Word replacements
Word replacements correct recurring errors, brand spellings, and fillers after recognition in the voice-typing and voice-translation text pipeline.
- Settings → Text processing → Word replacements
- The menu bar’s word-replacements entry
voicehotkey://word-replacements(see Deeplinks)
Edit rules
Section titled “Edit rules”| Field | Meaning |
|---|---|
| Original | Text to find; literal rules accept variants separated by commas, Chinese commas, or enumeration commas |
| Replacement | Result; leave empty to delete matches |
| Use Regular Expression | Supports capture groups such as $1; add any required word boundaries yourself |
| Timing | Before / After / Both, relative to polish or translation. Without AI, manual rules still run once |
Matching is case-insensitive. Literal rules handle word boundaries for ASCII words. New manual rules are global. Filter the list by All / Manual / 🤖 Auto-Learned; click to edit, or right-click to delete or convert a learned rule to manual. Rules with an app label can also be made global.
An existing rule’s app label is checked for whole-utterance shortcuts only; ordinary substitutions may still apply in other apps. Do not rely on that label for strict app isolation.
Ordinary replacement & whole-utterance shortcuts
Section titled “Ordinary replacement & whole-utterance shortcuts”An ordinary replacement can match part of a sentence and does not itself bypass the LLM. Ordinary voice typing also has a whole-utterance shortcut: the entire recognized utterance must equal the rule’s original after normalization, and the rule must allow fast output. It then uses the replacement directly and skips polish.
New manual rules allow this shortcut by default. Comma variants, regex, and partial matches do not trigger it through their usual matching syntax. The shortcut does not run in direct voice typing or voice translation. The overlay’s ⚡️0ms label means the LLM step was skipped, rather than measuring total recognition and insertion time.
Learned rules carry 🤖 Auto-Learned and a shortcut hit count. They participate in whole-utterance shortcuts, rather than ordinary word substitutions. See LLM polish.
Examples
Section titled “Examples”| Original | Replacement |
|---|---|
raycast, reycast |
Raycast |
air pods |
AirPods |
\b(um|uh)\b (regex) |
Leave empty |
Combine with recognition dictionaries
Section titled “Combine with recognition dictionaries”Engine context, prompts, or dictionaries help recognition hear the word correctly; word replacements correct output text. You can use both. These rules belong to the voice-typing pipeline; do not assume command resolution or file subtitles apply them too.