Voice dictation for Linux that learns you. Local, private, open source, and it fine-tunes itself on your own voice.
Try it in your browser at voiceio.dev: no install, the model runs in the tab.
- Types anywhere. Press your hotkey, speak, and the words stream into the focused app with a live underlined preview (IBus), on Wayland and X11: GNOME, KDE, Hyprland, sway, i3.
- Runs on your machine. Whisper decodes your speech locally. No account, no server, no telemetry; logs never contain your words.
- Learns your words. Keep your recordings (opt-in, on your disk) and voiceio fine-tunes its model on them while the machine is idle. It switches only when the new model wins on clips it never trained on. On the author's voice: 12.1% → 4.9% word errors after one night on a laptop CPU.
With a coding agent. Paste this into Claude Code, Codex or any agent with a shell:
Install voiceio on this machine following https://voiceio.dev/llms.txt. Ask me which hotkey I want and whether to keep my recordings on this computer so voiceio can learn my voice. Finish with voiceio doctor and fix whatever it reports.
By hand:
sudo apt install pipx build-essential python3-dev portaudio19-dev ibus gir1.2-ibus-1.0 python3-gi # Debian/Ubuntu
sudo dnf install pipx gcc gcc-c++ make python3-devel portaudio-devel ibus ibus-libs python3-gobject # Fedora
sudo pacman -S python-pipx base-devel portaudio ibus python-gobject # Arch
pipx install 'python-voiceio[desktop]'
voiceio setup # model, hotkey, autostartThen press your hotkey, speak, and press it again (or hold it while you speak). voiceio doctor shows what works and voiceio doctor --fix repairs what it can. NixOS, other extras and installing from source: docs/linux.md. Agent runbook: INSTALL.md.
voiceio vocab add Kalshi # a name, recognized from the next dictation
voiceio corrections add kalchi Kalshi # fix whatever still comes out wrong
voiceio learn schedule on # learn from your recordings while the machine is idle
voiceio learn status # what it learned and how each model scoresScheduled learning runs daily or weekly at a time you pick, only on an idle machine on mains power, with a notification you can stop. It never switches models unless the new one wins, and voiceio learn rollback undoes a switch. Details: docs/learning.md.
Everything lives in ~/.config/voiceio/config.toml; config.example.toml documents every option. voiceio config set section.key value edits it in place.
| Learning your voice | Vocabulary, corrections, labeling, fine-tuning, evaluation, schedule |
| Choosing a model | Whisper sizes, Parakeet, whisper.cpp servers, licenses |
| Commands | Every voiceio command |
| Your data | What is stored where, and what is off by default |
| Phone dictation | Run voiceio headless on a home server |
| Linux support and troubleshooting | Desktops, backends, NixOS, fixes |
| Comparisons | voiceio vs Voxtype, Wispr Flow, Superwhisper, Dragon and others |
- Unbounded vocabulary: a decoder whose biasing has no token cap (see model choice).
- GPU on AMD and Intel: let setup run a Vulkan whisper.cpp server for you.
- Typing without IBus on GNOME and KDE: a libei backend.
- Distro packages: AUR, .deb and .rpm.
- One model everywhere: sync your fine-tune across your machines.
Open an issue before a large PR. CONTRIBUTING.md covers the architecture, conventions and the reasoning behind the model choice. Install reports and ideas are welcome on the feedback board.
- Speech recognition: faster-whisper and CTranslate2, with OpenAI's Whisper models. The in-browser demo runs Moonshine through transformers.js.
- Ideas borrowed with thanks: modifier release, paste keys and media pausing from Voxtype; the landing page's look from Romario Kavin's KOLlateral.
MIT. Model weights are not part of voiceio and carry their own licenses (see docs/models.md).