Offline dictation on Windows: how on-device speech to text works
What "offline dictation" really means on Windows, how on-device speech recognition runs in real time on a normal CPU, and where the one and only network request happens.
“Offline dictation” gets used loosely, so it is worth being precise. In AuraScribe it means the entire path from your microphone to the finished text runs on your own computer. Your speech is captured locally, recognised locally, cleaned up locally, and typed into whatever app you are using. Nothing about what you said travels over the internet, because it does not need to.
Speech recognition small enough to run on your CPU
For a long time, accurate speech to text meant a large model on a powerful server, which is why so many tools stream your audio to the cloud. That has changed. A new generation of compact, open speech models is accurate and light enough to run in real time on an ordinary laptop processor, with no graphics card required. AuraScribe is built on these models, so the heavy lifting happens on the machine already in front of you.
To keep it responsive, the app transcribes your speech in short chunks while you are still talking, rather than waiting for you to finish. By the time you release the hotkey, most of the work is already done, so the text appears almost immediately.
The one and only network request
There is exactly one moment AuraScribe touches the network: the first time you choose a language, it downloads that speech model to your computer. That request pulls down model files. It does not upload your voice, your text, or anything about you. After the download, you can switch your Wi-Fi off and dictate all day. The app is designed so that this is the only outbound request it can make, which is a promise you can verify because the code is open.
Four engines, one simple experience
Different languages are best served by different open models, so AuraScribe uses four, each credited to the researchers who built it:
- Moonshine for fast, real-time English.
- NVIDIA Parakeet for twenty-five European languages, with automatic language detection.
- Dolphin for roughly forty Asian languages.
- AI4Bharat IndicConformer for Malayalam and Kannada.
You never have to think about which engine is running. You pick a language, the right model loads, and each one detects the specific language inside its own region automatically. You only download the ones you actually speak, which keeps the footprint small.
What you need
AuraScribe runs on Windows 10 and 11, sixty-four bit, and needs a microphone. The installer is about 8.6 MB, and the app sits at roughly 40 MB of memory when idle. There is no account to create and no subscription. After the first model download, no internet connection is needed at all.
If keeping your voice on your own machine matters to you, that is the whole point of running dictation offline. You can download AuraScribe for Windows and try it, or read more about what actually happens to your voice in cloud dictation.