README

VoxCtrl Documentation

VoxCtrl is a high-performance, privacy-first voice-to-text dictation application and programmable voice input broker. All processing happens 100% on-device with zero telemetry or cloud dependencies.

Global shortcuts are registered with your desktop through the XDG GlobalShortcuts portal, so VoxCtrl does not read your keyboard and needs no permission setup at all. See Privacy & Security for exactly what it can see and how to verify it.

---

Wiki Index

Document	Description
Overview	What VoxCtrl does, key features, and design principles
Privacy & Security	What VoxCtrl can see, what the installer touches, and how to verify both
Release signing	How releases are signed, how the updater verifies them, and the one-time key setup
Architecture	System design, crate layout, data flow, concurrency model
Audio Pipeline	Audio capture, device management, VAD, resampling
Speech Recognition	Whisper engine, models, inference pipeline, post-processing
Routing	Output targets, hotkey bindings, delivery types
Hotkeys	Global shortcuts via the desktop portal, gestures, platform support
Text-to-Speech	TTS engines, voice packs, playback
Integrations	MCP server, DBus service, OpenAI-compatible LLM API, webhooks
UI & Windows	Svelte frontend, overlay, settings
API Reference	Tauri IPC commands and frontend events
Configuration	All config files, schemas, and options
Installation & Setup	Dependencies, building, running
Development Guide	Dev environment, build system, crate structure
Windows Build	Building VoxCtrl on Windows
Windows Port Plan	Audit and phased plan for full Windows 11 parity
Windows Testing	Hand to a Windows tester: install, what to try, how to send a log
Bug Reports	What Settings → Bug Report collects, what it never collects, and the four ways to send one

---

Quick Summary

Microphone → Audio Capture → Speech Inference (whisper.cpp / Moonshine / Parakeet / Remote)
                                          │
                                          ▼
                               Post-Processing & S1-mini Normalization
                                          │
                                          ▼
                                   Output Router
                                          │
                  ┌───────────────────────┴───────────────────────┐
                  │                                               │
             Inject text                                 Clipboard / File / HTTP /
             to window                                   Webhook / Socket / DBus /
                                                         MCP / Exec / Pipe / Speak / Chat

Tech Stack:

• Frontend: Svelte 5 + Tailwind CSS 4 + Vite 5
• Desktop Shell: Tauri 2 (Rust + WebView); the dictation overlay is a second WebviewWindow, not a separate process
• Backend: Rust (Tokio async), 15 specialized workspace crates
• Speech Recognition: whisper.cpp (GGUF, CPU/Vulkan/CUDA), Moonshine (ONNX/WebGPU), Parakeet TDT (ONNX), Remote Speech Engine (OpenAI-compatible)
• Dictation Cleanup: S1-mini (Qwen3-0.6B) via Vulkan-accelerated voxctrl-llm-sidecar (llama.cpp)
• Text-to-Speech: Breeze-TTS-2, Piper, Pocket-TTS, Inflect-Micro-v2, VoxCPM2, eSpeak-NG (with on-demand idle unload)
• Config: TOML + JSON, hot-reloadable