Best Dictation Software for Linux (2026): 9 Options, and Why Your Text Never Arrives

Your desktop has no dictation built into it. Not GNOME, not KDE, not any distribution you can install today, and that is about to change but has not changed yet. What you can have instead, right now, is a free offline app that runs a modern speech model on your own machine and types into whatever window you are looking at.

By the VoiceBoard editorial teamUpdated August 28, 2026

The catch is not accuracy. Accuracy was solved years ago and has since got better again. The catch is the last two inches: getting recognised text out of the app and into the text field you were staring at. On Wayland that step fails, and it fails silently, which is why so many people conclude that Linux dictation does not work when in fact their transcription was perfect and simply went nowhere.

This page is organised around that problem. Prices, versions and package contents were read off the source on 27 August 2026.

Our picks

What changed, and why the other articles have not caught up

Four things moved recently, and between them they invalidate most of what is currently ranking.

Canonical is building dictation into Ubuntu. On 17 June 2026 an announcement went up on the Ubuntu Community Hub introducing Myna, described as "a new initiative to bring speech to text dictation to Ubuntu Desktop". It runs local models, the candidates named are Whisper, Nemotron and Qwen3-ASR, and "no internet connection is required once the necessary models are installed". The target is Ubuntu 26.10, and the first release "targets Ubuntu Desktop on Wayland, with GNOME as the primary validated environment". Do not go looking for a download: the canonical/myna repository was created on 5 June 2026, is GPL-3.0, and has no releases at all as of today. Voice commands, desktop control and automatic language detection are explicitly out of scope for the first version. Still, the opening line of every roundup in this category, the one about Linux being permanently left out, has a shelf life now.

Talon is walking away from Linux. Talon has been the answer for people who run a computer entirely by voice, and its own documentation for version 0.4.0 states the position plainly: "Wayland compositors lack the APIs necessary for Talon and Wayland support is not planned." The download page offers "Linux (X11)" and nothing else. On 31 May 2026 OSnews reported that Talon "will imminently remove ALL Linux support from the public release, as X11 continues to sunset", and quoted KDE developers disputing the technical reasoning: "Wayland implements all the APIs needed for Talon to work, but Talon's developers are simply not interested in using them." We have not seen a removal ship. We are flagging it because the people most affected are the people with the least room to switch.

"Just log into an X11 session" stopped being advice. This is the one that quietly invalidates the most pages. GNOME's own release notes tell the story in three lines: version 47 added the ability to disable X11 sessions, GNOME 49 disabled them by default when it shipped on 17 September 2025, and GNOME 50, released 18 March 2026, dropped support for running X11 sessions altogether, removing the display-server-specific unit files with it. Fedora 43 followed at the distribution level by deleting the GNOME X11 packages from its repositories and migrating those users to Wayland. KDE still offers an X11 session. GNOME does not. So every guide that shrugs and tells you to pick X11 at the login screen, including the installation guide of the tool most depended on by people who cannot type, is giving an instruction a large share of readers can no longer follow.

The tool sitting at the top of these search results is stalled. A short, honest post from January 2026 recommends Whispering, and it earned its position: the author dictated the whole post with the tool and said so. Since then its last release was v7.11.0 on 27 December 2025, and the README of the repository it lives in now says, word for word, "Whispering, vocab, skills, and the Epicenter host do not compile right now." The old binaries still download. Nobody is building new ones.

Why your text never arrives

This is the section the rest of the field skips, and it is the actual question.

Handy, the most used tool in this category, states it in one line in its own tracking issue: "Wayland has no single API for programmatic text insertion or registering global hotkeys." Both halves of dictation are broken independently, and Handy alone carries 48 open issues about it.

On X11, any program can ask the X server to synthesise keystrokes through the XTEST extension, which has been in the standard since 1992 and describes itself as "limited synthesis of input device events, almost as if a cooperative user had moved the pointing device or pressed a key". That is what xdotool type does. It works everywhere, in every app, with no permission of any kind. It also means any program you run can type into your bank's website and read every keystroke going to every other window. X11 has no notion of one app being isolated from another.

Wayland closed that hole on purpose, and it did not replace it with a restricted version. There is simply no request for "type this into whatever has focus", because a Wayland global cannot tell who is calling it, so anything exposed there is exposed to everything. xdotool's own README puts the consequence plainly: "Typing, window searching, and many other functions of xdotool do not work, and it is unclear if they could ever work."

