Open Source Dictation Tools, Checked One by One (2026)

Search for open source dictation and you land in one of two places, and neither of them is answering your question.

By the VoiceBoard editorial teamUpdated August 28, 2026

The first is a comparison of speech recognition models, ranked by word error rate, published by a company that sells speech recognition as an API. Useful if you are building a product. Useless if you want to talk into an email. The second is the home page of one specific app that happens to rank for the phrase, where the answer to "which open source dictation tool should I use" is, predictably, this one.

Sitting between them is a good curated list on GitHub with more than forty projects on it. No dates, no licences, no opinion. Forty is worse than none, because most of those forty were published in the last eighteen months by one person on a weekend, and you have no way to tell which.

So we did the boring part. We pulled the licence file, the star count and the last release date for two dozen projects on one day, and read the complaints of people who actually tried to install them. Here is what survives.

Checked 27 August 2026. Every licence, star count and release date below was read that day from the project's own repository or the vendor's own page.

The field at a glance

Three things in this field are not what the label says

This is the part no listicle does, because it takes ten minutes and annoys people.

FUTO Keyboard and FUTO Voice Input are source-available, not open source. The code is published, the app runs entirely offline, the privacy claim holds up, and for a lot of Android users it is genuinely the best answer. But the licence file is called the FUTO Source First License 1.1, and it says: "You may modify the software only for non-commercial purposes" and "You may distribute the software or any part of its source code only if you do so free of charge for non-commercial purposes." That is not an open source licence by any accepted definition, and if you are choosing FUTO because open source is your requirement, you should know that the requirement is not being met.

Talon's core is proprietary. Talon is the serious option for people who cannot use their hands and need to drive the whole machine by voice, and its user scripts are open and shared widely, which is where the confusion comes from. The application itself is not. Its EULA states that "Talon binaries, updates, files, code, and any works produced from reverse-engineered Talon code... are not permitted to be redistributed in public or private." One more thing worth knowing before you plan around it: the documentation says Linux support is X11 only, and that "Wayland compositors lack the APIs necessary for Talon and Wayland support is not planned."

Open source and free are not the same word. VoiceInk publishes everything under GPL-3.0. You can read it, audit it, and build it yourself at zero cost, and the BUILDING.md file in the repo tells you how. The developer also sells a signed build for $29 with automatic updates and support. Both of those statements are true at once, and this is the healthiest arrangement in the category rather than a catch: someone is paid to keep it working, and if they stop, the code is still yours.

Almost all of them run the same three engines

Once you notice this, the choice gets much easier.

Whisper (OpenAI, MIT, 108,018 stars) is the multilingual default, covering 99 languages, and it is what most desktop clients ship first. It is fine on a modern machine and slow on an old one.

Parakeet (NVIDIA, the parakeet-tdt-0.6b-v3 model on Hugging Face, CC-BY-4.0, commercial use allowed) is the fast one. It covers 25 European languages, needs about 2 GB of RAM, and runs at usable speed on a CPU, which is why Handy's author says he recommends it by default. A MacWhisper user running it on an M1 laptop put the difference this way on Hacker News on 18 August 2025: "Transcription speed of recordings feel at least 10x faster than Whisper, and the accuracy is better as well."

Vosk (Apache-2.0, 15,080 stars) is the old guard: tiny models around 50 MB that run on a Raspberry Pi, 30-plus languages, and the engine behind nerd-dictation. Note the date though. Commits are still landing (the most recent we saw was 9 August 2026), but the last tagged release is v0.3.50 from 22 April 2024. If you install a Vosk-based tool today, you are installing a two-year-old package.

The practical conclusion: transcription quality is mostly decided by the model, and everyone has access to the same models. What you are actually choosing between projects is the wrapper. Does the hotkey work, does the text land in the right window, does it survive an OS update, is anyone fixing bugs.

On a Mac

The Mac has the most open options because it has the most people writing them.

VoiceInk is the one we point people to, and the reason is boring rather than ideological: it is the only project here where the open code and the maintained build come from the same person and both are current. The licence file is plain GPL-3.0. Version 2.13 shipped on the day we checked. It runs Whisper and Parakeet on the Neural Engine, so nothing leaves the machine, and the optional cloud step touches only text you have already dictated, never the audio. In practice it feels like the paid apps rather than like a project: a personal dictionary for names and jargon, modes that switch settings depending on which app you are in, a keyboard shortcut that behaves the way you expect on the first day. There is an iOS app too.

The limits are hard and worth saying plainly. Apple silicon only, macOS 14.4 or later. On a 2019 Intel MacBook this is not an option at any price. Local models want RAM, and the first model download takes a while on a slow connection.

