Loquara

Loquara is a fast, private speech-to-text dictation application for Windows. Press a global shortcut anywhere, speak, and have your speech transcribed locally on your GPU or CPU using the NVIDIA Parakeet TDT 0.6B v3 model — and pasted directly into your active window, code editor, browser, or terminal.

Offline by default: audio and transcripts never leave your computer unless you switch to a cloud engine. No Python, PyTorch, or cloud accounts required.

Optional cloud engines: bring your own API key and dictate through xAI Grok Voice Transcribe, OpenAI, Groq, Mistral, ElevenLabs, Deepgram, or any OpenAI-compatible endpoint. Keys are encrypted with Windows DPAPI, and the connection can be verified from Settings before you dictate.

---

✨ Features

• Global shortcut: Ctrl+Space starts and stops dictation anywhere. Configurable in settings.
• Minimal floating overlay: Compact pill with a noise-adaptive meter during recording. Sound cues signal start, stop, and transcript readiness.
• Universal paste:
  • Standard apps: Ctrl + V
  • Linux SSH / Terminals: automatic Ctrl + Shift + V for Termius, Windows Terminal, WSL, Git Bash, Alacritty, WezTerm, Ghostty, etc.
  • PuTTY / KiTTY: automatic Shift + Insert
  • Custom paste mode selection available in Settings.
• Runs on any GPU or CPU: DirectML hardware acceleration covers NVIDIA, AMD, Intel, and CPU fallback out of the box with zero driver toolkits to install.
• Built-in model manager: Downloads the lightweight 670 MB Parakeet model in-app with 4 concurrent streams and live progress.
• Cloud transcription (optional): Pick a provider, a model, and paste an API key — settings can send a second of silence to check the connection and show how long the round trip took. Keys are stored encrypted with Windows DPAPI, the model list can be fetched from the provider, and your vocabulary is passed along as keyterms where the provider supports it.
• Live dictation (xAI): With "Type as you speak" enabled, Grok Voice Transcribe streams over a WebSocket and finished phrases land at the caret while you are still talking. If the live socket fails, the finished recording still goes through the ordinary request path.
• Small uploads: recordings are converted to 16 kHz mono before they are sent, which is the shape every recogniser consumes and several times fewer bytes than a 48 kHz stereo capture.
• Custom vocabulary & modes: Teach Loquara specialized jargon, brand names, and phonetic replacements. Transform output with custom mode rules.
• Audio history & retention: Review past recordings and transcripts. Set automatic audio cleanup (1 day, 7 days, 30 days, or forever).
• System tray integration: Start dictation, paste last transcription, or launch minimized to the tray at login.
• Bilingual interface: Polish and English UI matching system language with manual override.

---

🚀 Getting Started

Installation

1. Download the latest installer (Loquara_x.x.x_x64-setup.exe or .msi) from Releases.
2. Run the installer and launch Loquara.
3. On first start, click Download to fetch the offline speech engine model (670 MB).
4. Press Ctrl+Space anywhere in Windows to start dictating!

---

🛠️ Architecture & Tech Stack

Loquara is built as a lightweight, native Windows desktop application:

• Frontend: React 19, TypeScript, Vite, Fluent UI icons, Radix UI.
• Backend / Platform: Tauri 2 (Rust), Win32 API (window management, thread-attached input injection, scan-code mapping), cpal (audio capture), rusqlite (SQLite storage).
• Speech Engine: Native sherpa-onnx ONNX Runtime bindings with DirectML / CPU acceleration running nvidia/parakeet-tdt-0.6b-v3, or a cloud provider's HTTP transcription API when one is configured.

---

💻 Development & Building

Prerequisites

• Windows 10/11 x64
• Node.js 22+ & pnpm
• Rust (edition 2024 / stable toolchain)
• CMake 3.x (for compiling native ONNX runtime bindings)

Running locally

# Install frontend dependencies
pnpm install

# Run the app in development mode (with hot reload)
pnpm tauri dev

Running tests

# Run frontend tests
pnpm test

# Run Rust unit & platform tests
cargo test --manifest-path src-tauri/Cargo.toml --lib

Building the installer

pnpm tauri build

The resulting NSIS installer and MSI packages are created in src-tauri/target/release/bundle/.

---

📁 Local Data Paths

• Settings, SQLite database, and recordings: %APPDATA%\io.loquara.desktop\
• Engine models: %APPDATA%\io.loquara.desktop\models\

---

📄 License

MIT — see LICENSE.