Privacy-first local AI voice dictation with real-time speech-to-text, intelligent text editing, and global text injection.
I built VoxScribe because voice input should be private, fast, and work everywhere — without sending your words to the cloud. Goal: let users dictate into any app with real-time transcription and AI-powered text cleanup, all running locally.
Context:
- The average person types 40 WPM but speaks 130 WPM — voice is 3x faster.
- Cloud-based dictation tools send every word to remote servers, raising privacy concerns for sensitive work.
- Existing local solutions lack intelligent editing — they transcribe raw speech without fixing filler words, grammar, or tone.
Sources:
- https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6456571/
- https://hai.stanford.edu/news/speech-faster-typing
- https://www.eff.org/deeplinks/2023/08/your-voice-your-data
- Real-time voice capture with Web Audio API and live waveform visualization.
- Browser-native speech-to-text via SpeechRecognition API with interim results.
- Three edit modes: Raw (no changes), Light Edit (grammar + filler removal), Aggressive Rewrite (full tone transformation).
- Tone selector: Casual, Professional, or Concise output styles.
- Canvas-based waveform visualizer with bar and wave rendering modes.
- Dictation history with copy-to-clipboard support.
- Configurable model selection UI (STT + LLM) ready for Tauri/local model integration.
- Keyboard shortcut support (Space to toggle recording).
- Dark-mode-first glassmorphism UI designed to feel like a native desktop app.
- Next.js 16 + React 19 + TypeScript
- Tailwind CSS v4 + shadcn/ui
- Web Audio API (AnalyserNode for real-time frequency data)
- SpeechRecognition API (browser-native STT)
- Canvas API (waveform rendering)
- Tauri-ready architecture (IPC structure documented for Rust backend)
Data Flow: Mic --> Audio Capture --> Whisper STT --> Raw Text --> LLM Edit --> Edited Text --> Global Keyboard Injection --> Active App
┌─────────────────────────────────────────┐
│ Tauri Application │
│ │
│ ┌───────────────────────────────────┐ │
│ │ Frontend (React/Svelte) │ │
│ │ │ │
│ │ Floating Waveform Settings │ │
│ │ Mic Widget Display Panel │ │
│ └──────────────┬────────────────────┘ │
│ │ IPC (invoke/events) │
│ ┌──────────────▼────────────────────┐ │
│ │ Rust Backend (Tauri) │ │
│ │ │ │
│ │ ┌─────────────┐ ┌────────────┐ │ │
│ │ │ Audio │ │ Global │ │ │
│ │ │ Capture │─│ Hotkey │ │ │
│ │ │ (cpal) │ │ (rdev) │ │ │
│ │ └──────┬──────┘ └────────────┘ │ │
│ │ │ audio stream │ │
│ │ ┌──────▼──────┐ │ │
│ │ │Transcription│ whisper.cpp │ │
│ │ │ Engine │ [Background] │ │
│ │ └──────┬──────┘ │ │
│ │ │ raw transcript │ │
│ │ ┌──────▼──────┐ │ │
│ │ │ LLM Edit │ llama.cpp / │ │
│ │ │ Engine │ Ollama HTTP │ │
│ │ └──────┬──────┘ │ │
│ │ │ edited text │ │
│ │ ┌──────▼──────┐ │ │
│ │ │ Keyboard │ enigo │ │
│ │ │ Injection │ → active app │ │
│ │ └─────────────┘ │ │
│ └───────────────────────────────────┘ │
└─────────────────────────────────────────┘
app/
├── layout.tsx # root layout with dark theme + Inter font
├── page.tsx # main app orchestration + state management
└── globals.css # oklch color tokens + custom design system
components/
├── mic-button.tsx # animated mic toggle with glow + audio level
├── waveform-visualizer.tsx # canvas-based real-time audio visualization
├── transcript-display.tsx # raw vs edited text with diff highlighting
├── edit-mode-toggle.tsx # Raw / Light Edit / Rewrite + tone selector
├── settings-panel.tsx # model config, audio device, custom prompts
├── status-bar.tsx # recording state, duration, hotkey hints
└── history-panel.tsx # past dictations with copy + delete
hooks/
├── use-audio-engine.ts # Web Audio API capture + frequency analysis
└── use-transcription.ts # SpeechRecognition + local text editing pipeline
To have dictated text appear in Apple Notes, Terminal, or any app:
- Same shortcut everywhere — Hold Ctrl+D (Mac: Control+D) to record; release to stop and paste. From any app; text is pasted where your cursor is. Or use the tray / mic to toggle start and stop.
- macOS: Accessibility — Typing into other apps uses simulated keypresses (clipboard + Cmd+V). You must add the app under System Settings → Privacy & Security → Accessibility. When running from source (dev build) the app appears as "app" with a generic icon: add it via + → Cmd+Shift+G →
src-tauri/target/debug→ select app. After building (pnpm tauri build), the app appears as VoxScribe with the VoxScribe icon in the list, like other branded apps. Without Accessibility permission, paste will not go to the focused app. - macOS: No dock icon — On Mac, the app runs as an "accessory" so it does not steal focus when you press the global hotkey from Notes or another app. You open the window from the menu bar tray (click the icon). This keeps the target app in focus so the paste goes there.
You can also use the menu bar tray: Start / stop dictation (Ctrl+D).
pnpm install
pnpm devNo environment variables required — everything runs locally in the browser.