OpenVerb

Press a hotkey, talk, get text — anywhere on macOS, fully offline.

npm install -g openverb-app

That's the whole install. The npm postinstall hook fetches the source, builds the C++ engine, downloads the Gemma 4 weights, sets up a stable code-signing identity in your login keychain, and copies OpenVerb.app into /Applications. ~5 minutes on a clean machine.

Why OpenVerb (vs Superwhisper, Wispr, Whisper Flow)

| | OpenVerb | Cloud dictation tools | |--|--|--| | Price | Free, MIT | $5–25 / month | | Audio leaves your laptop | Never | Yes — every word is a POST | | Brains | One model. Gemma 4 transcribes and polishes. | Two. Whisper for transcription, GPT‑4o for polish. | | RAM | ~3 GB resident | ~6 GB (two model loads) | | Works on a plane | Yes | No | | Apple Developer account needed | No | n/a | | Source available | Yes | No |

The "one brain" thing matters. Competitors transcribe with Whisper, then ship the text to GPT‑4o for cleanup. That's two model loads, two privacy boundaries, two latency hops, and one cloud round‑trip. OpenVerb uses Gemma 4 E2B — Google's audio + text Gemma — for both. Half the RAM, no API hop, no subscription.

Requirements

| | | |--|--| | macOS | 13.0 Ventura or newer | | CPU | Apple Silicon (M1/M2/M3/M4). Intel is not supported. | | RAM | 8 GB minimum, 16 GB recommended | | Disk | ~6 GB for the model weights | | Node.js | any recent version (just for the install command) | | Permissions | Microphone, Accessibility, Input Monitoring (granted on first launch) |

Usage

| Action | Default shortcut | |--|--| | Start / stop dictation | ⌥Space | | Cancel current recording | Esc | | Open settings | menu bar → ⚙ |

CLI:

ov              # launch the app (installs first if not yet installed)
ov update       # fetch latest source and rebuild
ov uninstall    # remove /Applications/OpenVerb.app

Don't have / want Node? Build directly from source — see BUILD.md.

Privacy

• No network calls at runtime. Block OpenVerb at the firewall and it works exactly the same.
• No analytics, crash reporters, or "anonymous usage stats."
• Audio buffers are in memory only and zeroed when a session ends.

The only network call ever made is the one‑time model download from Hugging Face during install.

Acknowledgements

• llama.cpp — the C++ inference engine
• Google Gemma — the audio + text model family
• Silero VAD — voice activity detection

License

MIT. See LICENSE.