VoiceInk Refine
Clean up dictation on-device using the built-in Refine enhancement model.
VoiceInk Refine is a text enhancement model, not a speech transcription model. It cleans up the transcript after your selected transcription model finishes.
Requirements
- An Apple Silicon Mac with at least 16 GB of RAM.
- A downloaded Refine model from AI Models → Local.
- A working transcription model configured separately.
Refine uses VoiceInk Refine V1 and needs no provider API key. Once downloaded, cleanup runs locally.
Set Up Refine
- Open AI Models → Local and download VoiceInk Refine.
- Edit a Mode and enable AI Enhancement.
- Select VoiceInk Refine as the provider.
- Choose Paste or Custom Command output and save.
The model card lets you manage or delete the download and reveal its files in Finder.
What It Changes
Refine has a fixed cleanup task: improve punctuation and capitalization, handle spoken formatting, remove filler and repetition, and resolve spoken self-corrections while preserving meaning and tone.
It does not use custom prompts, selected-text context, clipboard context, or screen context. Those controls are unavailable for Refine. It cannot provide Respond output or review Dictionary Auto Learn corrections.
For live dictation, the shared Skip short transcriptions setting can skip Refine on short input. See Model Settings.
For email drafting, selected-text rewriting, or Assistant questions, choose a provider that supports prompts and context, such as Ollama or a compatible cloud provider.