Dictate into any app. Transform with local AI. No cloud. No subscription. Free forever.
Voxa is a free, open-source macOS dictation app that transcribes your voice and injects text into any application — then transforms it with a local LLM using customizable profiles. Powered by Whisper for speech-to-text and llama.cpp for intelligent post-processing, everything runs entirely on your machine.
No API keys. No subscriptions. No data leaves your device. Ever.
| Voxa | Wispr Flow | Superwhisper | VoiceInk | |
|---|---|---|---|---|
| Price | ✅ Free forever | $20/month | $5/month | $25 one-time |
| Fully local processing | ✅ 100% | ❌ Cloud | Partial | ✅ |
| Local LLM transformation | ✅ On-device | ✅ Cloud | ❌ | ❌ |
| Unlimited custom AI profiles | ✅ | Limited | ❌ | ❌ |
| Open source | ✅ | ❌ | ❌ | ✅ |
| No API key required | ✅ | ❌ | ❌ | ✅ |
| Works in any app | ✅ | ✅ | ✅ | ✅ |
| Apple Silicon native | ✅ | ✅ | ✅ | ✅ |
Voxa's biggest differentiator: it doesn't just transcribe — it thinks. Every dictation passes through a local LLM that reshapes your output based on the context you choose, without sending a single byte to any server.
Instead of raw transcription, Voxa passes your voice through a local LLM that reshapes the output according to a profile. Four built-in, unlimited custom:
| Profile | What it does |
|---|---|
| Elegant | Rewrites with perfect grammar and formal vocabulary. Keeps your ideas, elevates the expression. |
| Informal | Cleans up filler words and repetitions, keeps your natural tone. Great for Slack and chat. |
| Code | Acts as a prompt engineer. Turns your voice note into a structured, ready-to-use AI prompt (Role / Context / Task / Expected output). |
| Custom | Write your own system prompt. Full control over how the LLM processes your voice. |
Switch profiles instantly from the tray menu. Create as many custom profiles as you need.
- Works in every app — browser, VS Code, terminal, Slack, IntelliJ, Cursor. Just talk, Voxa handles the
Cmd+V. - VAD-Reactive Animation — the recording pill responds in real time to your microphone level.
- Focus Preservation — returns focus to the exact app you were in after injection, including Electron and JVM targets.
- Ultra-Compact Floating Pill — a minimal interface that lives at the edge of your screen, never in the way.
- Configurable Recording Limit — set auto-stop from 30s to 600s (Settings → General). Useful for long-form prompts.
- Launch at Login — start Voxa automatically when you log in (Settings → General → Behavior).
Every dictation is stored locally in SQLite:
- Review both the raw transcription and the LLM-refined version side by side.
- Edit any transcript after the fact.
- Delete individual entries or clear the full history.
Voxa learns from your corrections. When you fix a transcript, it automatically extracts new words and adds them to your personal dictionary — improving Whisper's recognition for domain terms, names, and jargon over time. Add or remove words manually from Settings at any time.
→ Download Voxa v1.6.0 for macOS (Apple Silicon)
Requires macOS 13+ on Apple Silicon (M1/M2/M3/M4).
- Download
Voxa_1.6.0_aarch64.dmgfrom the link above. - Open the
.dmgand drag Voxa to your Applications folder. - Before opening Voxa, run this in Terminal to remove the macOS quarantine flag:
xattr -cr /Applications/Voxa.app
- Open Voxa. On first launch, macOS will ask for Accessibility permission — click Open System Settings and enable the toggle for Voxa.
- Voxa automatically downloads the AI models (~1 GB) on first run.
- Start dictating: Alt+Space (push-to-talk) or F5 (hands-free toggle).
Voxa is not code-signed or notarized because the project doesn't have an Apple Developer account ($99/year). This means:
- macOS Gatekeeper will show "Voxa.app is damaged and can't be opened" — this is a false positive.
- The
xattr -crcommand removes the quarantine attribute added by macOS to downloaded files. - You also need to manually grant Accessibility permission in System Settings → Privacy & Security → Accessibility.
Voxa is 100% open source — you can audit every line of code and build it from source. No data ever leaves your machine.
If you have an Apple Developer account and want to help sign and notarize Voxa, see docs/code-signing.md.
| Permission | Why | How to grant |
|---|---|---|
| Accessibility | Capture global shortcuts and inject text into other apps via Cmd+V simulation |
System Settings → Privacy & Security → Accessibility → enable Voxa |
| Microphone | Record audio for voice dictation | Granted automatically on first recording attempt |
If shortcuts stop working after an update, remove Voxa from the Accessibility list and re-add it — macOS invalidates the permission when the binary hash changes.
- Core: Tauri v2 (Rust)
- Frontend: React + TypeScript + Vite
- Styling: Vanilla CSS (High-Performance Glassmorphism)
- Speech-to-Text:
whisper-rsviawhisper.cpp— local, GPU-accelerated on Apple Silicon - LLM:
llama-serverHTTP API viallama.cpp— Qwen2.5 running fully on-device
- Node.js (v22+)
- pnpm (v9+)
- Rust (stable)
- Tauri CLI (
pnpm add -g @tauri-apps/cli) - macOS with Xcode Command Line Tools (
xcode-select --install)
git clone https://github.com/lufermalgo/voxa.git
cd voxa
pnpm install
pnpm tauri devpnpm tauri build --target aarch64-apple-darwinThe .dmg will be in src-tauri/target/aarch64-apple-darwin/release/bundle/dmg/.
When running from
pnpm tauri dev, shortcuts work because the process inherits Terminal's Accessibility permission. For the built.app, grant Accessibility manually (see Installation).
Hotkey press → AudioEngine (cpal)
→ WhisperEngine (whisper-rs) → raw transcript
→ LlamaEngine (llama-server HTTP) → refined text
→ CGEvent simulate_paste() → target app
Voxa uses a decoupled MPSC channel to bridge the audio recording stream with the inference pipeline, keeping the UI fully responsive during transcription.
Key design decisions:
- No Dock icon — runs as
NSApplicationActivationPolicyAccessory(Alfred/Raycast model). Never steals focus. - Focus preservation — stores the frontmost app PID via
NSWorkspace, then restores it via PID-basedactivateWithOptions. Works reliably with Electron and JVM targets. - Native event tap — uses
CGEventTapat session level instead of Tauri's global shortcut plugin, which fails for system-reserved keys likeAlt+Spaceon macOS. - LLM inference —
llama-serverruns as a subprocess on a local port. Automatically selects Qwen2.5-3B (Apple Silicon) or Qwen2.5-1.5B (Intel) based on hardware. - Audio silence detection — uses peak amplitude instead of RMS to avoid false negatives on low-volume speech.
- macOS Native Event Tap & Shortcut Architecture
- VAD-Reactive Animation Architecture
- Long-Dictation Strategy for the Code Profile
Contributions are welcome. See CONTRIBUTING.md for guidelines.
Looking for a place to start? Check the open issues — issues labeled good first issue are a great entry point.
If you have an Apple Developer account and want to help with code signing and notarization, see docs/code-signing.md — this would remove the xattr installation friction for all users.
Voxa is built around a "Silent Conductor" philosophy: it lives at the edge of your screen, ready to translate your thoughts into text in any application, without friction, without noise.
It comes from an AI-first mindset — a constant curiosity about how AI fits better into daily work, and a conviction that the best AI tools are the ones you actually own. Local models are not a workaround. They are the right architecture: fast, private, and yours forever.