System-wide dictation pointed at your own server

Building a dictation server is half the job. The other half is being able to talk to any window on your computer and have the words appear, the way the paid dictation apps do, but pointed at your own machine instead of a company's. This is how I set that up, why it is private, and the one thing that is likely to trip you. New terms are explained on the back-to-basics page.

Why bother

The paid dictation tools are genuinely good now, and they all share one property: your voice goes to their computers and stays there, under their rules. For a work machine that is often somewhere between uncomfortable and not allowed.

I already had a Whisper dictation server running at home. What was missing was the desktop side: something that sits on a keyboard shortcut, records while you hold it, sends the audio to a server of my choosing, and types the result into whatever window I am in.

The tool

OpenWhispr is a free, open app that does exactly that, and the important part is that it lets you set your own server address instead of a company's. That one setting is what makes the whole thing private.

The setup that works

OpenWhispr has two separate address settings, and understanding the split is most of the battle:

• Transcription, which is the speech-to-text part. Point this at your Whisper server, using the address and the key you made when you set it up.
• Reasoning, which is optional. If you want the app to tidy your dictation afterwards, that is a second, separate program (a language model), and it has its own address. You can leave this off entirely and just get straight transcription.

A nice touch: the app checks your address by asking the server what it can do, before you have said a word. If the list fills in, your address and key are right. If it stays empty, they are not, and you have found the problem early.

"Will it even allow a home address?"

This was the first thing I checked, because web-based apps increasingly refuse to use anything that is not a secure https:// address, and a home server usually is not.

The answer here is sensible. OpenWhispr allows plain home-network addresses (the ones that start 192.168, 10. and so on) over ordinary connections, while it insists on the secure kind only for addresses out on the public internet. For a home setup that means no certificates, no extra plumbing. Type your home address, done.

The one snag: it may not connect the first time

If you turn on the optional reasoning feature, the app's check for what the server can do may quietly fail, even though the same server answers fine when you test it another way. This is a permissions thing between the app and the server, and it has a one-line fix on the server side that tells it to accept requests from the app. If your transcription works but the optional tidy-up does not, that is almost always this, and the forum has the exact fix for the tool you are running.

What you end up with

A keyboard shortcut on your laptop. Hold it, speak, let go, and the words appear in whatever window you were in, having travelled to a machine you own and no further. With a graphics card behind the server it is effectively instant; even without one it is fine for note-length dictation.

Two honest limits. First, the clever features in paid tools, like telling speakers apart in a meeting, are their own hard problems; this setup is for dictation, not for taking minutes. Second, OpenWhispr has its own optional account features for syncing notes, and those do talk to its makers; the dictation itself is the part you have made private, so simply skip signing in if that matters to you.