs and 625 KB allocated (was 164 µs and 2.6 MB before the 0.2.1 hot-path work) |
| 48k-character cleanup chunking | 30 µs and 191 KB allocated (down 21% time, 49% allocation) |
AI cleanup, fully offline (phi-4 via Foundry Local) |
~1.6 s median added latency, best on-device quality grade |
AI cleanup, cloud default (gpt-5.4) |
~1.8 s median added latency, grade B+ across the 46-model golden suite |
Details and methodology: the model leaderboard (52 models against Scribe's real cleanup pipeline with a golden-reference judge), the GPT-5.6 phonetic benchmark (sound-alike transcript challenges and prompt A/B results), and the local performance benchmark (BenchmarkDotNet, production code paths).
- Audio never leaves your machine. Ever. It is captured and transcribed locally, then discarded unless you explicitly enable local audio history.
- Transcription is 100% local (Parakeet via sherpa-onnx on CPU).
- AI cleanup is optional and yours to control. The on-device provider (Foundry Local) is fully offline. If you choose Azure or a custom endpoint, only the transcribed text (never audio) is sent to the server you configure, under your credentials.
- Even the stats are local. Performance and Usage are computed from history already on your disk. Usage AI insight runs only when you click it and sends bounded aggregate data without transcripts, audio, application names or timestamps.
See the full Scribe AI Privacy Policy for data storage, optional transmissions, security, retention, and user controls.
For help, feature requests, or bug reports, use GitHub Issues. Do not include transcripts, audio, credentials, or other sensitive information in a public issue.
You'll need: Windows 11 (x64) and the .NET 10 SDK.
git clone https://github.com/ChrisMcKee1/scribe.git
cd scribe
# One-time: fetch the speech model (~670 MB). The installer ships with it bundled,
# but the model files are too large to live in git, so source builds download it once.
pwsh ./scripts/Download-Models.ps1
dotnet build Scribe.slnx -c Debug
dotnet run --project src/Scribe.AppWant to contribute? Everything else you need (project layout, code style, tests, the pull-request workflow, the AI-cleanup eval harness, and how releases are packed) lives in CONTRIBUTING.md.
Preparing a Store release? Use the checked-in Microsoft Store submission checklist for Partner Center answers, listing copy, screenshots, certification notes, and package validation.
Scribe is released under the MIT License.
It stands on the shoulders of excellent open work:
- Parakeet TDT 0.6b v3: © NVIDIA, CC-BY-4.0
- Moonshine: © Useful Sensors, MIT
- sherpa-onnx: Apache-2.0 (Next-gen Kaldi / k2-fsa)
- Silero VAD: MIT
Download Scribe AI
Built for people who'd rather talk than type. 🎙️