Skip to content

Offline voice dictation on Linux

Short answer: Linux has no built-in dictation comparable to Windows Speech Recognition or macOS Dictation. YazSes fills that gap with hold-to-talk voice typing that runs entirely on your own CPU, works on both X11 and Wayland, and types into any focused application — editor, browser, terminal or chat.

sudo snap install yazses    # or: pipx install yazses
yazses doctor               # check mic, keyboard access and text injection
yazses enroll               # calibrate to your voice and room (~30 s)
yazses start                # hold the hotkey, speak, release

Why Linux dictation is usually hard

Three separate problems have to be solved, and most tools only solve one:

Problem What it means How YazSes handles it
Recognition Turning audio into text without a cloud API On-device faster-whisper, CPU int8 — no GPU, no network
Triggering Detecting a held key globally, across every window Reads keyboard events directly, and listens on all connected keyboards (laptop + external)
Injection Getting text into the focused app — the part Wayland broke Runtime probe picks ydotool, wtype, xdotool or clipboard paste

That third row is the one that trips up most Linux dictation projects. Wayland deliberately prevents one application from synthesising input into another, so tools built for X11's xdotool simply stop working.

X11 and Wayland

YazSes probes your session at runtime and selects a working injection backend rather than assuming one:

  • X11 — xdotool. Works out of the box on most distributions.
  • Wayland (GNOME, KDE, sway) — ydotool (via the ydotoold daemon) or wtype. yazses setup provisions these for you; yazses doctor tells you which backend was selected and whether it is functional.

You can pin the choice explicitly if the automatic probe picks badly:

[injection]
backend = "auto"   # auto | type | clipboard | wtype

On Wayland, typing (rather than clipboard pasting) is the default because it also works inside terminal emulators, where clipboard paste is frequently a no-op.

Terminals

If text appears in your editor but not in your terminal, you are almost certainly on a clipboard backend. Switch to type. See troubleshooting.

Distributions

Installation is covered in detail in the Linux install guide. In brief:

bash <(curl -fsSL https://raw.githubusercontent.com/MSKazemi/yazses/main/install-apt.sh)

sudo snap install yazses

pipx install yazses

Two Linux-specific gotchas

  • libportaudio2 is required for microphone capture and is not pulled in by pipx. Install it from your package manager if yazses start fails immediately.
  • Reading the keyboard needs membership of the input group. After sudo usermod -aG input $USER you must log out and back in for it to take effect.

yazses doctor checks both and tells you exactly which one is missing.

What "offline" actually means here

The model is downloaded once at install time. After that, dictation works with the network cable unplugged. No audio, no transcript, and no telemetry leave the machine — see the privacy statement for the specifics of what is stored and where.

This matters beyond privacy preference: it means dictation keeps working on a train, on a plane, on a locked-down corporate network, and on air-gapped machines. See private offline dictation.

Beyond typing

Once dictation works, the same daemon also recognises voice commands — "undo that", "save file", "go to line 42" — mapped to real key sequences by a fast regex grammar. Browse everything available with:

yazses features            # 139 capabilities, what's on, what's recommended
yazses features info code  # what one capability does, with an example

Where it falls short on Linux

  • Wayland window control is limited. Anything that needs to focus another window (such as the experimental Glance-Type gaze targeting) works on X11 only, because Wayland does not permit it.
  • No GPU acceleration path is configured by default — YazSes targets CPU int8 so it runs on modest hardware. If you have a strong GPU and want maximum speed, a whisper.cpp-based tool may transcribe faster.
  • Voice scripting is not the goal. If you want to program your desktop with a deep voice-scripting ecosystem, see Talon in the comparison.