Transcription / Speech models

Speech models

Install the model Local mode listens with, and check that it keeps up on this computer.

The models

Set Mode to Local on Settings → Transcription to see them. Recommended marks the model this computer should use; In use marks the one Local mode runs.

  • NVIDIA Nemotron + Parakeet (fast, accurate): the default. About 1.2 GB. English only.
  • Small (accurate): about 465 MB. Every language on the list.
  • Base (fast): about 141 MB. For computers where Small cannot keep up.

Install a model

  1. Open Settings → Transcription and set Mode to Local.
  2. Press Install beside the model.
  3. When it finishes, the model is selected. To switch between installed models, press Use.

Install before the service. One model installs at a time, and NVIDIA Nemotron + Parakeet is over a gigabyte.

Remove a model

Press Remove beside an installed model, then Remove again to confirm.

The model Local mode is set to, or is running, has no Remove. Choose another model first.

The speech engine

Base and Small need the speech engine, the row above the models. On Windows and Linux, installing Base or Small gets it too; Get engine gets it on its own.

  • Ready · fast: installed and working.
  • Ready · graphics card: running on the graphics card.
  • Ready · slow: an older engine, about four times slower. Update engine replaces it.
  • Needed: no engine yet. Until there is one, Local uses Cloud.
  • For Base and Small: you are using NVIDIA Nemotron + Parakeet, which does not need it.

Locate engine

Locate engine picks a speech engine program already on the computer. Once one is set, the button reads Change engine, and Clear forgets it. On a Mac, install the engine with Homebrew, then press Locate engine.

brew install whisper.cpp

Graphics cards

Only Base and Small can use a graphics card. On Windows with an NVIDIA graphics card and driver 527.41 or newer, the button reads Get engine · graphics card. AMD and Intel graphics are not supported.

Check speed

Check speed times an installed Base or Small model on 10 seconds of audio, and shows the verdict beside it. Press it again after changing the engine or the computer. It is not offered for NVIDIA Nemotron + Parakeet.

  • Runs comfortably here: twice real time or faster.
  • Keeps up here: at least 1.2 times real time.
  • Only just keeps up here: likely to fall behind when the computer is busy.
  • Too slow here: the model will lose words. Use a smaller model.