Speech models

A speech model is the file that turns your voice into text. It runs on your computer. The Models page lists every model with its size, languages, speed and accuracy.

The models

  • Whisper: large-v3 turbo, large-v3, Distil large-v3.5 (English only), medium, small, base and tiny. Whisper covers 99 languages.
  • NVIDIA Parakeet: English only, or one that covers 25 European languages.
  • Moonshine: English only, built for slower processors.

Parakeet and Moonshine don’t use your vocabulary list. A smaller model suits a slower computer. See Languages and translation for which models can translate.

The recommendation

HushWise checks your processor, memory and graphics card and suggests the best match. The speed test on the Models page runs on a recording that comes with the app. You can use one model for dictation, one for live preview and one for commands.

Size and download

The one-time download is typically 260 MB to 1.6 GB, depending on your machine. The download can be paused and resumed. Every file is checked against a SHA-256 fingerprint before it is used. After that, HushWise works offline.

The GPU pack (Windows)

With an NVIDIA graphics card, an optional 1.3 GB GPU pack of NVIDIA’s libraries speeds up the big models. Download it from the Models page. Quit HushWise from the tray and open it again afterwards. Unpacked, it needs about 2 GB in your models folder. On a Mac, the graphics cores are used automatically.

Where models are kept

Models can live on any drive. Change the folder on the Models page, or Settings → Storage. Moving them copies the files, checks them, then removes the old copy. If you cancel or something goes wrong, the old folder stays as it was. Pause or finish downloads before moving.

If the drive with your models is missing, such as an external drive, plug it back in or choose another folder.

Download mirror

Models come from Hugging Face by default. If that is blocked or slow where you are, enter a mirror in Settings → Storage. It must be an https address.

Memory

Settings → Advanced lets you keep the model always loaded, unload it when idle, or load it only while you dictate. The least memory is used when it loads only while you dictate, and then each dictation waits for the model to load.