Your Fcitx5 input method learns to listen

What it does fcitx5-vinput adds speech-to-text to Fcitx5, the input-method framework most Linux desktops use for CJK typing. Tap Alt_R to toggle recording, or hold it for push-to-talk; the transcript lands wherever your cursor is. Recognition runs either fully offline via sherpa-onnx models, or through cloud providers (Doubao, Aliyun Bailian, ElevenLabs, anything OpenAI-compatible), and you can switch engines at runtime from a command palette.

The interesting bit The transcript doesn’t have to be the end product. An LLM post-processing layer handles error correction, formatting, and translation through configurable “scenes” — and a command mode lets you select text, speak an instruction, and have the model apply it. Dictation becomes editing.

Key highlights

• Two trigger modes: tap-to-toggle and hold-to-talk, keys remappable in Fcitx5 settings
• Offline ASR via sherpa-onnx; a “lite” build ships without the ONNX runtime for cloud-only users
• LLM scenes for correction, formatting, and translation; voice commands on selected text
• Both a GUI (vinput-gui) and a CLI (vinput) for control
• Packaged for Arch, Fedora, Ubuntu/Debian, Nix, and Flatpak, on x86_64 and aarch64

Caveats

• Flatpak installs need extra permission overrides and an Fcitx5 restart — the sandbox doesn’t cooperate out of the box
• Local ASR requires downloading a model first via the GUI
• The README doesn’t say which languages the offline models cover, so check the docs before assuming yours is supported

Verdict Worth a look if you type CJK on Linux and want dictation without shipping your voice to a datacenter by default. If you never touch Fcitx5, this is someone else’s excellent idea.