So dictation tools pick a workaround, and each one works in some places and not others. There are five in the wild.

The virtual keyboard protocol (zwp_virtual_keyboard_v1), which is what wtype uses. Clean, Unicode-safe, and implemented only by wlroots-derived compositors. GNOME will not implement it, and said so in writing. Jonas Ådahl in 2021: "The virtual keyboard protocol is unlikely something we are going to support, where instead the plan for handling things like xdotool, Synergy, and similar things is using libei via a portal." Sebastian Wick, closing the follow-up request on 26 May 2025: "Yes, wayland doesn't have access control and we would have to expose this to every client. Having and maintaining multiple ways to do the same thing is also not sustainable." KDE has not implemented it either. Someone testing on Ubuntu 25.10 with KDE Plasma in November 2025 got the error verbatim: "Compositor does not support the virtual keyboard protocol." wtype's own tracker has carried this as issue #45 since 25 November 2022, and wtype itself has had no real commit since January 2022. If you are on GNOME or KDE, wtype is not your answer, however many articles list it.

The kernel route (/dev/uinput), which is what ydotool and dotool use. You create a virtual input device in the kernel and type through it. It works on every compositor, on X11, even in a TTY, for one reason: it sits below the compositor, so the compositor never gets a say. The clearest account of this came from a February 2026 write up about getting dictation running on COSMIC: both the hotkey detection and the text output "operate at the Linux kernel level, below the Wayland compositor".

The price is a permission most guides wave through in one line, and it deserves better. ydotoold needs write access to /dev/uinput, which normally means a udev rule and adding yourself to the input group. But systemd's default rules put every input device in that group. Joining it does not just let you write a virtual keyboard, it lets you read every keystroke from every real one, permanently, with no prompt. You are trading a compositor-mediated permission for exactly the capability Wayland was built to take away. Do it if you want, but do it knowing.

There is also a quality difference between the two tools that nobody mentions. ydotool sends raw kernel keycodes and its own docs admit it "does not recognize if the user is using a custom keyboard layout". dotool links libxkbcommon and is layout-aware. If your layout is anything other than US QWERTY, prefer dotool.

An input method (IBus). Instead of pretending to be a keyboard, the tool registers as the machinery that feeds text to applications in the first place, the same path that handles Chinese and Japanese input. It carries text rather than keycodes, so layouts stop mattering. This is the route Vocalinux takes on GNOME and KDE, and it is the most native-feeling of the lot when it works.

The clipboard plus a paste chord. Copy, then send Ctrl+V. It sounds trivial and it is the single messiest option in practice. The configuration guide for hyprwhspr is the best public catalogue of what goes wrong: terminals need Ctrl+Shift+V rather than Ctrl+V, so the tool has to know what app you are in, which Wayland also will not tell it, so on GNOME it resorts to the accessibility bus to find out; Emacs uses Ctrl+Y and treats Ctrl+V as scroll; on Dvorak or bepo the physical keycode for v is not v; on Thai, Russian or Arabic layouts no key produces a v keysym at all, so the tool briefly switches your input source to a Latin one and switches it back. And on GNOME specifically, because Mutter implements neither of the data-control protocols, putting text on the clipboard requires opening a surface and taking keyboard focus, which means momentarily stealing focus from the app you are dictating into.

The desktop portal with libei, which is the route the Wayland project actually intends. Your app asks the desktop for permission once, through a dialog, and then sends input down a sanctioned channel where the compositor knows who is sending it. This is the right answer and it is not finished. Until libei 1.6.0, released 15 May 2026, the interface carried keycodes and a keymap, so a dictation tool still could not type a character absent from your active layout. That release added an interface for sending UTF-8 text directly, which removes the problem entirely. It has not reached you yet: KDE has it on master only, shipping in Plasma 6.8 in October 2026, and the GNOME merge request has been open since 24 March 2026 and is currently in conflict.

Which one works where

Verified on 27 August 2026 against protocol support tables and compositor source trees.

COSMIC earns its own line, because it is the best documented failure in the category. There, wtype does not error out. It types the wrong thing. The February 2026 write up describes saying "hello world" and getting back 84 101 108 108 111, the keycodes themselves rendered as digits. And launched through COSMIC's own spawn action, wtype fails with no message at all.

