Voice input on Linux, Wayland, and Omarchy
Doubao Say listens on a global shortcut, shows a live transcript, and sends the text to the current app. Optional polishing uses an OpenAI-compatible model you configure. It is not an offline engine. Install steps stay in the install guide and INSTALL.md.
Boundaries
You need a network and the recognition service you pick. The default backend is a Doubao web login, which is not an official client protocol. The app is MIT-licensed. That license is not a promise that cloud recognition is free or uncapped. This page does not state accuracy, GPU use, or download counts.
The project is based on doubao-input-for-linux. Recognition and login components are credited to doubao-murmur in the repository README and NOTICE.
App or plugin
They share one engine. Install one.
- Omarchy: prefer the plugin. Finish first-run setup before
omarchy plugin enable. If add-time enable is offered, skip it. - Arch or Hyprland without Omarchy: the standalone app, from GitHub Releases or
./install.sh. - Native X11: clipboard paste is the documented path. Direct typing through
wtypeis not. - Other Wayland compositors: if the focus cannot be checked, the text stays for you to copy. Do not read that as automatic paste everywhere.
- Offline on-device ASR is not a documented backend.
Requirements: Python 3.11 or newer, GTK4, WebKitGTK, and PipeWire. The installer lists missing Arch / Omarchy packages before it asks to install them.
Three backends
If optional polishing fails or times out, the original transcript is kept.
What each desktop can do
XWayland is not treated as native X11. The microphone is selected through PipeWire.
First launch
The install guide order is: Doubao web login or an API key from the table above, a PipeWire microphone plus a three-second local check, then a shortcut and a trial utterance. Upgrades and uninstall stay in that guide.
Price
The shop lists Doubao Say as free and MIT, with a Doubao login and optional models billed separately. Free software is not a free speech quota.