Is AuraScribe really free?

Yes. AuraScribe is free forever and open source under the MIT license. There is no paid tier, no subscription, and no account to create.

Does my voice get sent to the cloud?

No. Every word is transcribed on your own PC. The only network request the app ever makes is downloading a speech model once, after which it works fully offline.

How is it different from Wispr Flow, Superwhisper, or Dragon?

AuraScribe is free, open source, and 100% offline, with no account required. Wispr Flow and Superwhisper are paid and cloud-based; Dragon is expensive. It is the private, local alternative.

Which languages does it support?

English plus 25 European languages, roughly 40 Asian languages, and Malayalam and Kannada, using four on-device engines. Each auto-detects the language within its region.

What are the system requirements?

Windows 10 or 11 (64-bit) and a microphone. The installer is about 8.6 MB and runs faster than real time on an ordinary CPU. No internet is needed after the first model download.

Does it work in every application?

Yes. Press the hotkey in any window and the finished text is typed straight into whatever app has your cursor: browsers, editors, terminals, chat boxes, and form fields.

On-device Malayalam (and Kannada) speech to text, with no cloud

Most dictation tools skip Malayalam and Kannada or send them to a server. How AuraScribe runs both fully offline using AI4Bharat IndicConformer, and what to expect.

Malayalam and Kannada are where most dictation tools quietly give up. They either do not support the scripts at all, or they only handle them by sending your audio to a cloud service. AuraScribe runs both on your own machine, with no internet connection needed after setup. Here is how, and what to actually expect.

Why these languages are usually missing

The fast English and European speech models simply were not trained on Malayalam or Kannada, so they cannot transcribe them at all. The tools that do support these languages tend to lean on large cloud models, which means your voice leaves your device. For anyone dictating personal messages, notes, or work in these languages, that is exactly the wrong trade.

Running IndicConformer locally

AuraScribe uses IndicConformer, an open speech model from AI4Bharat built specifically for Indian languages, and runs it entirely on-device. Getting a model like this to run fast and offline on a normal Windows CPU took real work, and Malayalam was verified against actual spoken dictation rather than assumed to work. In testing it transcribed full Malayalam paragraphs into clean, coherent text, and handled code-switched “Manglish” sensibly. Kannada uses the same engine and the same on-device path.

What to expect

It is fast, but not instant. On-device recognition for these languages runs at real-world speed on a normal CPU, quick enough for comfortable dictation.
It is fully offline. After the one-time model download, your speech never leaves your computer.
It is free. There is no subscription and no account, the same as every other language.

Giving the work back

Reaching this point produced a packaged Malayalam model that was contributed back to the open-source speech community, so the next person trying to build offline Malayalam dictation does not have to start from scratch. That is the kind of thing being open source is for.

If you have wanted to dictate in Malayalam or Kannada without handing your voice to a server, you can download AuraScribe for Windows, choose the Malayalam or Kannada model, and try it. It is honest work in progress, and real feedback from native speakers is the most useful thing you can send back.

All posts