Light Whisper

A lightweight desktop app for global voice-to-text. Press Alt+Space anywhere to start recording, press again to transcribe and paste the result into the active application.

Built with Tauri v2 (Rust backend, vanilla JS frontend).

Features

• Global hotkey (Alt+Space) works from any application
• Two STT engines:
  • Whisper (OpenAI) via whisper.cpp — models from 40 MB to 500 MB
  • Parakeet TDT v3 (NVIDIA) via ONNX Runtime — ~670 MB, 25 languages, auto-detection
• French & English support (and more with Parakeet)
• Auto-paste: transcribed text is automatically pasted via clipboard + keyboard simulation
• Minimal UI: frameless overlay during recording, settings accessible from the tray icon
• Auto-opens settings on first launch if no model is downloaded

Dependencies

• Rust (1.70+)
• CMake + C/C++ compiler
• macOS 11+, Windows 10+, or Linux (X11 recommended)

Prerequisites install

# Install Rust
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh

# Install Tauri CLI
cargo install tauri-cli --version "^2"

MacOS: install Homebrew (https://brew.sh/) then CMake

# Optional if you already have homebrew
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

brew install cmake

Linux (Ubuntu 22.04+ / Debian 12+)

sudo apt-get install -y \
  libgtk-3-dev \
  libwebkit2gtk-4.1-dev \
  libayatana-appindicator3-dev \
  librsvg2-dev \
  libssl-dev \
  libasound2-dev \
  libxdo-dev \
  cmake \
  build-essential

Linux (Fedora 39+)

sudo dnf install -y \
  gtk3-devel \
  webkit2gtk4.1-devel \
  libayatana-appindicator-gtk3 \
  librsvg2-devel \
  openssl-devel \
  alsa-lib-devel \
  libxdo-devel \
  cmake \
  gcc-c++

Build & Run

# Development (hot reload frontend, debug backend)
cargo tauri dev

# Production build (creates .app + .dmg on macOS, .msi on Windows)
cargo tauri build

# Debug build (faster, no bundling optimization)
cargo tauri build --debug

# Rust backend only (no frontend bundling)
cd src-tauri && cargo build

First build takes ~5 minutes due to whisper.cpp compilation via CMake.

macOS note: CMAKE_OSX_DEPLOYMENT_TARGET=11.0 is required for whisper.cpp's std::filesystem usage. This is set automatically via src-tauri/.cargo/config.toml.

Usage

1. Launch the app — if no model is downloaded, the Settings window opens automatically
2. Choose an STT engine (Whisper or Parakeet) and download the model
3. Save settings and close the window
4. Record — press Alt+Space to start, Alt+Space again to stop
5. Result — transcribed text is automatically pasted into the active application
6. Access settings anytime via the tray icon (left or right click)

Available Models

Whisper (OpenAI)

Model	Size	Speed	Quality
tiny	~40 MB	Very fast	Basic
base	~60 MB	Fast	Recommended
small	~200 MB	Medium	Good
medium	~500 MB	Slow	Excellent

Parakeet TDT v3 (NVIDIA)

Model	Size	Speed	Quality
int8 quantized	~670 MB	Fast	Excellent (WER 7.7% FR)

Parakeet supports 25 European languages with automatic language detection. Models are downloaded from HuggingFace and stored in ~/lightwhisper/models/.

Permissions (macOS)

Light Whisper requests both permissions on first launch.

• Microphone: required for audio capture. Go to System Settings > Privacy & Security > Microphone and enable Light Whisper. Without this permission, macOS silently feeds empty audio to the app — recording appears to work but the waveform stays flat and no transcription is produced.
• Accessibility: required for auto-paste (keyboard simulation). Go to System Settings > Privacy & Security > Accessibility and enable Light Whisper. The app will prompt on first launch; if denied, transcribed text cannot be pasted automatically.

Project Structure

light-whisper/
├── src/                        # Frontend (vanilla HTML/JS/CSS)
│   ├── index.html              # Recorder overlay (280x80 frameless window)
│   ├── settings.html           # Settings (engine, model, device, language)
│   └── settings.js / .css
├── src-tauri/                  # Rust backend
│   ├── src/
│   │   ├── lib.rs              # Tauri setup, commands, tray, shortcut handler
│   │   ├── audio.rs            # Audio capture (cpal) on dedicated thread
│   │   ├── stt.rs              # STT engine dispatch (Whisper + Parakeet)
│   │   ├── paste.rs            # Clipboard + keyboard simulation (enigo)
│   │   ├── config.rs           # JSON config I/O, directory paths
│   │   └── model_manager.rs    # Model download with streaming progress
│   └── Cargo.toml
└── README.md

Storage

~/lightwhisper/
├── config.json                     # {audio_device, model_size, language, engine}
├── models/
│   ├── ggml-{size}.bin             # Whisper models
│   └── parakeet-tdt/               # Parakeet ONNX models
│       ├── encoder-model.onnx
│       ├── decoder_joint-model.onnx
│       └── vocab.txt
└── temp/                           # Temporary WAV files

Windows uses %TEMP%\lightwhisper\ for temp files.

License

MIT