Settings
Speech Engine
The model that turns your voice into raw words. Three to choose from, all running on your own machine.
The speech engine is the first half of the pipeline: it turns audio into raw words. The second half is AI Refine, which tidies them.
Every engine runs fully on-device, on both platforms.
Open Settings → Speech Engine.
The three engines
| Engine | Size | Languages | Notes |
|---|---|---|---|
| Fast | ~0.6 GB | 25 | The default. Runs on the Neural Engine on Apple silicon, the GPU on Intel Macs, and the processor on Windows |
| Most accurate | ~1.5 GB | ~100 | Best with accents and background noise. Slower |
| Built into macOS | — | Follows macOS | macOS only. Uses the speech model macOS downloads itself. Needs macOS 26 or later |
Which should you pick?
Start with Fast. It is quick enough that transcription is not what you are waiting for — on Apple silicon, and on an ordinary Windows processor with no graphics card at all — and 25 languages covers most people.
Switch to Most accurate if you have a strong accent, dictate in a noisy room, or work in a language outside the 25. It is noticeably better in those conditions, at the cost of speed and about a gigabyte more disk.
Built into macOS is worth trying on macOS 26 or later. It uses the model Apple already ships, so there is nothing extra to download. It has no Windows counterpart — Windows offers the first two.
Downloading a model
A model downloads the first time you select it. This is the only moment Voxy needs the internet in order to dictate.
Once it is on disk, dictation is completely offline.
Switching engines
- Open Settings → Speech Engine.
- Pick an engine.
- Wait for the download if it is the first time.
Your Vocabulary, Phrases and Styles are unaffected — they apply whichever engine is running.