Offline-first desktop voice dictation — no cloud, no telemetry, just your voice.
Record with a global shortcut, transcribe locally with whisper.cpp, parakeet.cpp, or an optional NVIDIA NeMo runtime, and send text straight to your clipboard or active app.
LocalVoice is a lightweight desktop app that turns your voice into text — entirely on your machine. There's no account to create, no audio sent to a server, and no subscription. Just press a shortcut, speak, and your words appear wherever your cursor is.
It's built for developers, writers, and anyone who wants fast, private voice input as part of their daily workflow.
- Global hotkey recording — start and stop dictation from anywhere on your desktop
- Push-to-talk mode — hold the shortcut to record, release to stop; configurable per session
- 100% local transcription — powered by bundled Whisper/parakeet.cpp sidecars or an optional local NeMo runtime; your audio never leaves your machine
- Multiple model runtimes — download and switch between Whisper GGML, Parakeet GGUF, and optional
.nemomodels per language - Smart output — insert directly into the active app, copy to clipboard, or preview first
- Custom dictionary — teach LocalVoice your vocabulary, acronyms, and corrections
- Filler word removal — automatically strips "um", "uh", and other fillers
- Ambiguity detection — flags low-confidence phrases for your review
- Session history — browse, search, and filter past transcriptions with bulk select, delete, and export
- Audio playback — replay the original recording directly from the session detail view
- Pagination & filters — configurable page size (25/50/100), "has audio" toggle, and date quick-presets
- CSV / JSON / TXT export — export single or multiple sessions in your preferred format
- Dashboard & analytics — WPM trends, language breakdown, correction metrics
- Session reprocessing — re-run post-processing on past sessions with updated rules
- Compact pill UI — a small floating window that stays out of your way
- Themes & shortcuts — light, dark, or system theme; fully configurable hotkeys
- No telemetry — zero data collection, ever
| Layer | Technology |
|---|---|
| Desktop framework | Tauri v2 |
| Frontend | React 19, TypeScript 5.8, Vite |
| Styling | Tailwind CSS v4, shadcn/ui, Radix UI |
| State management | Zustand |
| Backend | Rust (stable ≥ 1.77) |
| Database | SQLite (bundled via rusqlite) |
| Transcription | whisper.cpp sidecar, parakeet.cpp sidecar, optional NVIDIA NeMo worker |
| Audio capture | cpal, hound |
Dashboard — usage stats, WPM trend & language breakdown |
Model Manager — download local transcription models (dark theme) |
Model Manager — per-language defaults & installed models (light theme) |
Log Viewer — filterable in-app debug logs with JSON export |
| Platform | Status | Notes |
|---|---|---|
| Windows 10/11 x64 | First-class | Fully supported, signed installers |
| macOS Apple Silicon (arm64) | Supported | Unsigned builds; Accessibility permission required for auto-insert |
| macOS Intel (x86_64) | Supported | Same as Apple Silicon |
| Linux x86_64 | Supported | Requires xdotool (X11) or wtype (Wayland) for auto-insert; libappindicator for tray |
Security notices by platform
- Windows: You may see a SmartScreen warning ("unknown publisher"). Click "More info" → "Run anyway". Release binaries are Authenticode-signed via SignPath.
- macOS: The app is not notarized. Right-click the
.app→ Open → Open to bypass Gatekeeper. Auto-insert (paste) requires granting Accessibility permission in System Settings → Privacy & Security → Accessibility.- Linux: No code signing. Mark the binary executable and run directly. Tray icon requires
libayatana-appindicator3orlibappindicator3.
Release builds check https://github.com/iptoux/localvoice/releases/latest/download/latest.json for the latest stable GitHub release. LocalVoice shows a native notification and an in-app banner when an update is available, but it only downloads and installs after the user clicks Update Now. The automatic startup check can be disabled in Settings -> System -> Automatic updates.
Updater-capable public releases must be built by the GitHub release workflow so the signed update bundles, .sig files, and latest.json manifest are uploaded together. Local release scripts are useful for manual installer builds, but they do not publish a complete updater channel by themselves.
| Tool | Version |
|---|---|
| Node.js | ≥ 20 |
| pnpm | 10.34.4 |
| Rust | stable ≥ 1.77 |
macOS additional requirements:
- Xcode Command Line Tools:
xcode-select --install - cmake (for building whisper.cpp):
brew install cmake
Linux additional requirements:
sudo apt-get install -y \
libwebkit2gtk-4.1-dev libssl-dev libgtk-3-dev \
libayatana-appindicator3-dev librsvg2-dev libasound2-dev \
cmake build-essential
# For auto-insert (X11):
sudo apt-get install -y xdotool
# For auto-insert (Wayland):
sudo apt-get install -y wtypeClone the repo and run the bootstrap script. It handles dependencies, Whisper/parakeet.cpp sidecars, and build verification automatically.