One more trap worth naming: XWayland does not rescue you. Modern Xwayland can route XTEST calls through the portal, so xdotool key ctrl+v will still work inside an X application. But it forwards the raw keycode and lets the compositor interpret it, which silently discards xdotool's trick of rewriting the keymap to reach arbitrary Unicode. xdotool type with anything outside your active layout produces nothing. And wtype aimed at an XWayland window has its own separate long-running bug, filed repeatedly against Chromium and Electron apps since 2022.

The person with the best claim to be annoyed about all this is Jordan Sissel, who maintains xdotool. Testing the state of things in November 2025, he found GNOME 48 popping up "Allow remote interaction" every few minutes with no way to grant it permanently, and wrote: "I want to make this work but am at a loss for how to proceed with all the fragmentation. I don't mind what the protocol is, but I sure would love to have any protocol that does what I need. Is it worth it to continue?"

How to tell what broke, in about a minute

If dictation seems to do nothing, find out which half is broken before you change tools.

1. echo $XDG_SESSION_TYPE tells you x11 or wayland. Half the advice on the internet assumes the other one.
2. ydotool type "hello" with a cursor in a text box. If "hello" appears, injection works and your problem is audio or the model. If nothing appears, injection is your problem.
3. groups | grep input tells you whether you are allowed to use the kernel route at all.
4. pactl list sources short shows whether your microphone is the source the tool is actually recording from.

That sequence is worth more than any comparison table, including ours.

And know the shape of the silent failure, because it is everywhere in 2026 bug trackers. Vocalinux issue #523, filed 15 July 2026, is titled: the app "reported 'Text injection completed successfully' while the text never reached the app". Issue #485, from 30 June 2026, reports injection that "silently does nothing on COSMIC / wlroots compositors when IBus is active". Spokenly's tracker opens with the same class of bug on COSMIC, filed 8 August 2026. A green tick in the dictation window is not evidence that anything was typed.

Whisper is not the answer any more, and that changes what to install

Every article in this category assumes Whisper. That assumption aged out during 2025.

The Open ASR Leaderboard on Hugging Face is public, reproducible and updated continuously. We read it on 27 August 2026, when it held 67 entries. Here is where the models these tools ship actually sit, measured as word error rate, lower being better, alongside RTFx, which is how many seconds of audio the model processes per second of compute:

Parakeet is a full point of word error rate better than the biggest Whisper model and roughly an order of magnitude faster, which on a laptop CPU is the difference between text appearing as you finish talking and text appearing after an awkward pause. Handy ships 69 models across 14 families now, including Parakeet, Canary, Moonshine, Nemotron, Granite and Qwen3-ASR, and Whisper is one option among them rather than the point of the app.

Two consequences worth acting on. First, if a tool offers Parakeet V3, try it before you try Whisper. Second, anything built on Vosk is a generation behind, and that includes the tool most of these articles put in second place.

One caveat that keeps the numbers honest: the same model in two different applications does not perform the same, because the engine wrapping it and the post-processing after it differ. A user comparing two apps running identical Parakeet V3 put it precisely in June 2026: "This was with the same underlying Parakeet model, I think the differences are in the surrounding engine and post processing etc." Treat a leaderboard rank as a floor, not a promise.

At a glance

The tools

Handy

Best for: most people, most of the time.

Price: free, MIT licensed, 30,459 stars, last release v0.9.6 on 24 August 2026.

Handy is a small desktop app that records while you hold a key and drops the finished text where your cursor is. Its model catalogue has grown to 69 models across 14 families, all running locally: Parakeet, Canary, Moonshine, Nemotron, Granite, Qwen3-ASR and Whisper among them. Parakeet V3 is the one to try first, described in its own docs as CPU optimised with automatic language detection. There are AppImages for x86_64 and ARM, a .deb and two .rpms in the latest release, so installation is a double click on most systems. Nothing leaves your machine. It is, by a wide margin, the most used tool of its kind, and there is a reason people keep landing on it.

Now the part its marketing page leaves out and its documentation says out loud. Handy has "limited support for Wayland display server". You install the injection helper yourself: xdotool on X11, wtype or dotool on Wayland, and specifically ydotool on Ubuntu 26.04, which uses Wayland by default. Global shortcuts "must be configured through your desktop environment" rather than inside the app. The recording overlay ships disabled because "certain compositors treat it as the active window". And the docs admit that Whisper models "crash on certain system configurations".

