1. Start from what Linux desktops actually include
There is no dictation switch in GNOME Settings or KDE System Settings as of October 2026. When Canonical announced Project Myna on June 17, 2026, it wrote that speech recognition “has become a common feature on modern platforms” and that it should be a first-class experience on Ubuntu Desktop as well, which is the clearest statement that the gap exists today.
Myna’s first release targets Ubuntu 26.10 on Wayland, with GNOME as the validated desktop. Canonical says it runs speech models locally, needs no internet once the models are installed, only opens the microphone when you start dictation, and processes audio in memory before discarding it. Voice commands, desktop control, translation and automatic language detection are out of scope. Ubuntu 26.10 entered beta on October 1, 2026 and is scheduled for release on October 15, 2026.
People searching for “Dragon for Linux” will not find one. Nuance sells Dragon Professional v16 as a Windows product, optimized for Windows 11 and backward-compatible with Windows 10, so the practical replacement on a Linux desktop is one of the native tools below.
2. Know whether your session is X11 or Wayland
Every Linux dictation tool does three jobs: it records from PipeWire or PulseAudio, runs a recognizer such as VOSK, whisper.cpp or NVIDIA Parakeet, and then fakes keystrokes into whatever window has focus. The last job depends on the display server, and it is where most setups fail.
On X11, xdotool can type into any window. On Wayland, xdotool only reaches X11 programs running under XWayland, so native Wayland apps receive nothing. Wayland desktops need a different helper: wtype on compositors that implement the virtual-keyboard protocol, such as Sway and Hyprland; kwtype on KDE Plasma; or dotool and ydotool, which create a virtual keyboard through the kernel’s uinput device and therefore also work under GNOME, at the cost of giving your user access to /dev/uinput. The Wayland guide covers each helper and its permissions.
- Run echo $XDG_SESSION_TYPE in a terminal; it prints x11 or wayland.
- X11 session: install xdotool (sudo apt install xdotool on Debian-based systems).
- Sway, Hyprland and other wlroots-style compositors: install wtype.
- KDE Plasma on Wayland: build kwtype, or use a uinput tool.
- GNOME on Wayland: dotool or ydotool, plus write access to /dev/uinput.
3. nerd-dictation: one script you wire up yourself
nerd-dictation is a GPL-3.0 Python script built on the VOSK API, and it keeps no background process running. You start recording with nerd-dictation begin and type the result with nerd-dictation end; the README tells you to bind begin, end and cancel to shortcut keys yourself. Setup is pip3 install vosk, a git clone of the repository, and a VOSK model such as vosk-model-small-en-us-0.15 unpacked into ~/.config/nerd-dictation/model.
It records with parec by default (sox and pw-cat are alternatives) and types with xdotool by default; the --simulate-input-tool option switches to DOTOOL, YDOTOOL or WTYPE for Wayland. A Python configuration file can rewrite text before it is typed, and output never presses Enter. The README is candid about limits: VOSK output is all lower-case, and the first start can be slow on older disks.
4. Speech Note: a Flatpak with many offline engines
Speech Note, published on Flathub as net.mkiol.SpeechNote under the MPL-2.0 license, combines speech to text, text to speech and machine translation, all processed on the computer. For recognition it offers Coqui STT, Vosk, whisper.cpp with Parakeet, Faster Whisper and april-asr, with separate NVIDIA and AMD add-on Flatpaks for GPU acceleration.
It is a note-taking app first, but flatpak run net.mkiol.SpeechNote --action start-listening-active-window turns it into voice typing that inserts text into the focused window. Its README says that works out of the box on X11, while on Wayland the ydotool daemon must be installed and running and the Flatpak must be allowed to reach its socket. Global keyboard shortcuts on Wayland go through the desktop portal, which the README says only recent KDE Plasma and GNOME releases support.
5. Handy and OpenWhispr: press a key and the text appears
Handy is a free, MIT-licensed app that transcribes on your computer: press a shortcut, speak, and the words land in the active field. It offers Whisper models and a CPU-friendly Parakeet V3 model, filters silence with Silero VAD, and as of version 0.9.8 (October 2026) publishes .deb, .rpm and AppImage builds for x86-64 and ARM64. Its Linux notes recommend xdotool on X11 and wtype or dotool on Wayland, describe Wayland support as limited, and tell Wayland users to bind handy --toggle-transcription in their desktop’s shortcut settings.
OpenWhispr, also MIT-licensed, runs local Whisper, Parakeet and other models or sends audio to cloud providers, and adds meeting transcription, notes and AI agents. Its Linux guide offers .deb, .rpm, AppImage and tar.gz packages for x86-64 only. On Wayland its own paste helper needs /dev/uinput access; otherwise it falls back to wtype, xdotool for XWayland apps, or ydotool, and if none is available it leaves the text on the clipboard for you to paste. Its help center also covers accounts and paid plans, so check that the features you rely on run locally.
6. Where IraVoice fits on Linux, and where it does not
IraVoice 1.2.2 is a commercial app built on Handy, and its Linux version is a beta. Its AppImage runs on any 64-bit distribution from 2022 or later, Fedora and Arch included, and updates itself; a .deb and an .rpm are there for people who prefer their package manager. There is no Flatpak, Snap or AUR package. Recognition runs on-device with NVIDIA Parakeet: the default Parakeet Unified EN model (about 700 MB) downloads once, and dictation then works offline.
Hold Ctrl+Space to talk, or tap it once for hands-free dictation and again to stop. IraVoice picks its typing helper by itself, using wtype, kwtype, dotool or ydotool on Wayland and xdotool on X11, though you may need to install one first. On Wayland its global shortcut may not reach the app, so bind iravoice --toggle-transcription to a desktop shortcut instead. Because it types into the local window, it also works in a terminal running SSH. When a word comes out wrong, Ctrl+Alt+Shift+F or History → Fix lets you correct it once and IraVoice types the fix from then on. The code-name spelling IraVoice offers on Mac and Windows is not available on Linux.
After a 3-day free trial that starts with your first dictation, IraVoice Local costs $79 once for one computer. If a free, open-source tool already types reliably on your desktop, keep it. IraVoice is for people who want a packaged .deb with Parakeet, replacements and snippets, and correction learning, paid for once rather than by subscription.