Windows (PowerShell):
git clone https://github.com/your-username/localvoice.git
cd localvoice
.\scripts\bootstrap.ps1macOS / Linux:
git clone https://github.com/your-username/localvoice.git
cd localvoice
./scripts/bootstrap.shThe script will:
- Check for Node.js, Rust, and pnpm (and install pnpm if missing)
- Check Linux system packages (Linux only)
- Install frontend dependencies
- Download or build whisper.cpp binaries for your platform (skip with
--skip-whisper) - Download and verify the pinned parakeet.cpp CLI, build the streaming worker, and stage Parakeet runtime DLLs (skip with
--skip-parakeet) - Verify the Tauri CLI is available
- Run a Rust compilation check (skip with
--skip-verification)
# Install frontend dependencies
pnpm install
# Start the dev server (hot-reload frontend + Rust watch)
pnpm tauri dev
# Release build with updater artifacts
pnpm tauri buildOn Windows, pnpm tauri build runs pnpm run tauri:prepare before Tauri bundles the app. That preparation builds the frontend, downloads the pinned Parakeet CLI when missing, builds parakeet-stream-worker.exe, and stages required DLLs under src-tauri/parakeet-runtime/.
pnpm tauri build requires TAURI_SIGNING_PRIVATE_KEY because updater artifacts are enabled. For a Windows installer build without publishing a GitHub release or creating updater artifacts, use the release helper's local mode:
.\scripts\create-release.ps1 -LocalBuildUse .\scripts\create-release.ps1 -LocalBuild -UpdaterArtifacts only when you intentionally want signed updater artifacts locally. In that mode the script reads the updater key from %USERPROFILE%\.tauri\localvoice-updater.key if TAURI_SIGNING_PRIVATE_KEY is not already set.
CI and release jobs install dependencies with pnpm install --frozen-lockfile.
When updating Tauri, keep the Rust tauri crate and frontend
@tauri-apps/api package on the same major/minor version before running a
release build.
The bootstrap script and Windows Tauri build preparation handle this automatically. For manual setup, you need to place platform-appropriate sidecar binaries in src-tauri/binaries/:
| Sidecar | Target name example |
|---|---|
| whisper.cpp | whisper-cli-x86_64-pc-windows-msvc.exe |
| parakeet.cpp | parakeet-cli-x86_64-pc-windows-msvc.exe |
| parakeet.cpp streaming worker | parakeet-stream-worker-x86_64-pc-windows-msvc.exe |
Parakeet sidecars are pinned to mudler/parakeet.cpp v0.3.2 and must be checksum-verified or built from the pinned source before release packaging. Public installers bundle the CPU/portable sidecars only; model weights, .nemo checkpoints, CUDA stacks, and Python/NeMo environments are never bundled in the base installer.
The Parakeet streaming worker can also produce small runtime libraries during CI/release builds. Those files are staged under src-tauri/parakeet-runtime/ and bundled as Tauri resources; installed Windows builds load them from the bundled resources/parakeet-runtime/ directory. The source directory is kept in Git, but generated runtime files are ignored.
Windows:
- Download
whisper-bin-win-x64.zipfrom the whisper.cpp v1.7.1 release - Extract and copy files:
| File | Destination |
|---|---|
Release/whisper-cli.exe |
src-tauri/binaries/whisper-cli-x86_64-pc-windows-msvc.exe |
Release/ggml.dll |
src-tauri/ggml.dll |
Release/ggml-base.dll |
src-tauri/ggml-base.dll |
Release/whisper.dll |
src-tauri/whisper.dll |
Release/SDL2.dll |
src-tauri/SDL2.dll |
Some whisper.cpp releases may also include ggml-cpu.dll; the setup action copies it when present, but the Tauri bundle does not require it because current Windows assets do not ship it.
macOS (Apple Silicon):
git clone --depth 1 --branch v1.7.1 https://github.com/ggerganov/whisper.cpp
cd whisper.cpp && cmake -B build -DCMAKE_BUILD_TYPE=Release -DWHISPER_BUILD_EXAMPLES=ON
cmake --build build --target whisper-cli -j$(sysctl -n hw.logicalcpu)
cp build/bin/whisper-cli ../src-tauri/binaries/whisper-cli-aarch64-apple-darwinmacOS (Intel): same steps, use target name whisper-cli-x86_64-apple-darwin.