None of that is disqualifying. It is a twenty minute setup instead of a two minute one. But if you are on GNOME or KDE with Wayland, note that the recommended helper, wtype, is the one your compositor refuses, so read the previous section before you start.

Vocalinux

Best for: GNOME or KDE on Wayland, which is the case most tools handle worst.

Price: free, AGPL-3.0. 783 stars, v0.16.0 released 23 August 2026, with nightly builds going out daily.

This is a Linux-first project rather than a cross-platform app with a Linux port, and it shows in the choices. Instead of trying to fake a keyboard, it registers as an IBus input method on GNOME and KDE, uses wtype where wtype actually works, falls back to xdotool for XWayland windows, and has a clipboard path for compositors that offer nothing else. That is the correct set of answers to the matrix above, and no other tool in this list has all four.

It runs whisper.cpp by default with Vulkan acceleration, so AMD, Intel and NVIDIA GPUs are all in play, and it will fall back to VOSK or PyTorch Whisper. Installation is an interactive script, an AUR package, a Flatpak build or pip install vocalinux. It wants Python 3.11 or newer and Ubuntu 24.04 or equivalent.

The honest counterweight: this is a young project with a small team, and its own tracker is where you can see the edges. Issue #738, still open from 25 August 2026, describes the keyboard layout desynchronising between GNOME, IBus and XWayland mid dictation. Issue #664 from 10 August has dictation quietly switching a Brazilian ABNT2 layout to US. Issues #485 and #523 are the silent injection failures quoted earlier. The website's compatibility table marks every major compositor "Full support", which is more confident than the tracker justifies. We would still start here on GNOME Wayland, because the alternative is doing the same work by hand.

Voicy

Best for: people who want dictation, not a project.

Price: 30 minutes of recording free with no card. Pro is $8.49 a month, with 20% off if you pay yearly, and a lifetime licence at $260. Teams are $6.79 a seat from three seats. Students and people with a disability get 20% off. Checked on the pricing page, 27 August 2026.

Here is what buying it actually feels like. You install a package the way you install any package, you pick a hotkey during setup, and then you talk into whatever window you are in and finished text appears with the punctuation already in place. Firefox, VS Code, a terminal, a Slack box, the search bar. No model to download, no GPU to think about, no helper daemon to configure. If you have spent an evening on ydotool permissions you know exactly what that is worth.

We wanted to know whether the Linux build was real or a checkbox, so we went and looked, and this is the part nobody else in this category does. All three downloads answer with HTTP 200 and are substantial: the .deb is 134 MB, the .rpm 119 MB, the AppImage 174 MB. All three carry a Last-Modified date of 26 August 2026, the day before we checked. Inside the .deb, the control file reads Version: 2.4.5, Architecture: amd64, and an installed size of about 465 MB. The dependency line is the interesting bit: libxtst6, libnotify4, libasound2, libappindicator3-1, wl-clipboard, gjs. libxtst6 is XTEST, the X11 injection path. wl-clipboard is the Wayland path. And the post-install script says the rest out loud in a comment: it grants "the active user write access to /dev/uinput so Voicy's Wayland paste injector can open it", loading the uinput module, writing a udev rule tagged uaccess, and creating the input group if it does not exist. AppImage users get the same thing through a pkexec prompt at runtime. It also adds its own apt repository so it updates with the rest of your system.

One detail in that script is worth crediting, because it is better than what most guides tell you to do. The udev rule is tagged uaccess, which hands the ACL to whoever is logged in at the seat, with the group as a fallback for systems that do not support it. So it does not put you permanently in the input group, and you do not end up able to read every keystroke from every keyboard as a side effect. It grants the ability to write synthetic input, and stops there. That is the narrower and more correct version of the trade.

So: it is an Electron app that solves Wayland the same way ydotool does, and its whole Linux value proposition is that it performs the privileged setup for you instead of leaving you a wiki page. That is a fair thing to charge for. It is not magic, and anyone telling you it is a different class of technology is selling.

Two limits stated plainly. There is no offline mode, and the package proves it. Voicy's own comparison table says "Works offline: No", and there is not a single local inference artifact inside the download: no ggml weights, no ONNX, no whisper.cpp, nothing. What is in there are hardcoded WebSocket endpoints, wss://us.usevoicy.com and wss://eu.usevoicy.com. Its security document describes the route: audio goes from your machine to Voicy's servers on Heroku in the USA, is "immediately forwarded to Groq.com, which hosts the open-source Whisper V3 transcription model", and neither party stores it. Take that model name as the vendor's own statement rather than as something we could verify from outside, but if it is accurate then you are paying for the plumbing and the speed, not for a better model than the free tools run locally. Zero retention is a real and meaningful policy. It is still not the same as your voice never leaving the room.