If you are on a recent Mac and you want local dictation you do not have to babysit, try VoiceInk. It is $29 once for one Mac, $49 for two, $69 for three, with a 14-day refund, or free if you would rather open Xcode and build it. Our longer write-up is in the VoiceInk review.

FluidVoice is the free Mac option with the most momentum: GPL-3.0, 10,970 stars, commits on the day we looked, and a Windows build now too. OpenSuperWhisper (MIT, 2,727 stars) is smaller and simpler. Handy works on the Mac as well and is the one to try if you want the same tool on every machine you own.

On Windows

Handy is the practical default. MIT, 30,459 stars, v0.9.6 on 24 August 2026, installable through winget or a normal installer, and it does the one thing you want: press a shortcut, talk, release, the text appears where your cursor is. Silence gets filtered out before transcription, and you pick between Whisper sizes and Parakeet in the settings.

Two honest notes about it. The author is candid on the project's own about page, quoted in his Hacker News thread on 28 September 2025: "This is my first Rust project, and it shows. There are bugs, rough edges, and architectural decisions..." And it is worth knowing that open source desktop apps of this type set off antivirus software. When Whispering shipped version 7.3.0 in August 2025, a commenter posted VirusTotal results on 18 August showing AVG and Avast flagging the installer as potentially unwanted. As another commenter put it in the same thread, an app that types into other windows behaves exactly like the thing security software is built to catch. Unsigned builds from small projects will keep triggering this, and you will have to decide whether you trust the source.

OpenWhispr is the other current option: MIT, 5,780 stars, macOS, Windows, Linux and iOS from one codebase, local models by default with a paid cloud tier you can ignore.

Windows also has free built-in dictation (Win+H), which is worth knowing about before you install anything, but it is a cloud service. Microsoft's own documentation says it requires an internet connection. Free and private are different problems, and the built-in tool solves only one of them. Our full Windows rundown is in the best dictation apps for Windows.

On Linux

This is where open source is not a preference but the only category that exists, and the honest picture has some sharp edges in it.

Handy runs on Linux and is the biggest project with real momentum. Wayland is the caveat. As of late August 2026 there are open issues and pull requests for global shortcuts and text insertion under GNOME Wayland, filed between 27 April and 13 August 2026, several of them still under discussion. On X11 it behaves. On Wayland, check the issue tracker for your compositor before you plan your day around it.

Speech Note is the calmest option on Linux and the least discussed. MPL-2.0, distributed as a Flatpak on Flathub, version 4.9.0 from 28 June 2026, and it packs speech to text, text to speech and offline translation into one app with optional GPU add-ons for AMD and NVIDIA. Everything runs locally with no network connection. It is a note-taking app first, so system-wide dictation is not its centre of gravity, but for writing in a window and getting text out it works without you compiling anything.

nerd-dictation deserves a paragraph mostly as a warning about how these lists are written. It is a well-made single-file Python script over Vosk, it converts spoken numbers into digits, and its configuration is plain Python so you can rewrite the output any way you like. It also has no tagged release, ever, and its last commit was 10 October 2025. A commenter on Hacker News summed it up back in January 2022 and the description still fits: "Definitely more diy/hacker focused than actually being a solution though." It appears near the top of nearly every "best open source dictation for Linux" listicle published this year, usually with no indication that it sits on an engine whose last release is from 2024.

If you like scripts, that is a legitimate path and it is about twenty lines. One Hacker News reader posted his in August 2025: bind a key to start ffmpeg recording, bind the release to kill it, pipe the file through whisper-cli from whisper.cpp, type the result. No installer, no antivirus warning, nothing to update. Whether that is a delight or a chore is the whole question in this section. Our platform guide is at best dictation apps for Linux.

On Android

Transcribro (ISC licence, 739 stars) is the clean answer: a keyboard and recognition service that runs entirely on the device, built on whisper.cpp, available through F-Droid. Development is quiet, with the last commits we saw dated 29 August 2025, so treat it as stable rather than actively growing.

FUTO Keyboard and FUTO Voice Input are the well-known pair. They are fully offline, take no account, and are funded by an optional one-time payment instead of a subscription. Just read the licence section above before you file them under open source, and take the project at its own word that it is still filling in features and languages.

The free baseline on Android is Gboard's own voice typing, which works offline once you download a language pack, and which most people already have installed.

Where open source still loses, specifically

Not vaguely. Here are the five places, with sources.

1. Nobody cleans up your speech. This is the real gap and it explains most of the disappointment. Paid cloud tools do two steps: transcribe, then rewrite. The rewrite turns "so I think we should, um, maybe move the meeting, actually let's move it to Thursday" into a clean sentence. Local models do the first step only. The maintainer of Whispering said it plainly on Hacker News on 19 August 2025: "Recent transcription models don't have very good built-in inference/cleanup, with Whisper having the very weak 'prompt' parameter", and the best available workaround is to pipe the transcript into a separate language model afterwards. You can build that. Several projects now ship it. But out of the box, what you get is what you said, filler and all.

