VeloScribe is a desktop application that gives any text field on your Mac a supercharged voice-to-text pipeline. Press a global hotkey, speak, and your words are transcribed in real time, then polished by an LLM before landing in the active window — no copy-paste, no context switching.
What it does
- Global hotkey capture — a configurable shortcut triggers dictation from any app, no window focus needed.
- Live floating HUD — a transparent, click-through glassmorphic overlay shows transcription progress and AI status without interrupting your workflow.
- Command / Edit Mode — highlight existing text anywhere in the OS, hold a hotkey, speak rewrite instructions, and the result replaces the selection via OS-level clipboard bindings.
- Custom prompt override — users supply a system-prompt instruction (e.g. "keep URLs untouched, no filler words") that routes through to the LLM for every dictation session.
How it's built
The client is a Tauri v2 desktop app using React, TypeScript, and Vite. A
WebSocket AudioWorklet streams raw PCM mic input to a FastAPI backend, where
Deepgram transcribes in real time and Claude/Haiku polishes via an agentic
routing layer. A Rust clipboard sentinel with 250 ms polling guards against
race conditions with slow-responding apps. Settings persist to localStorage
and sync hotkeys through @tauri-apps/plugin-global-shortcut.