Open-Source Dictation Software: Free Apps on GitHub

Free open-source dictation apps on GitHub are real, usable and easy to check. Four active projects are worth comparing: Handy and OpenWhispr use the MIT License and run on Windows, macOS and Linux ^{1 2 3 4}. VoiceInk is a native Mac app whose source code is under the GNU General Public License v3.0 ^{5 6}, and its README says a purchase supports development and gives access to automatic updates ⁷. HEX is an MIT-licensed app for Apple silicon Macs ^{8 9}. Each app can run an open speech model on your computer. The trade-offs are more setup, fewer integrations and reliance on volunteer maintenance. This guide is part of our overview of dictation software, which compares every kind of app.

Open-source dictation apps run open speech models on your own computer

A dictation app records your voice, passes the audio to a speech recognition model and types the text where your cursor is. An open-source app publishes the code for that whole path. The main speech models these apps use are OpenAI's Whisper and NVIDIA's Parakeet. Parakeet's model card says use of the model is governed by CC-BY-4.0 ¹⁰, so the model's license is separate from the license of each app.

Each step in that path is a place where privacy can be lost, and each one can be checked. The microphone step is controlled by your operating system's permissions. The model step is where the audio is turned into words, and it is the step an open-source app makes visible in its code. The insertion step is where the text lands, and some apps add an optional rewrite there before the text is typed. An app that keeps all three steps on your computer is a local app. An app that sends the audio or the text elsewhere is a cloud app, whatever its marketing says.

Running the model on your own computer is the main difference from a cloud dictation service. HEX's README says that when you turn on post-processing with OpenCode in Modes, the app sends the transcript, the foreground application and the Brave website hostname to the model provider you configure ¹¹. Check each app's settings and privacy notes before you assume your audio stays local.

Open source here means the code is public and reusable under its license. It does not mean the app is free of cost in every case. VoiceInk's source is open, but its README says a purchase supports development and gives access to automatic updates ⁷. For a general comparison of paid and free options, see best dictation software and free dictation software.

The active projects compared by platform, model and license

The table compares the four projects on the facts their own repositories state. Each row cites the README or license file, and the last column gives the newest release tag that the GitHub releases page marked Latest on 16 October 2026 ^{12 13}.

Handy and OpenWhispr are the two that cover all three desktop systems. VoiceInk and HEX are Mac apps. VoiceInk's README calls it a native macOS application ⁵, and HEX requires Apple silicon ⁹. If you use a Windows PC, that leaves Handy and OpenWhispr. For Windows-specific options, see best dictation software for Windows.

The license column needs a careful read. VoiceInk's repository lists its license as GPL v3.0 in the README, but its LICENSE file is only a short header that links to the full text ¹⁹. GitHub's automatic license detector may therefore fail to recognize it. The project is not unlicensed. It is a copyleft license, which means people who distribute modified copies must share their source under the same terms.

Two names cause confusion. The original Swift app called Hex was archived by its owner on 14 October 2026 and is now read-only ²⁰. The current project is HEX at anomalyco/hex. If you searched for "Hex" and found an older install, check which one you have. The questions what is the Hex dictation app and what is the Handy dictation app cover each project in more detail.

What this comparison does not measure

The table does not rank the apps by transcription accuracy. Accuracy depends on the model, the language, the microphone and the room, so test each app on your kind of speech. A model that is larger in parameters, such as Parakeet v3 at 600 million ²¹ or Whisper large at 1,550 million ²², is not guaranteed to be more accurate in your app, because the app also decides how the audio is captured and processed. Treat the size figures as a guide to memory and compute, not a verdict on quality.

The comparison also leaves out popularity. Star counts and download totals on GitHub measure attention, not whether a program is correct or safe. They can help you find projects, but they should never decide which one you trust with your voice. Each project's own README and license are the source for the claims in this guide, and the checks in the last section are the way to confirm them.

Open speech models made local dictation practical

The app is only as capable as the model inside it. The sizes differ a lot. These apps use models that you download and run on your own computer. Whisper's smallest model, tiny, has 39 million parameters and needs about 1 GB of memory ²³. The small model has 244 million parameters and needs about 2 GB ²⁴. The large model has 1,550 million parameters and needs about 10 GB ²². Parakeet v3 has 600 million parameters ²¹, which puts it between Whisper's small and large models on size.

Parameter count is not the same as accuracy or speed. A larger model usually needs more memory and more time, but how fast a given app transcribes depends on the hardware and on how the developer packages the model. This is the reason to test an app on your own machine before you commit to it. The speech recognition guides explain how these models work and how they are measured, starting with open-source speech recognition and how to run Whisper locally.

