Clean Up Transcripts With AI

This guide is for people who want TypingVoi to polish rough transcripts after speech-to-text. It explains when Cleanup AI helps, how to choose between a local and remote LLM, and how to set the profile so cleanup improves readability without changing what you meant.

1. Know when Cleanup AI is available

Cleanup AI is a paid feature. You can use it during the 14-day trial or with Pro. If you are on the free tier without an active trial, TypingVoi shows a locked Cleanup card instead of the full settings.

That means:

• you can still record and transcribe without cleanup
• you need the trial or Pro to turn on Run cleanup
• you also need the trial or Pro to save cleanup model settings in a profile

If you are unsure which plan features are unlocked, check Understand Free Trial and Pro Features.

2. Open the Cleanup AI settings for a profile

Cleanup is configured per profile, not once for the whole app. Open Settings, select the profile you want to tune, then open the Cleanup AI card.

Screenshot placeholder: Profile settings page with the Cleanup AI card expanded, showing the Run cleanup toggle, Keep original on failure toggle, LLM Model selector, temperature slider, and Prompt Preset menu

This card controls whether TypingVoi sends the transcript through an LLM after transcription and before final output or saved history.

3. Turn cleanup on only when the profile needs it

Start with the two main toggles:

• Run cleanup enables the cleanup step for that profile.
• Keep original on failure tells TypingVoi to keep the raw transcript if the cleanup step fails.

Keep original on failure is usually the safer default. It reduces the chance that a temporary model or network problem leaves you with no usable result.

Use cleanup when:

• your transcript is readable but needs punctuation and sentence boundaries
• you want filler words reduced without changing meaning
• you dictate business writing, meeting notes, support replies, or technical text that benefits from polish

Leave cleanup off when:

• you want the fastest possible turnaround
• you need the raw wording exactly as spoken
• you are checking recognition quality before adding any AI rewrite step
• you are troubleshooting language, vocabulary, or model accuracy first

4. Choose between a local LLM and a remote LLM

The LLM Model menu is a shared selector. It can show:

• None
• installed Local LLMs
• configured Remote LLMs

TypingVoi uses only one cleanup model per profile at a time.

Local LLM

Choose a local LLM when you want cleanup to run on your Mac. This is the best fit when you prefer an on-device workflow and have already installed a local cleanup model from Model Library.

Practical tradeoffs:

• no remote endpoint setup is needed
• the model must be installed locally first
• availability depends on what your Mac can run comfortably

If the local model you want is missing or not installed, go to Download and Manage Models.

Remote LLM

Choose a remote LLM when you want cleanup to use a hosted model. Remote entries are shared across profiles, then selected from the same LLM Model menu inside each profile.

Screenshot placeholder: Model Library LLM Models area showing the Remote LLMs manager with at least one saved remote endpoint and the form fields for name, base URL, model ID, and API key

Practical tradeoffs:

• you need to add the remote endpoint first
• the selected entry needs a valid API key
• remote cleanup depends on network access and the remote service being available

Use a remote LLM if you do not want to install a local model, if your Mac is not a good fit for local cleanup, or if you want to reuse one hosted endpoint across several profiles.

5. Pick the model carefully

The model selector is intentionally simple: TypingVoi does not combine multiple cleanup models in one profile. Pick the single model that matches the job.

A good rule:

• use a lighter cleanup model for quick everyday writing
• use a stronger or more specialized model for longer meeting notes, technical dictation, or more sensitive wording
• keep None selected if you only want raw transcription

If the selected model shows as unavailable later, re-open the selector and choose it again, or update the model from Model Library.

6. Set the temperature slider

The temperature slider changes how conservative or flexible cleanup should be.

• closer to Accuracy keeps cleanup tighter and more predictable
• closer to Creative allows more rewriting freedom

For transcript cleanup, start near the accuracy side. That is usually better because the goal is to clean up what you said, not restyle it heavily.

Use a higher setting only if your current prompt and model are too stiff and keep producing awkward spoken phrasing.

7. Start with a prompt preset

The Prompt Preset menu gives you built-in cleanup styles. Presets replace the current instructions, so choose one first and then edit only if needed.

Common examples include:

• General Cleanup for balanced everyday cleanup
• Light Touch for minimal rewriting
• Natural Conversation for spoken language that should still read naturally
• Business Professional for polished work writing
• Meeting Transcript for readable meeting output
• Technical Dictation for commands, product names, and technical terms

Presets are written in English, but they explicitly tell the model to keep the output in the same language as the source transcript.

8. Add custom instructions only when the preset is close

Use Cleanup Instructions when a preset is almost right but not quite. Keep your edits narrow and practical.

Good reasons to customize:

• keep specific product names or internal terms untouched
• stay extra conservative with legal, medical, or compliance language
• remove more filler words for a certain workflow
• preserve technical commands exactly

Avoid turning the cleanup prompt into a general writing assistant. The more you ask it to rewrite for tone, summarize, or invent structure, the more likely it is to drift away from the original transcript.

9. Test the cleanup before you rely on it

The Test Clean-up Transcript Here area is the fastest way to validate the current model, temperature, and prompt before you start using the profile in real work.

Screenshot placeholder: Cleanup AI test area with rough sample transcript text entered, the Test Selected Model button visible, and a Result box showing cleaned output below

Use this workflow:

1. Paste rough transcript text with obvious punctuation or spacing mistakes.
2. Confirm the right local or remote model is selected.
3. Click Test Selected Model.
4. Read the Result box and decide whether the cleanup is too weak, too aggressive, or about right.
5. Adjust the temperature, preset, or instructions and test again.

If the test fails:

• check that a model is selected
• for local cleanup, make sure the model is installed
• for remote cleanup, make sure the endpoint and API key are valid
• if you still need the raw transcript later, leave Keep original on failure turned on

10. Decide which profiles should use cleanup

Cleanup works best when you treat it as a profile-level choice, not a universal default.

Good candidates for cleanup:

• writing profiles
• meeting review profiles
• support or email reply profiles
• technical dictation profiles that need punctuation but must keep exact terms

Profiles that often work better without cleanup:

• quick capture profiles where speed matters most
• raw transcript review profiles
• debugging profiles used to compare speech-to-text accuracy across models

If you use several profiles, cleanup does not need to be on for all of them.

Related guides

• Download and Manage Models
• Choose Languages, Models, and Custom Vocabulary
• Review, Search, and Reuse Session History
• TypingVoi Settings Reference by Pane

Best next step: enable cleanup in one profile you use often, run a few tests with rough real transcript text, and keep it off in profiles where raw speed matters more than polish.