Second, we did not test it on a live desktop. We read the package, not the experience. In particular we cannot tell you how the clipboard paste path behaves in a terminal, where Ctrl+V is not paste.

The 99%+ accuracy figure on the marketing pages is the vendor's own number, unverified by anyone including us.

See Voicy

Speech Note

Best for: turning recordings into text, and for languages nobody else supports.

Price: free, MPL-2.0. 1,610 stars, code updated the day we checked. Current desktop release is 4.8.4 from 15 April 2026: version 4.9.0 exists but shipped for Sailfish OS only, with the desktop build promised later.

Speech Note is a note taking application first and a dictation tool second, and it is worth having for the first job alone. It carries 146 speech recognition models across 42 languages, split across five engines: Vosk, Whisper, Faster Whisper, April-ASR and the old Coqui STT. It also does text to speech and offline machine translation. Everything runs locally: "no data is sent to the Internet." Install it from Flathub, or from the AUR, openSUSE Packman, or OpenRepos if you are on Sailfish. Budget the disk: the full Flatpak is around 1 GiB to download and 3.6 GiB installed, though the Tiny variant is 48 MiB.

It does have an "insert into active window" mode, so it can act as system-wide dictation, and the documentation is specific about the price on Wayland: "the external ydotool daemon must be installed and running for it to work. If you are using Flatpak, also make sure that the application has permission to access the ydotool daemon's socket file."

The honest limitation is the model generation. That library is the pre-2025 lineup: no Parakeet, no Canary, no Nemotron. For breadth of languages nothing here comes close. For raw accuracy in English, a Parakeet-capable tool will beat it.

Where it fits: if your job is a folder of interview recordings, or a language the Whisper wrappers have not thought about, this is the strongest free thing on Linux and it is not close. If your job is typing into Slack all day, one of the push-to-talk apps will feel better.

Voxtype

Best for: Sway, Hyprland, niri, river, and people who like a config file.

Price: free, MIT. 1,270 stars, last push 25 August 2026.

Voxtype is push to talk built by someone who clearly uses a tiling compositor. It runs whisper.cpp by default and can swap in Parakeet, Moonshine, SenseVoice, Paraformer and others through ONNX, all locally, and it claims 9 to 11 times realtime on CPU. It ships through the AUR, .deb, .rpm, AppImage, Homebrew and cargo. It needs glibc 2.38 or newer and your user in the input group.

What makes it worth naming separately is that it is honest about the matrix. It tries wtype first, switches to dotool when your keyboard layout is not US, and drops to ydotool for X11 and TTYs. Its own documentation states that on KDE and GNOME wtype is unsupported and you need dotool or ydotool. That single sentence is more useful than the Wayland coverage in every listicle currently ranking.

Speed of Sound

Best for: people who want the officially blessed route rather than a clever hack.

Price: free, MIT. 197 stars, last push 13 July 2026.

Alone in this list, Speed of Sound types through XDG desktop portals, which is the mechanism the Wayland project actually intends applications to use for this. No input group, no udev rule, no virtual keyboard protocol lottery. Transcription is local through Sherpa ONNX with Whisper, Parakeet or Canary, and there is optional text polishing through an LLM, which you can point at your own Ollama or llama.cpp if you want it to stay in the building. Flathub, Snap, AppImage, .deb and .rpm.

The reason it is not our top pick is momentum. It is a younger, smaller project with 61 open issues and a quieter release rhythm than Handy or Vocalinux, and portal coverage still varies by desktop. But architecturally it is where this category should end up, and if the portal route matters to you, install this one first.

Google Docs voice typing in Chrome

Best for: one document, on a machine you cannot install software on.

Price: free.

Open a Google Doc in Chrome, press Ctrl+Shift+S, talk. Google's own help page lists the requirement as "the latest versions of: Chrome, Edge, Safari" and covers around 80 languages and variants; in Slides it works on speaker notes and captions. Note that Linux is neither listed nor excluded on that page, so treat it as unsupported-but-working rather than officially supported.

