Dictation that sends nothing to the cloud

There is a commercial tool that does exactly this and does it well. You press a shortcut, speak, release, and the text appears where your cursor was. The only thing that bothered us about it was that the voice goes to somebody else's server — including the moments when we are dictating the contents of an internal document.

So we built our own version. A small tray application captures a global hotkey, records for as long as the key is held, and sends the recording for transcription. The difference is where it sends it: to our own server running a speech model. The recording never leaves the network and is deleted after transcription.

The text is inserted via the clipboard, which is inelegant and simultaneously the only approach that works reliably across applications. The previous clipboard contents are restored afterwards.

A practical finding after several months of use: the quality of Slovak transcription on our own hardware is good enough for email, notes and task descriptions, and not good enough for anything where exact terminology matters without an editing pass. Diacritics and punctuation are inferred well; technical abbreviations almost never.

The same speech model then also serves a voice assistant on a phone, which is a side effect we did not anticipate. Once speech transcription is a service you own, you start using it in places where you would not have paid for a cloud service.