Voice-to-text for Windows.
Beta. Unsigned. Testers can grab a build from GitHub Releases. This is not a signed store build and not a stable public release.
VocaWin is native Windows voice typing. After a Whisper or ONNX model is on disk, recording and speech-to-text stay on this PC. No Voca account, no hosted speech API.
It sits next to VocaLinux, VocaMac, and VocaPhone. The family directory is vocahq.com.
- On this PC - After the model download, transcription is designed to run locally
- No Voca cloud - There is no hosted speech service to sign up for
- Open source - The repository is public
- No telemetry in the product - The app does not phone home
- Windows native - Tray app, hotkey, WASAPI, text at the caret
- Honest status - Beta. Unsigned. Windows will likely say the publisher is unknown.
Testers can install an unsigned beta today. Download the NSIS .exe from GitHub Releases. MSI is paused while the version is 0.2.1-beta. Windows will likely say the publisher is unknown. That is SmartScreen. More info, then Run anyway if you trust the file. Read the setup guide first.
Want today's main instead of the last tagged Release? Use the nightly. Same unsigned NSIS, rebuilt when app source on main changes. MSI is paused there too. Prefer the latest tagged Release if you want the last cut we named.
This is a tester build you can run today. It is not a store listing and not a stable public release.
- Hold a hotkey, speak, text at the caret - Default is Right Alt, the same hold as VocaLinux. Double-tap toggles. You can change the hotkey in Settings.
- Tray - Idle, recording, and processing icon states. Close goes to the tray. Show window / Quit.
- Settings - Grouped pages with one search box: shortcuts, models, audio, formatting, dictionary, snippets, history, stats, and power. Export and import a settings backup.
- Local models - In-app Download for Whisper/whisper.cpp, Distil-Whisper, Voca Hinglish, Parakeet, Moonshine, SenseVoice, GigaAM, Canary, and Cohere Transcribe.
- GPU - whisper.cpp on Vulkan with CPU fallback. ONNX Runtime on DirectML with CPU fallback.
- Clipboard stays yours. Insertion types at the caret and does not replace what you copied. Turn on Copy to clipboard in Settings if you want the transcript left there. Clipboard paste is only a fallback, and that path restores the previous clipboard.
- More ways to dictate - Escape cancels a take. A hands-free shortcut starts and stops without holding a key. A middle or side mouse button works like the hotkey. A paste-last shortcut types your last dictation again.
- On-screen pill - A "VocaWin is ready" pill for a few seconds after launch, and a live level while you speak. It never takes focus.
- Text rules on this PC - Removes "um" and "uh" and keeps only the fix when you correct yourself ("tomorrow, no, Wednesday"). Optional spoken numbers ("twenty three" → 23), symbols, and emoji ("party emoji" → 🎉). A personal dictionary and snippets work with every model. Numbers and emoji are English only.
- History and stats - Search, copy, replay, and retry past takes. Audio is saved before transcribing, so a crash does not lose what you said. Usage stats with streaks and time saved. All local, with 1-day to forever retention.
- Audio - Silence is trimmed before the model runs, found by the Silero voice detector as on VocaMac (MIT, 1.8 MB, fetched with your first model Download); a take with only keyboard noise or silence is not decoded. Other apps can be muted while you record.
- Unsigned. SmartScreen is expected. There is no purchased CA signature and no Microsoft Store listing.
- The installer does not bundle a speech model. First run needs a network once to download one.
- Windows blocks typing into apps running as administrator. VocaWin then leaves the text on the clipboard and says so; press Ctrl+V.
- Parakeet CTC and Vosk stay out of the catalog until they work.
- No auto-update. Expect bugs. File an issue if something breaks.
The recognizer is not a cloud API. VocaWin only invokes a locally installed or downloaded engine. Model downloads may use the network once. Audio and transcription do not.
Tauri UI (TypeScript)
└─ Rust command layer
├─ Settings + model catalog
├─ Global push-to-talk shortcut + UI recording coordinator
├─ CPAL microphone capture + 16 kHz resampling
├─ whisper.cpp adapter (Whisper-family models)
├─ ONNX adapters (Parakeet, Moonshine, SenseVoice, GigaAM, Canary, Cohere)
├─ System tray (Show / Quit, close-to-tray)
└─ Windows text injector (SendInput)
| Engine | Initial models | Windows acceleration |
|---|---|---|
| whisper.cpp | Tiny through Large v3 Turbo, Distil-Whisper, Voca Hinglish | Vulkan, CPU fallback |
| ONNX Runtime | Parakeet, Moonshine, SenseVoice, GigaAM, Canary, Cohere | DirectML for Parakeet and SenseVoice, CPU fallback; others CPU |
The development build does not bundle a large model. Prefer the in-app Download buttons on the Models page. For Whisper testing from a shell, you can also place a whisper.cpp GGML model into VocaWin's local model folder:
$models = Join-Path $env:APPDATA "com.vocahq.vocawin\models"
New-Item -ItemType Directory -Force $models
Invoke-WebRequest "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.bin" -OutFile (Join-Path $models "whisper-tiny.bin")See the ONNX model guide for Parakeet, Moonshine, SenseVoice, GigaAM, and Canary layouts.
Same privacy bar, different machines. Start at vocahq.com for the map.
| Platform | Project | Website | GitHub | Status |
|---|---|---|---|---|
| Family | VocaHQ | vocahq.com | VocaHQ | Directory |
| Linux | VocaLinux | vocalinux.com | VocaHQ/vocalinux | Available |
| macOS | VocaMac | vocamac.com | VocaHQ/vocamac | Available |
| iPhone / Android | VocaPhone | vocaphone.vocahq.com | VocaHQ/vocaphone | Android beta / iOS TestFlight |
| Windows | VocaWin | vocawin.com | VocaHQ/vocawin | Beta |
| Infrastructure | VocaGateway | vocagateway.vocahq.com | VocaHQ/vocagateway | Early |
VocaGateway is optional self-hosted compute for other Voca clients. VocaWin does not expose a gateway mode today.
