How to dictate text into any Windows app

Why people give up on dictation

Almost everyone tries dictation at some point, and almost everyone stops within a week. The reason is rarely accuracy in the abstract. It is the friction around the accuracy.

You open a separate window to speak into, then copy the result across. You lose the paragraph you just dictated because the app moved focus. A client name comes back spelled three different ways in the same document. You write in two languages during a normal working day and the tool only knows one of them. Each of these is small on its own. Together they cost more time than typing.

Typing at the cursor beats a dictation window

This is the single design decision that decides whether you keep using a dictation tool.

A tool with its own window makes you leave the thing you were writing, speak, then carry the text back. A tool that types at the cursor does not interrupt anything. You are already in the reply box, you hold a hotkey, you say the sentence, and the sentence appears in the reply box. Nothing was opened, nothing was pasted.

The practical consequence is that it works everywhere without integrations. Because the text arrives as ordinary keyboard input, the receiving program does not have to know anything about it. Outlook, Word, Excel, a browser form, a chat window, a code editor, some internal admin panel from 2011: they all accept it, because they all accept typing.

The name problem

Speech recognition is good at ordinary sentences and bad at the words that matter most in your work: client names, product names, the abbreviations your industry uses, your colleagues’ surnames. These are exactly the words you cannot leave wrong, so you end up proofreading every dictated line, and proofreading every line is why dictation stops feeling faster.

The fix is a word list of your own. You write down how a name should be spelled and which mishearings to replace, and the tool applies it every time. In BeltoVox that list also feeds a second pass that catches new mishearings which merely sound like something on the list, so you do not have to enumerate every possible way a name can be mangled. The list stays on your machine.

Two languages, one keystroke apart

If you work in more than one language, the switch has to be instant or you will not use it. Digging into a settings dialog to change language, then digging back twenty minutes later, is not a workflow anyone keeps.

BeltoVox supports 51 languages and lets you nominate two of them as the pair you actually work in. Ctrl+L swaps between them, and you can put the swap on a mouse button instead. Answering a Hungarian email and then an English one costs a keystroke.

Where your voice goes, and what it costs

Dictation means sending a recording of your voice somewhere. If that somewhere is a vendor’s own server, you have to trust the vendor’s policy, and in a lot of professional work you cannot make that promise on a client’s behalf.

BeltoVox uses your own API key, from OpenAI or from Groq. The recording goes from your machine to your account with that provider. It does not pass through a Beltorion server, and there is no Beltorion account holding your transcripts.

That choice also changes the price shape. Because we are not reselling someone else’s transcription minutes, the licence is a one-time 59 USD rather than a monthly fee, and you pay the recognition provider directly for what you actually use. The app has a usage tab that shows a running estimate of that spend, so it is not a mystery line on a card statement.

Getting started

Three things, in this order. Get an API key from OpenAI or Groq and paste it into the app. Pick your two working languages. Add the five or six names you write every week to the word list, before you need them.

Then use it for one full day on real work rather than testing it on a sample sentence. Dictation is a habit, and habits are decided in the first day.

The full setup and everything the app can do is in the