The limitation is total and obvious: the words land in the Google document and nowhere else. For an essay, a locked down work laptop or a library machine, that is a fair trade and it costs nothing. For daily use you will get tired of copying and pasting by about Thursday. It also sends your voice to Google, which for a portion of this article's readers ends the conversation on its own.

Talon

Best for: running an entire computer without hands, if you can stay on X11.

Price: free to download. A paid Patreon tier unlocks beta builds and priority support.

Talon is not a dictation app, it is a hands-free input system: voice commands, noise control where a sound becomes a click, eye tracking, and Python scripting for anything you want to define. Its speech engine, Conformer D, runs on your machine. For people who code by voice because typing hurts, it has been the answer for years, and nothing else in this list attempts what it does.

Read the status carefully before you invest the weeks it takes to learn.

The download page offers "Linux (X11)". The documentation for version 0.4.0 says: "Wayland compositors lack the APIs necessary for Talon and Wayland support is not planned." On 31 May 2026 OSnews reported that all Linux support is being removed from the public release, and quoted KDE developers pushing back on the reasoning: "Wayland implements all the APIs needed for Talon to work, but Talon's developers are simply not interested in using them."

There is also a fact that puts the rest in context. We checked the headers on all three public downloads, Linux, macOS and Windows. Every one carries Last-Modified: 24 July 2023. The public stable build is three years old on every platform, not just this one. Active development lives in the paid beta channel. That is a legitimate way to run a project, and it is worth knowing before you build a workflow on the free build.

And here is where it gets genuinely bad, rather than merely awkward. Talon's installation guide says: "Talon, like many tools for automation or accessibility, does not support Wayland. You will have to select an X11 session from your login manager." As of GNOME 50 in March 2026, there is no X11 session in your login manager to select. For a GNOME user who runs their computer by voice because they cannot use a keyboard, the documented workaround no longer exists on their machine. Nothing else in this category attempts what Talon does, which is why this is worth stating bluntly rather than folding into a pros and cons list.

If you need voice control rather than dictation and you are already on Wayland, look at numen, free software aimed at the same job, which types through /dev/uinput and therefore works on X11, Wayland and in a TTY alike. It is a smaller thing than Talon. It is also a thing that runs.

nerd-dictation

Best for: nothing, in 2026, and it is still in every article.

Price: free, GPL-3.0. 1,913 stars.

This is the tool the whole genre recommends, and it deserved that once. It is a single Python file with almost no dependencies that pipes VOSK into xdotool, and it was the first thing that made Linux dictation feel possible. Its author was clear about the deal, including the line in the README that explains most of the complaints about it: "Text from VOSK is all lower-case."

Two things have changed underneath it. Its last commit was 10 October 2025, roughly ten months ago, with 83 issues open, several about Wayland and keyboard layouts that have been open since 2021. And VOSK is no longer competitive. It does not appear on the Open ASR Leaderboard at all, its last release was v0.3.50 in April 2024, and its accuracy falls apart outside clean audio: the small English model sits near 9.85% word error rate on clean read speech, and the large model hits 29.78% on call centre recordings. A developer working on Arch with Wayland in January 2026 described the practical result, Vosk producing "sporadic hallucinations: during silence, Vosk would randomly output 'the' or 'there'", along with punctuation arriving as literal words, and concluded that you should "skip Vosk entirely".

One more thing worth knowing if you ship anything built on it: Vosk's models are not uniformly permissive. vosk-model-small-fr-pguyot-0.3 is CC BY-NC-SA, which forbids commercial use, and vosk-model-fr-0.6-linto is AGPL. Check the licence of the specific model, not the toolkit.

If you already run it and you are happy, nothing here obliges you to move, and Elograf gives it a maintained graphical front end. If you are starting today because a listicle told you to, start somewhere else.

Where your voice goes

For a Linux audience this is not a footnote, so here it is flat.

Audio never leaves your machine: Handy, Vocalinux, Speech Note, Voxtype, Speed of Sound (unless you switch on LLM polishing), Talon, nerd-dictation.

Audio goes to a server: Voicy, Google Docs voice typing, NovaVoice.

Read that column twice. Every tool in the list above that keeps your voice at home is free, and every one of them is open source. That is not the case on Windows or macOS, and it is the single strongest thing this platform has going for it in this category. If local processing is the reason you are on Linux at all, you do not have to compromise and you do not have to pay.