Linux:
git clone --depth 1 --branch v1.7.1 https://github.com/ggerganov/whisper.cpp
cd whisper.cpp && cmake -B build -DCMAKE_BUILD_TYPE=Release -DWHISPER_BUILD_EXAMPLES=ON
cmake --build build --target whisper-cli -j$(nproc)
cp build/bin/whisper-cli ../src-tauri/binaries/whisper-cli-x86_64-unknown-linux-gnuUse the pinned v0.3.2 CPU/portable release assets from mudler/parakeet.cpp, then rename the extracted CLI to Tauri's target-triple sidecar format:
| Platform | Destination |
|---|---|
| Windows x64 | src-tauri/binaries/parakeet-cli-x86_64-pc-windows-msvc.exe |
| macOS arm64 | src-tauri/binaries/parakeet-cli-aarch64-apple-darwin |
| macOS Intel | src-tauri/binaries/parakeet-cli-x86_64-apple-darwin |
| Linux x64 | src-tauri/binaries/parakeet-cli-x86_64-unknown-linux-gnu |
| Linux arm64 | src-tauri/binaries/parakeet-cli-aarch64-unknown-linux-gnu |
All sidecar binaries are excluded from version control (
.gitignore). Every contributor must provide them manually or run the bootstrap script.
macOS — Auto-insert (paste):
LocalVoice.app must be granted Accessibility permission:
System Settings → Privacy & Security → Accessibility → enable LocalVoice.
macOS — Autostart:
The autostart toggle writes a launchd plist to ~/Library/LaunchAgents/com.localvoice.app.plist.
Linux — Auto-insert (paste): Install the tool matching your display server:
- X11:
sudo apt-get install xdotool - Wayland:
sudo apt-get install wtype
Linux — Tray icon:
Requires libayatana-appindicator3 or libappindicator3. Install via:
sudo apt-get install libayatana-appindicator3-devLinux — Autostart:
The autostart toggle writes an XDG .desktop file to ~/.config/autostart/localvoice.desktop.
LocalVoice stores all settings in a local SQLite database — no config files to edit by hand. Everything is configurable through the app's Settings page:
| Setting | Description |
|---|---|
| Recording shortcut | Single key or key combination to start/stop recording |
| Output mode | Insert to active app, clipboard, or preview |
| Default language | Language used for transcription |
| Default transcription engine | Preferred engine (whisper-cpp, parakeet-cpp, or optional nemo) |
| Preferred transcription runtime | Bundled sidecar or optional local NeMo runtime |
| Active model | Per-language model selection |
| Streaming transcription | Enables low-latency streaming for capable Parakeet GGUF models; optional live insert writes cleanup-filtered worker deltas |
| Pill mode | Recording overlay (default) or Classic pill |
| Theme | System, light, or dark |
| Filler words | Language-specific list of words to strip |
| Audio retention | Whether to keep raw audio after transcription |
| Logging | Enable/disable in-app debug logging and the persistent app log file |
- Launch LocalVoice. By default the main window opens; the recording overlay stays hidden until you record.
- Press your configured shortcut (default: customizable in Settings) to start recording.
- Speak. The default recording overlay appears bottom-center with a waveform only.
- Press the shortcut again (or let silence detection stop it automatically).
- LocalVoice transcribes locally and sends the text to your active app or clipboard.
Switch Settings -> Appearance -> Pill mode to Classic pill if you prefer the older persistent compact/expanded pill with transcript preview.
You can open the full dashboard at any time to browse history, manage models, edit your dictionary, review ambiguous phrases, and view usage stats.
src/ React/TypeScript frontend
src-tauri/
src/
commands/ Tauri IPC command handlers
db/ SQLite layer (migrations, repositories)
audio/ Recording and device management
transcription/ hybrid engine orchestration, sidecar protocols, NeMo worker bridge
postprocess/ Text cleaning, filler removal, corrections
dictionary/ Custom vocabulary and correction rules
os/ Tray, hotkeys, clipboard, text insertion
state/ AppState shared across commands
errors/ AppError / CmdResult types
docs/
user/ User-facing guides
dev/ Developer and architecture docs
scripts/ Bootstrap and utility scripts
Full developer reference: docs/dev/index.md
Contributions are welcome. Here's how to get involved:
Reporting bugs
Open an issue with a clear description, steps to reproduce, your OS, and the LocalVoice version. Attach logs from the in-app log viewer or the plain-text log file at the app data directory (%APPDATA%\com.localvoice.app\localvoice.log on Windows) if relevant.
Suggesting features Open a discussion or issue describing the use case and why it would be valuable. Check existing issues first to avoid duplicates.
Submitting a pull request
- Fork the repo and create a branch from
main:git checkout -b feat/your-feature - Follow the existing code style — Rust uses standard
rustfmt, TypeScript uses the project's ESLint config - Keep PRs focused — one feature or fix per PR
- Update or add documentation if your change affects user-facing behavior
- Open the PR with a clear description of what changed and why
Good first issues
Look for issues tagged good first issue — these are scoped and well-documented entry points.
Free code signing provided by SignPath.io, certificate by SignPath Foundation.
Official Windows release artifacts for LocalVoice are built from this public repository and submitted for signing only from the project's release workflow.
Committers and reviewers:
- @iptoux
Approvers:
- @iptoux
Only release artifacts produced from the LocalVoice repository are submitted for signing. Third-party upstream binaries included in release packages remain attributed to their original upstream projects and are not individually signed under the LocalVoice project certificate.
MIT — see LICENSE for the full text.