macOS menu bar app for audio transcription that works system-wide in any application.
Two recording modes:
- Toggle Mode: Press your shortcut (Right Option or your favorite modifier like
Fn) to start, press again to stop - Hold to Speak: Hold your shortcut key, speak, release to transcribe
Text appears instantly in your current application—no copy/paste needed.
Status: Core features are functional. Planned enhancements tracked in docs/features/.
- 100% offline - Audio never leaves your device
- No data collection - Zero telemetry, zero tracking
- Open source - Verify for yourself what the code does
- Automatic language detection - Switch between languages mid-sentence
- 25+ European languages with Parakeet v3 (German, French, Spanish, Italian, etc.)
- No manual language switching - Just speak, FluidVoice adapts
- Sub-second transcription: 0.1-0.3 second response times
- 10x faster than traditional approaches with optimized model loading
- Instant response - Models preloaded and ready
- Real-time feel - Near-zero latency for short audio clips
- Ultra-fast pattern matching: 3-5ms correction (vs 1500-3000ms for LLM-based approaches)
- Privacy-first processing: All vocabulary correction happens locally, zero network requests
- Default presets: Includes common technical terms - "api" → "API", "github" → "GitHub", "typescript" → "TypeScript"
- Domain adaptable: Easily customize for any field - legal, medical, scientific, or business terminology
- JSONC configuration: Developer-friendly config with inline comments at
~/.config/fluidvoice/vocabulary.jsonc - Live reload: Edit vocabulary while FluidVoice runs - changes apply instantly
- Parakeet v3 Multilingual: NVIDIA's 600M parameter model optimized for speed and accuracy
- 25 European languages with automatic detection
- Daemon mode for zero cold-start latency
- Sub-second transcription on Apple Silicon
- Runs entirely offline via MLX
- Features: See
docs/features/for planned features anddocs/features/done/for completed implementations - Bugs: See
docs/bugs/for known issues anddocs/bugs/done/for resolved bugs
Looking for other voice transcription tools? Check out ALTERNATIVES.md for a detailed comparison of 8 competing products including VoiceInk, Spokenly, SuperWhisper, and more.
- macOS 14.0 (Sonoma) or later
- Install Xcode Command Line Tools:
xcode-select --install - Clone repository:
git clone https://github.com/mazdak/FluidVoice.git cd FluidVoice - Setup code signing (required for microphone access):
- Follow the guide: docs/setup-code-signing.md
- Takes ~5 minutes, completely free, no Apple account needed
- Build:
./build.sh - Run: Open
FluidVoice.app
Why code signing? macOS requires signed apps for microphone permissions. Self-signed certificates work perfectly and cost nothing.
Using just commands:
just # List all available commands
just dev # Build and run development version
just release # Build release version and install to /Applications
just test # Run tests
just logs # Stream app logs
just kill # Kill running app processes"Unidentified Developer" Warning
- Right-click the app and select "Open" instead of double-clicking
- Click "Open" in the security dialog
Microphone Permission Issues
- Go to System Settings → Privacy & Security → Microphone
- Ensure FluidVoice is enabled
- If permissions don't work after rebuilding, reset them:
tccutil reset Microphone com.fluidvoice.app
- Then restart the app and grant permission again
Parakeet Setup Issues
- Click "Download Parakeet v3 Model" in settings
- Check that download completed (~600MB)
See CLAUDE.md for development setup and guidelines.
Based on mazdak/AudioWhisper. Built with SwiftUI, AppKit, MLX, and Parakeet v3. MIT License.