Settings
AI Refine
A small local model fixes punctuation, filler and phrasing. It runs on your own machine and never calls out.
Raw speech-to-text is accurate but messy: no punctuation, odd capital letters, every "um" intact. AI Refine is the pass that tidies it.
It runs a small language model locally through Ollama. It talks to 127.0.0.1 only — there is no cloud service behind it and no request ever leaves your computer.
Open Settings → AI Refine.
Turning it on and off
Enable refine pass switches the whole thing on or off. With it off, Voxy still dictates — you get the raw transcription, plus the built-in filler removal and line-break commands, which run in plain code.
Ollama
Refine needs Ollama installed and running. Voxy shows its state here and can install it for you if it is missing.
The two models
| Model | Memory | Character |
|---|---|---|
| Standard · sharper | ~4.9 GB | The default and the better all-rounder. Handles long dictations reliably, and is actually the quicker of the two on Apple silicon |
| Light · fast | Smaller | Worth choosing if your machine is short on memory. On longer dictations it occasionally drops a whole sentence |
Note Standard being both sharper and faster looks like a trick answer, but it is what measuring on Apple silicon showed. Light's only real advantage is its size.
Prose model and code model
They are set separately:
- Prose model — everyday writing.
- Code model — editors and terminals.
The code model defaults to Standard, because it handles identifiers far more reliably. "use auth hook" becoming useAuthHook is exactly the kind of thing the lighter model gets wrong.
See Code mode.
Keep Voxy ready
Voxy's AI has to be running before it can tidy your words. Left idle, your system shuts it down, and the next dictation waits about 20 seconds for it to start again.
Keep Voxy ready avoids that wait.
| Setting | Behaviour |
|---|---|
| Automatic | The default. Holds the model when your machine has room to spare |
| Always | Holds it whatever the machine |
| Never | Never holds it. Accept a slow first refine after a break |
Automatic is the right answer for nearly everyone. Voxy watches memory pressure and stands aside if your machine comes under real strain — a promise made at launch does not survive you opening a virtual machine an hour later.
Debugging
Show raw vs refined puts each dictation's before-and-after in the menu bar (the tray menu on Windows). Useful when you are tuning your Vocabulary or Styles and want to see what the refine pass actually changed.
What refine will never do
It is a text filter, not an assistant. It never answers a question you dictate, never completes your thought, and never adds content you did not say.