About
Your voice, your machine.
AuraScribe is a free, open-source dictation app for Windows. You press a hotkey, speak, and clean text appears wherever your cursor is. Every word is transcribed on your own computer, and it stays there.
Why it exists
Good dictation had quietly become something you rent. The best tools are subscriptions that stream your microphone to a server you do not control, the classic desktop option costs hundreds, and the one built into Windows sends what you say to Microsoft. For a lot of people that is a fine trade. For anyone who dictates private notes, medical or legal work, source code, or simply does not want their voice sitting on someone else’s computer, it is not.
AuraScribe started from a stubborn question: what if none of it ever left your machine, and it still cost nothing? Speech recognition is now good enough and light enough to run in real time on an ordinary laptop CPU. The only thing missing was someone willing to build the whole thing around privacy instead of bolting it on afterwards.
How it works, honestly
The app has exactly one job: hotkey, speak, text at your cursor, in any application. The recognition runs entirely on-device. The only network request AuraScribe ever makes is a one-time download of the speech model you choose; after that it works with your Wi-Fi off. There is no account, no telemetry, no analytics, and no cloud fallback that quietly kicks in. The permitted-network rule is enforced in the app’s own configuration so that a regression would be obvious, not silent.
Under the hood, four on-device engines cover different parts of the world, each one open and credited to the researchers who built it:
- Moonshine handles english (Useful Sensors).
- NVIDIA Parakeet handles 25 european languages (NVIDIA).
- Dolphin handles ~40 asian languages (DataoceanAI).
- IndicConformer handles malayalam & kannada (AI4Bharat).
Between them that is English, twenty-five European languages, roughly forty Asian languages, and Malayalam and Kannada, all running locally. Malayalam and Kannada were the hardest to reach and were verified against real dictation, not just assumed to work.
What actually makes it good
The speech models are not a secret sauce. They are open, and every serious local-dictation tool draws from the same public research; nobody in this space trained their own, and doing so would cost millions and break the promise of staying free and lightweight. So AuraScribe does not pretend the model is the product. The product is everything around it:
- Latency you do not feel. Speech is transcribed in chunks while you talk, so the text is nearly ready by the time you stop.
- A local cleanup pass. Punctuation, capitalisation, and filler words are tidied on-device, so what lands reads like writing, not a raw transcript.
- Your words. A personal dictionary and text snippets teach it the names, terms, and boilerplate you actually use.
- It types into anything. Browsers, editors, terminals, chat boxes, form fields. If your cursor is there, the text goes there.
Free, open, and giving back
AuraScribe is released under the MIT License and free forever, with no paid tier and no account to create. The full source lives on GitHub, so anyone can read exactly what it does with your microphone, which is the only real way to trust a privacy claim. The work of getting Malayalam running well produced a packaged model that was contributed back to the open-source speech community, so the next person building an offline tool does not have to start from nothing.
What it is not, yet
AuraScribe runs on Windows 10 and 11 today. macOS and Linux are planned and honestly not built yet, so rather than fake support the app returns clear errors on those platforms until the real thing ships. It is a focused dictation tool, not a meeting transcriber, not a chat assistant, and it will never grow a cloud option or a subscription. If those limits are the price of software that keeps your voice on your own machine, they are the right ones.
Jeswin Thomas Jestin, the maker
AuraScribe is designed, built, and maintained by Jeswin Thomas Jestin. The work spans the whole stack: the local-first architecture, the four on-device speech engines (Moonshine, NVIDIA Parakeet, Dolphin, and AI4Bharat’s IndicConformer), the on-device cleanup pipeline that turns raw speech into clean text, the personal dictionary and snippets, and the Windows text injection that types into any application.
Reaching genuinely good offline Malayalam was the hardest part, and it produced something worth sharing: a packaged Malayalam speech model contributed back to the open-source speech community, so the next person building offline Indic dictation does not have to start from nothing. AuraScribe is built in the open, shipped as real, versioned releases rather than demos, and improved through the people who use it and report what breaks.
How to help
If you want to support the project, the most useful things are simple: use it daily, file an issue when something is wrong, star the repository so others can find it, or sponsor the work so it can keep being free for everyone else.