2. Whisper invents things during silence. In "Careless Whisper: Speech-to-Text Hallucination Harms" (FAccT 2024), Koenecke and colleagues found that 1.4% of audio segments produced a hallucinated sequence, text that was never spoken, and that 38% of those hallucinations contained explicit harms such as invented violence or false attributions. Hallucinations clustered around pauses in speech. For dictating an email this is a curiosity. For anything you will not reread, it is not.

3. Installation is a real cost. Antivirus flags on unsigned builds, model downloads that fail without a useful error, permissions dialogs, Wayland. One user described trying the Whispering AppImage on Linux in August 2025, failing to download a model, wiring the models up by hand, and then getting a zero-length audio file, concluding on 21 August: "this software isn't ready for everyday use yet, at least not on Linux." That is one project on one day and it has moved on since, but the shape of the experience is common enough that pretending otherwise would be dishonest.

4. Support is an issue tracker. There is no help desk. If the hotkey stops working after an OS update, your options are to wait, to read the thread, or to fix it yourself.

5. Most of these projects will be gone. Between November 2025 and August 2026 we counted more than a dozen near-identical local dictation apps posted to Hacker News: Ito, Vocalinux, Toice, Purr, Fonos, TypeWhisper, Recitey, TongueType, Plainsay, Velora, Dictata and others. Most collected fewer than five points and zero comments. This is not unique to small projects either. Mozilla's DeepSpeech, with 26,772 stars, is archived. Coqui STT, the project that carried it forward, has had no commits since March 2024, and the company behind it wound down that year. Star counts measure enthusiasm on the day of launch. Release dates measure whether anyone is still there.

Where it wins, and this part is not small

Your audio does not go anywhere. Not as a policy on a security page, but because there is no network call. That is the one privacy claim that does not depend on trusting anybody, and it is why doctors, lawyers and people working under NDA end up here.

No caps, no account, no renewal. Paid dictation runs $8 to $15 a month, and the free tiers stop at a couple of thousand words a week, which is about fifteen minutes of talking. Local tools have no word counter because there is no server counting.

You choose the model. Whisper Medium may suit your voice better than Large. Parakeet may be five times faster on your laptop. The Handy author found Medium better than Turbo for his own speech, which is exactly the sort of thing you can only discover if you are allowed to swap models.

It keeps working. No vendor can discontinue your setup, change the pricing, or decide your country is no longer supported. The Dragon owners who watched the consumer line wind down know what that is worth.

You can read it. Not that you will. But somebody can, and in a category where an app watches everything you type, that matters more than usual.

If you just want dictation that works today

Worth saying without flinching: if you do not enjoy this kind of thing, open source is not currently the shortest path to good dictation. The paid tools are better at cleanup, they are signed and notarised, they have a support address, and they cost between two and fifteen dollars a month.

If that is you, skip the whole category and start with a platform guide instead: best dictation apps for Mac, for Windows, or for Linux. Each one lists the free built-in option first, because on two of those three platforms it is genuinely enough for a lot of people.

What we left out and why

Amical (MIT, 1,508 stars), VoiceTypr (AGPL-3.0, 692), OpenSuperWhisper (MIT, 2,727) and about thirty others. All real, all currently maintained. They are not in the main table because they do the same thing as the tools above with fewer people behind them, and recommending a young project for something you will use every day is how you end up reinstalling in six months. If you want the full field, the awesome-voice-typing list on GitHub is the most complete one we found.

Whispering. Quoted several times on this page and still a good tool, but note that the standalone repository is archived: it now lives inside the Epicenter monorepo, which is where you should look if you go searching.

Model libraries: faster-whisper, WhisperX, sherpa-onnx, NeMo. These are components, not applications. If you are building something they are excellent. If you want to dictate, they are the wrong shelf.

Anything whose page says open source without a licence file. Paid closed-source products rank in these searches with pages built entirely around the phrase. The check takes a minute: find the repository, open the LICENSE file, read the first line.

How we checked this

On 27 August 2026 we queried the GitHub API for every project on this page and recorded four things: the contents of the licence file, the star count, the date of the last tagged release, and the date of the last commit. Prices and system requirements came from the vendors' own pages on the same day. Licence claims are quoted from the licence file or the EULA, not from the marketing copy, which is how the FUTO and Talon entries ended up looking different here than they do elsewhere.

We did not run an accuracy bake-off, and there are no word error rates of our own on this page. Where speed or accuracy is described, it comes from a published study or from a named user on a dated public thread, and it is labelled as such. Every complaint we quote is attributed to a person and a date so you can judge how stale it is, and several of them are from 2025 and may already be fixed.