Winpepper

Local, private voice dictation for Windows — hold a key, speak, and your words are typed into any app.

Download and install

Download Winpepper for Windows (64-bit installer)

If that link doesn't work, go to the latest release page and click the file ending in .msi.

Steps

1. Download the .msi file using the link above. It will land in your Downloads folder.
2. Double-click the downloaded file to start the installer.
3. Windows will show a blue "Windows protected your PC" box. Click More info, then click Run anyway.
4. Follow the installer. It takes a few seconds. You will not be asked for an administrator password — Winpepper installs just for you.

Why does Windows warn me?

Because the installer is not signed with a paid certificate. Certificates cost money every year, and Winpepper is a free, open-source project that doesn't have one yet. Windows shows that same warning for any unsigned app, new or old, harmless or not — it is not a virus alert, and it doesn't mean Windows found anything wrong.

Get-FileHash "$env:USERPROFILE\Downloads\winpepper-x64.msi" -Algorithm SHA256

First launch: one download of the speech models

Winpepper does its listening on your own PC, so it needs to fetch the speech files it thinks with — about 760 MB once (more if you add the optional backup or cleanup models).

1. Start Winpepper from the Start Menu.
2. A short setup walks you through picking your microphone and your hotkey, and offers a Download models step. (You can also do this later from the Models tab: click Download Missing Models.)
3. Wait for the download to finish. You need an internet connection only for this one download — after that, Winpepper works fully offline. If the download is interrupted, it resumes where it left off.

Until the models are downloaded, Winpepper will tell you "Speech model not installed. Open the Models tab to download it."

Live streaming: Winpepper also downloads the Nemotron Speech Streaming model (~760 MB, English by default — a multilingual variant is available; NVIDIA Open Model License for the English model, OpenMDW-1.1 for the multilingual one) automatically in the background on first run — dictation works immediately without it, and you can install or repair it yourself from the Models tab (or skip it entirely by turning streaming off there). Once installed it enables real local streaming — transcription runs while you speak and the text is ready almost the moment you release the hotkey. It uses the MIT-licensed transcribe.cpp engine (downloaded alongside the model, pinned and checksum-verified; see THIRD-PARTY-NOTICES.md). Without it, local dictations are transcribed in one pass after you stop — same results, just slower to appear. All local speech audio stays on your machine either way. Two practical notes: local dictation requires the Microsoft Visual C++ x64 Redistributable — both local engines need it (preinstalled on most PCs; if local dictation fails, install it from aka.ms/vs/17/release/vc_redist.x64.exe), and after your first streaming dictation the model stays loaded (~1 GB of memory) until you close Winpepper.

How to dictate

Hold Right Ctrl + Right Shift, speak, then let go. A moment later your words appear in whatever window you were typing in. Press Esc while it's working to cancel.

That key combination is the default; you can record your own during setup. There's also a second shortcut you tap once to start recording and tap again to stop, if holding keys down isn't comfortable.

What you need

• A 64-bit PC running Windows (nearly all PCs)
• About 1.5 GB of free disk space for a default install (roughly 700 MB for the app, 760 MB for the speech model); up to ~3 GB with the optional backup speech and text-cleanup models
• A microphone
• A graphics card or chip that supports DirectX 12 — nearly all PCs from the last several years. Without one, Winpepper still works, just more slowly.

Coming soon: winget install obra.Winpepper. The submission is under review by Microsoft; until it's accepted, use the download link above.

How to uninstall

Open Settings, go to Apps > Installed apps, find Winpepper, and choose Uninstall. No administrator password needed.

Your settings and dictation history are deliberately left behind so a reinstall picks up where you left off. To erase those too, delete the folder %LOCALAPPDATA%\winpepper (paste that into the File Explorer address bar).

What it does

• Types into any app. Whatever window has your cursor — email, chat, browser, Word, a code editor — that's where the text lands.
• Stays on your PC. Your voice and your words are processed locally. No account to create, no subscription, no telemetry, nothing sent anywhere once the models are downloaded.
• Cleans up what you said. A small local language model adds punctuation and capitalization and trims the "um"s. If you'd rather have the raw transcript instantly, turn it off on the Cleanup tab.
• Learns your vocabulary. Add names, jargon, and words it keeps mishearing on the Corrections tab so it gets them right next time.
• Keeps a history. Review past dictations on the History tab, and re-run one through different settings on the Lab tab.
• Lives in the tray. It starts hidden with Windows and waits quietly for your hotkey.
• Free and open source. Apache License 2.0. Read the code, change it, keep it.

Optional: if you'd rather use a cloud speech service, there's built-in support for AssemblyAI with your own API key. It's off by default — nothing goes to the cloud unless you deliberately turn it on.

Try it risk-free

Windows Sandbox is a feature built into Windows Pro and Enterprise editions that gives you a clean, throwaway copy of Windows in a window — anything you install inside it disappears when you close it, leaving your real PC untouched.

The scripts in scripts/windows-sandbox/ launch a sandbox, install Winpepper inside it, run the self-test, and show you the result. See scripts/windows-sandbox/README.md for the one command to run.

One catch: Windows Sandbox has no microphone, so you can confirm the app installs and starts, but not that dictation itself works.

If something goes wrong

Nothing happens when I hold the hotkey. Check that Winpepper is running (look for its icon in the system tray, near the clock), that the speech models finished downloading (the Models tab), and that the right microphone is selected.

The text takes several seconds to appear. That's usually the cleanup step, which leans on your graphics chip; it's much slower on built-in graphics than on a dedicated graphics card. Turn off Enable cleanup LLM on the Cleanup tab for near-instant results with slightly rougher text.

Windows warned me the app is unrecognized. Expected — see why above. Click More info, then Run anyway.

Something else is broken. Open the Diagnostics tab and click Copy diagnostics bundle. It creates a zip of logs and system information — never your audio, never your transcripts — that's safe to attach to a bug report. Please file one at github.com/obra/winpepper/issues.

Project status

Winpepper is alpha software: in active development, in real-world use, and improving quickly. Expect an occasional rough edge — and when you hit one, the Diagnostics tab builds a safe-to-share bug report bundle in one click (see above).

Quality is taken seriously. Over 1,400 automated tests run on every change, every release is installed and self-tested on a clean Windows machine before it is published, and the speech engine is verified against real recorded audio.

One thing worth knowing about how it's made: Winpepper is written almost entirely by an AI coding agent working under human direction, with people testing the results and reviewing the work.

For developers

Winpepper is a single .NET 9 / WinUI 3 app: a low-level keyboard hook, WASAPI audio capture, NVIDIA Parakeet TDT v3 speech recognition through ONNX Runtime, a small local Qwen model for cleanup, and SendInput to type the result.

• docs/DEVELOPMENT.md — architecture, repository layout, building from source (including the WinAppSDK XAML compiler workaround), known issues, and dev VM notes.
• docs/testing-windows-from-wsl.md — running the full Windows test suite from a WSL2 checkout.
• docs/windows-smoke-test.md — the long-lived real-hardware release smoke test.
• docs/release.md — the tag-triggered release process and winget submission.

License and credits

Apache License 2.0. See LICENSE. Copyright 2026 Jesse Vincent.

Companion to pepper-x — same problem, a native rewrite for Windows.