Use OpenWhisper
AI cleanup
Polish dictation with a cloud model, a local server, or your own coding agent, and teach it your words.
On this page
AI cleanup runs after transcription. It can fix punctuation, apply your spelling and style rules, or turn dictated speech into a reusable format such as an email or support ticket. It's off until you turn it on.
Connect a text model#
Open Settings → Dictation → AI cleanup. Under Who runs AI cleanup, pick one of your installed coding agents or the built-in API / local model; the Cleanup model card below holds the endpoint, model, and Thinking level. That gives you three kinds of choice:
| Kind | Options | Setup |
|---|---|---|
| Cloud provider | OpenRouter, OpenAI, Groq, OpenCode Go/Zen | Add the key in Settings → App → API keys. Keys live in your operating system's credential store. |
| Server on this computer | Ollama, LM Studio, vLLM, llama.cpp, LocalAI | Preset to each server's standard local address; no key needed. Just run the server. Edit changes the address or sets an optional key variable. |
| Your coding agent | Claude Code, Codex, OpenCode | Uses the agent you've already installed and signed in to, with its own default model or one you pick. |
Custom… adds any other OpenAI-compatible endpoint; choose the API it speaks, Chat Completions or Responses. The picker marks local servers that are running, reads each model's context window and thinking support from the server itself, and lists only models that generate text.
Thinking level lists only the levels the chosen model accepts, from Off or Minimal up to Max (Codex models go up to Ultra). A saved level the model lacks runs at the nearest one it takes, and Settings tells you so.
A new text-model choice starts on OpenRouter's free models. The first time cleanup is on and you haven't picked what runs it, OpenWhisper looks for Claude Code and Codex. If one is installed and signed in, it asks once whether cleanup should run through it, with the account you already use there and no API key. Nothing switches without your answer, and a model you already chose is never touched.
Choose a cleanup level#
Cleanup level sets how much the model may change:
- Light: punctuation, capitals, filler words, and obvious recognition errors.
- Medium: also spoken corrections ("no, make that Tuesday"), lists, and paragraph breaks.
- High: also removes false starts and tightens wording while keeping the meaning.
A saved custom prompt shows as Custom and replaces the level's instructions; pick a level to return to a preset.
Enable and check cleanup#
Turn cleanup on, make a short recording, and compare the cleaned result with what you said before using it in a message or document. These are instructions to a language model, not a guarantee of perfect edits.
Teach spelling and style#
The Settings → Personalize section holds six pages:
- Dictionary: add names and jargon, star important words, and enter Sounds like replacements. It steers recognition where the engine supports hints, learns from your corrections, and its spelling replacements work even without AI cleanup.
- Snippets: save text behind a spoken trigger. A trigger on its own skips cleanup; one inside a sentence is protected during cleanup. Snippets apply to live dictation.
- Styles: a Formal, Casual, or Very casual tone for each kind of app. Styles need AI cleanup. The optional, off-by-default Read text near the cursor on Windows sends nearby text to your cleanup model for context.
- Learned rules: other spelling and style preferences, in your own words.
- Profiles and Commands: covered below.
Keep instructions concrete: preferred spellings, punctuation, or a specific structure. Review names, numbers, and commitments in the result. The personalization guide has the details.
Create a cleanup profile#
A profile turns speech into a fixed format. Support ticket and Email are editable starters.
- In Quick Record, check AI cleanup to show Cleanup profile, then choose Manage…, or open Settings → Personalize → Profiles.
- Pick a starter or New profile. Name it and describe the structure, tone, and details you want.
- Optionally include learned rules and give it a dedicated shortcut.
- Save profile, choose it in Quick Record, and record.
Profile instructions replace the standard cleanup prompt, and the profile's format wins over conflicting formatting rules. Profiles use your cleanup model and always run cleanup, even when it's off for standard dictation. Profile shortcuts always toggle: press once to start and again to stop. Editing a profile doesn't affect a recording already in progress.
Rewrite selected text#
Set a Command Mode shortcut in Settings → Personalize → Commands, select text in any app, and speak an instruction such as “make this more concise.” With nothing selected, it can write new text at the cursor if that option is on.
Transforms are saved rewrites with optional shortcuts; they also work from the Scratchpad and can prepare words for actions. Both use your cleanup model even when standard cleanup is off, and send the selected text with your instruction. If the app can't safely paste back into the original spot, it copies the result for you to paste.
Recover from a cleanup failure#
If cleanup fails, OpenWhisper keeps the raw transcript and the status says why. Check the provider, address, model, key, and network connection, or for an agent, that it's installed and signed in.
History records the profile used and keeps the raw transcript whenever cleanup changed it. Meeting AI insights have separate model settings; see Meeting mode.
Get help from the community, or tell us what went wrong.