There is a second privacy question specific to Linux that nobody frames properly. Giving a dictation app the /dev/uinput access that most Wayland-capable tools here need buys more than the ability to synthesise input. Because of how systemd's default rules group input devices, joining the input group also lets anything running as you read every keystroke from every physical keyboard on the machine. That is the exact capability Wayland removed on purpose, handed back by a one-line install step. The tools ask for it because there is no better universal option yet, and the portal route is the way out, which is part of why we rate Speed of Sound's approach even though the app is smaller. If a tool sets this up with a uaccess udev tag instead of putting you in the group, as Voicy's package does, it is asking for the narrower version of the permission and that is worth preferring.

While we are being even-handed about it: KDE Plasma has a hole of its own that nobody has closed. KWin still exposes a legacy fake input protocol to every Wayland client unconditionally, and its authorisation function is a stub with a // TODO: make secure comment above a line that approves the request. On Plasma Wayland, any application you run can inject keystrokes with no prompt. That is not an argument against KDE, it is an argument for reading source instead of security pages, and it is the same reason we opened the Voicy package instead of trusting its marketing.

And one warning that applies to every Whisper based tool on this page: Whisper hallucinates on silence. It will occasionally produce "Thank you" or a stray phrase when you said nothing at all, a behaviour tracked in Vocalinux issue #331 from 18 March 2026 and complained about widely. It is a property of the model, not a bug in the app you chose.

What we checked, and what we did not

We did not run an accuracy benchmark, and we are not going to imply otherwise.

Here is the thing about this particular field: nobody else has either. We read every article ranking for these queries and there is not one measured number in any of them. No word error rate, no latency, no memory figure, no timing, and almost no screenshots. The comparison tables all carry an "Accuracy" column, and every cell in every one of them is an adjective. Which is doubly hollow, because two apps running the identical Parakeet V3 model produce visibly different results depending on their engine and post-processing.

So instead of inventing numbers, we went and got the ones that already exist and are reproducible, from the public Open ASR Leaderboard, and we verified everything else that is checkable at the source. On 27 August 2026 we:

• Read the Open ASR Leaderboard on Hugging Face, 67 entries at the time, and pulled the rank, word error rate and throughput for the models these tools actually ship.
• Downloaded the first 400 KB of the Voicy .deb, extracted control.tar.xz, and read the version, architecture, dependency list and post-install script. That is where the libxtst6, wl-clipboard and /dev/uinput findings come from, and you can repeat it in two minutes with curl -r and ar x.
• Sent HTTP HEAD requests to all three Voicy Linux download URLs and recorded the status, size and Last-Modified header of each, then did the same for Talon's three platform downloads, which is how we know the public build is from July 2023 everywhere.
• Searched the Voicy package for local inference artifacts and found none, and for network endpoints and found two.
• Queried the GitHub API for every open-source candidate: stars, licence, creation date, last push, open issue count, archive status, and the release list with dates.
• Read the projects' own documentation for how each one injects text, and quoted the limitations the maintainers state themselves rather than paraphrasing them.
• Read the vendor documentation for the products people ask about but cannot have, including Wispr Flow's supported devices page, and WineHQ's application database entry for Dragon.
• Read the Ubuntu Community Hub announcement for Myna and checked the repository for releases.
• Read dated user reports where they could be linked: GitHub issue threads, Hacker News comments through the Algolia API, Unix StackExchange, and a Linux Mint forum thread that runs from June to October 2025.

Two things we are flagging rather than glossing over. We did not install and use any of these tools on a live desktop, so the compositor matrix here is assembled from documentation, issue trackers and other people's dated hands-on reports, not from our own machine. And Reddit is not reachable from where we work, so every community voice on this page comes from a source you can click.

The competition

Tools and answers we looked at and left out, and why.

Wispr Flow is the most searched name in this whole category by people on Linux, and the answer is no. Its documentation lists supported devices as "Mac, Windows, iOS, Android" and then says: "iPad, Linux, Chromebooks, and virtual machines or remote desktop environments are not supported." A separate help article is blunter still: "There is no native Linux application." If you run Linux in WSL or a VM on a Windows or Mac host, Flow can paste into it from the host side, and that is the whole of its Linux story.

The unofficial Wispr Flow repackage at wispr-flow-linux exists, has around 100 stars, and pairs the Electron app with a clean-room Rust helper that reimplements the text injection Wispr ships only on macOS and Windows. It even has a --doctor command that checks your session type and /dev/uinput access, which tells you what the hard part was. We are naming it because you will find it, not recommending it: it repackages someone else's proprietary application, and that is a decision to make with your eyes open.