Parakeet v3 covers a defined set of languages. NVIDIA's model card says it extends language support from English to 25 European languages ²⁵. If you dictate in a language outside that list, check which model the app offers for it.

The two sources also disagree about hardware. Handy's README describes Parakeet V3 as a CPU-optimized model ¹⁵. NVIDIA's model card says the model is designed and/or optimized to run on NVIDIA GPU-accelerated systems ²⁶. Both statements can be true of different builds of the model, so time the app on your own computer rather than trusting either claim. For background on the terms, see automatic speech recognition and on-device speech recognition.

Open code lets you verify privacy claims instead of trusting them

Each project makes privacy claims in its README, and those claims are the place to start. VoiceInk says its "100% offline processing ensures your data never leaves your device" ²⁷. OpenWhispr says it has "no data collection, no telemetry" and is fully open source ²⁸. HEX's post-processing, which you turn on in Modes, sends data to a model provider ¹¹.

A claim is only as good as your ability to check it. The Open Source Definition requires that "the source code must be the preferred form in which a programmer would modify the program" ²⁹. That is what makes the code a real check. You can search the source for network calls, read how the app handles audio, and see what a model download contacts. You can also watch the app's network use while you dictate with a network monitor. A clean result proves little about future versions, so repeat the check after each update.

A simple test fits in one sitting:

1. Install the app and list every speech model it offers, noting which ones run locally.
2. Start a network monitor, then dictate a paragraph with a local model and note any outbound connections.
3. Turn on each optional cloud or post-processing feature in turn, and repeat the same dictation. Note which ones send data and what they send.
4. Compare what you saw with the README's privacy section. Any claim you could not reproduce is a question to raise with the project before you rely on it.

Open code does not settle every privacy question. The app can be private while the model host, the operating system or a cloud provider you chose is not. The voice privacy guide on whether voice to text is safe covers the wider questions about what happens to voice data. For a single app's claims, the question whether VoiceInk is free and open source shows how to read the license and the paid tier together.

Expect more setup and fewer integrations than paid apps

Open-source dictation asks more of you. You may need to download a model, grant microphone and accessibility permissions, pick a shortcut and check the app's language settings. Paid apps often package those steps into one installer and may add AI cleanup, sync, team features and support. Open-source projects vary in how complete their setup is, and many depend on a small group of volunteer maintainers.

The feature sets also differ by design. OpenWhispr describes itself as dictation "with AI agents, meeting transcription, and notes" ⁴, which is broader than a pure dictation tool. If you need dictation into every app with a company behind it, a paid app or the built-in free dictation may fit better. The AI dictation apps guide compares the paid tools that add AI rewriting, and dictation without AI covers apps that only transcribe.

VoiceInk shows the trade-off in its license. Its README says a purchase helps support continued development and gives access to automatic updates, priority support and upcoming features ⁷. Building from source costs time rather than money, and the paid route is what keeps updates arriving without a build step.

Choose by what you will tolerate:

• Open source fits if you can handle a setup step, a few rough edges and occasional bugs, or if you want to read the code and build on it.
• A paid or built-in tool fits if you need dictation to work out of the box in every app, want a company to answer support questions, or need features such as team management.

Switching between these apps costs little in money, since each one is a download, but learning one app's shortcuts and settings takes your own time. Pick the one whose trade-offs you can live with, and test it on the passages you actually write.

Judge a project by its releases, license and archive status before you install it

Check four things before you install an open-source dictation app. Each check is quick and tells you more than a star count or a review.

1. Read the license file, not just the badge. A LICENSE file that is a short header with a link to the full text, as VoiceInk's is ¹⁹, means you need to open the linked license to know the terms.
2. Find the newest release and its tag. On 16 October 2026 the Handy releases page marked v0.9.8 as the Latest release ¹². A project with no recent tag may still be maintained, so also read its issues and commit history on GitHub.
3. Check whether the repository is archived. The original Hex repository displays a notice that its owner archived it on 14 October 2026 and made it read-only ²⁰. An archived repository will not receive fixes.
4. Download from the project's own releases or the package manager its README names. Avoid copies on third-party download sites, which may not match the published code.

Once you have the file, a Mac app can be checked further. Mac apps distributed outside the App Store are often signed by their developer. The codesign command in Terminal can show both. codesign -dv --verbose=4 followed by the app's path reports who signed it, and codesign --verify --deep --strict followed by the same path checks that the contents have not changed since signing. A valid signature shows who built the file and that its contents are unchanged. It does not show that the app is free of privacy problems, so the network check described earlier still matters.

If you want an on-device dictation app without building the setup yourself, Yap, which we make, runs on-device dictation for Mac and Windows. Its prices are on the pricing page. For a free and open route, the four projects above are the ones to test first. Dictate the same passage in each, time it, and keep the one whose output needs the least correction.