- Speech Engine: whisper.cpp and ONNX Runtime, both local
- Platform: Windows 10/11
- GPU: Vulkan for Whisper, DirectML for ONNX, CPU fallback for both
- Languages: determined by the downloaded model
A macOS or Linux host can validate the frontend and the Rust command layer (npm run check, cargo test). A real desktop build (npm run tauri dev / npm run tauri build) needs Windows 10/11 plus the same toolchain CI installs:
- Node.js 20+ (CI uses 22)
- Rust stable
- Tauri Windows prerequisites: Microsoft C++ Build Tools (MSVC) and WebView2
- LLVM Windows installer (
LLVM-*-win64.exe), solibclangis on disk. CI setsLIBCLANG_PATHtoC:\Program Files\LLVM\bin. - LunarG Vulkan SDK for whisper.cpp Vulkan. Confirm
VULKAN_SDKpoints at the SDK root (CI usesC:\VulkanSDK\1.4.363.0). - CMake for the whisper/ggml native build. Add it to PATH. CI also installs Ninja and sets
CMAKE_GENERATOR=Ninja. - A way around Windows
MAX_PATH(260 characters). whisper.cpp Vulkan shader nests go deep even if the repo sits atC:\vocawin. Enable OS long paths, or use a shortCARGO_TARGET_DIRthe way CI does (C:\t).
Open a new terminal after installing so PATH and env vars refresh. Chocolatey (choco install cmake llvm ninja) matches the CI installers if you already use it.
npm install
npm run tauri dev # desktop development
npm run tauri build # NSIS .exe on Windows (MSI is paused while the version is X.Y.Z-beta)
npm run check # TypeScript build + Rust testsWindows injection, WASAPI, Vulkan, DirectML, and installer artifacts must be exercised on Windows 10/11.
These are the failures that show up when a CI dependency is missing locally.
Unable to find libclang (couldn't find any valid shared libraries matching ['clang.dll', 'libclang.dll'])
Install LLVM from the Windows installer. Then, if the crate still cannot see it:
$env:LIBCLANG_PATH = "C:\Program Files\LLVM\bin"Set that as a user environment variable so it survives a reboot. Confirm clang.dll or libclang.dll is in that folder.
Please install Vulkan SDK and ensure that VULKAN_SDK env variable is set
Install the LunarG Vulkan SDK and reopen the terminal. Check:
echo $env:VULKAN_SDKIf it is empty, point it at the SDK root, for example C:\VulkanSDK\1.4.363.0 (use the version you installed). CI pins 1.4.363.0. A current LunarG Windows SDK is fine for local work.
is cmake not installed? (failed to execute command: program not found)
Install CMake and tick "Add CMake to the system PATH" (or winget install Kitware.CMake). Confirm cmake --version in a new shell. The transcribe.cpp/ggml crate build needs it.
exceeds the OS max path limit / The fully qualified file name must be less than 260 characters
transcribe-cpp-sys builds ggml's Vulkan shaders in nested folders under target\debug\build\transcribe-cpp-sys-*\out that can hit MAX_PATH (it shortens them itself, but a short target dir is still the safe setting). Either of these is enough:
- Point Cargo at a short target dir, matching CI:
$env:CARGO_TARGET_DIR = "C:\t"
$env:CMAKE_GENERATOR = "Ninja" # optional; CI uses Ninja to avoid MSBuild path racesPut Ninja on PATH if you set CMAKE_GENERATOR=Ninja.
- Enable Windows long paths, then reboot:
- Group Policy: Computer Configuration, Administrative Templates, System, Filesystem, Enable Win32 long paths
- Or registry:
HKLM\SYSTEM\CurrentControlSet\Control\FileSystem, DWORDLongPathsEnabled=1
Windows CI uses CARGO_TARGET_DIR=C:\t and CMAKE_GENERATOR=Ninja. It does not enable OS long paths on the runner.
Windows CI builds an unsigned NSIS installer on source-changing pushes to main and on workflow_dispatch, then uploads it as a GitHub Actions artifact. App-source pull requests run cargo test only; they do not package the setup wizard. Docs-only PRs skip the Windows VMs via the paths-filter job so the required check is not left pending. The installer stays unsigned. SmartScreen can still warn.
Pushing a v* tag builds the same NSIS installer and attaches it to a GitHub Release. While the app version is X.Y.Z-beta, tagged cuts are NSIS only (MSI is paused because WiX rejects the -beta marker). Testers should use Releases, not the workflow artifact. The build is unsigned, not a purchased CA or store signature. Windows will likely still warn. More info, then Run anyway. There is no Microsoft Store listing and no auto-update. Read the setup guide before you install, and file an issue if something breaks. vocawin.com points at the same download.
- Windows 10 version 1809 or later, or Windows 11
- 4 GB RAM (8 GB+ recommended for larger models)
- Microphone
- GPU recommended for faster transcription (NVIDIA, AMD, or Intel)
The landing page at vocawin.com lives in web/ and deploys through GitHub Pages.
cd web
python3 -m http.server 4173
node --test tests/site.test.mjsGitHub Actions publishes web/ on pushes to main. The web/CNAME file maps vocawin.com.
VocaWin is in early development. Download the unsigned beta from Releases if you want to try it. File bugs on Issues. The family directory is vocahq.com.
The product design and model support are informed by VocaMac, VocaLinux, VocaPhone, VocaGateway, Handy, and Dictus.
AGPL-3.0-or-later. Copyright (C) 2026 Jatin Kumar Malik.