Superwhisper, VoiceInk, Aqua Voice and Typeless all get searched with "linux" appended, and none of them has a Linux build. VoiceInk is macOS on Apple Silicon only.

Dragon NaturallySpeaking under Wine is the oldest answer in this category and it has never worked the way people hope. WineHQ's own application database, which is where this gets settled rather than argued, rates Dragon 13, 15 Home and 15 Professional all as Garbage. The best rating any version ever earned is Silver, for version 12.5, marked obsolete, last tested on 27 January 2021. And even that entry contains the sentence that ends the idea: "Dictation works only within the wine environment... Text can be cut and pasted into any application." So the best case, on a version from over a decade ago, is dictating into a Windows window and pasting out of it. Be fair to the record though: nobody has retested Dragon on a current Wine build since 2022, so "it fails today" is an assumption, not a measurement. The workaround people actually adopted was a Windows virtual machine relaying text over a socket, which a developer described building on Hacker News in 2023. If you came here from a Dragon search, take the local Parakeet or Whisper route instead. First-pass accuracy will surprise you and it costs nothing.

Buzz has 21,162 stars and turns up in these lists constantly. It transcribes files and microphone input into its own window. There is no type-into-the-focused-app feature and nobody has asked for one. Excellent at its job, which is not the job on this page.

Wspr is a $14.99 one-time purchase with local whisper.cpp and a .deb, which sounds ideal until you read the requirements: it types through xdotool via XWayland, and its own documentation says "pure-Wayland setups are not supported yet". Cheap, honest, and specifically not for the setup most readers have.

Spokenly deserves a mention for having a genuinely free local tier with bring-your-own-key, Pro at $99.99 a year, and the best-organised Wayland section of any vendor blog. It also has an open issue, filed 8 August 2026, reporting that text does not paste into the target app on COSMIC because the overlay steals focus. It sells the fix to a problem it has not finished fixing.

NovaVoice is the other commercial app with a Linux client, .deb and AppImage for Ubuntu 20.04+, Debian and Fedora, at $10 a month. Cloud only, and its Linux page says nothing at all about X11 or Wayland, which for this category is the whole question.

Simon, CMU Sphinx, Julius and IBM ViaVoice appear in roundups published this year. They are being copied out of a Wikipedia page whose sources are mostly from 2012 and which does not contain the word Wayland anywhere. ViaVoice left the market in 2002. DeepSpeech was archived after v0.9.3 in December 2020. Coqui STT wrote its own epitaph in its README, "this project is no longer actively maintained... we've seen focus shift towards newer STT models such as Whisper", and the company behind it shut down on 3 January 2024. Kaldi is the exception in that list: it is still getting commits, but it is a research toolkit for building recognisers, not something you install to dictate an email. None of them appears on the ASR leaderboard. Skip any article that recommends them as dictation apps, which unfortunately means skipping several of the ones above this page.

Whispering has our sympathy and a caveat. Good tool, real fans, and the top-ranking article in this category dictated itself with it. Its upstream repository currently says it does not compile. The 27 December 2025 binaries still install fine if you want it.

Myna is not here as a pick because you cannot install it. Come back at Ubuntu 26.10.

The wrapper crowd is now an industry. OpenWhispr has 5,781 stars, OpenLess 3,334, hyprwhspr 1,176, and Dictee, which is tiny at 56 stars but is the most interesting design in the pile: Wayland-first, KDE plasmoid, Parakeet TDT v3 by default, layout-aware injection through evdev. Beyond those: Sotto, VOXD, Blurt, whisrs, lokstt, hyprvoice, waystt, LinuxWhispr, Vibe Typer, and more arriving monthly, several written by someone who wanted Wispr Flow and could not have it. Some are genuinely good. We are naming the category rather than pretending to have tested each member of it. If you want to go exploring, our open source dictation guide goes deeper, and the full catalogue of what we cover lives in dictation apps.

One pattern to arm yourself against while you explore: at least three of the paid Linux dictation apps each describe themselves as "the only" paid dictation app for Linux. They cite each other's blogs. Check the download page, not the claim.

Prices, versions, package contents and project activity were read from vendor pages, official documentation, the GitHub API and the packages themselves on 27 August 2026. This page is rechecked quarterly, and sooner when Ubuntu 26.10 ships.