Feature decisions

Choose local text cleanup for dictation

Choose Light cleanup for routine punctuation and filler cleanup, or None for a rawer transcript with vocabulary and shortcuts still applied. Medium and High can add a configured local LLM for polishing or rewriting. More rewriting requires more review, especially for numbers, negations, and exact wording.

By WhisperJot · Documentation reviewed

Choose the smallest editing step you need

Choose the smallest editing step you need
LevelDocumented behaviorUseful starting point
NoneVocabulary and shortcuts; skips extra cleanupQuotes or wording you intend to edit yourself
LightOn-device rules for fillers, capitalization and punctuationEveryday messages; the default
MediumRules plus optional local LLM polishDraft prose that needs light editing
HighRules plus optional local LLM rewriteText where rephrasing is acceptable

Separate cleanup from meaning

Judge the result by whether it preserves the instruction. A more fluent sentence is not useful if it drops a negation or changes a deadline. Quotes and verbatim material deserve particular care because even removing a filler changes the wording.

Authored editing example — not a captured model output
Spoken draft: Um, we should not ship the payment change until the retry test passes.
Illustrative edited version: We should not ship the payment change until the retry test passes.
Review: Keep 'not', the payment scope, and the condition about the test.

Set up optional local polish

The configured local endpoint is separate from the speech engine. Choosing local cleanup does not make an opt-in cloud transcription step local. Cleanup rules are English-focused; do not assume equivalent editing quality for other languages.

The local runtime adds setup and memory use. If it does not materially reduce your editing effort, Light may remain the better choice. Consult the help center for recovery options after an unwanted AI edit.

  1. Start with Light and measure whether ordinary rule cleanup solves the problem.
  2. For Medium or High, follow the local runtime setup in the help center: MLX-LM or Ollama on Mac; a supported local server on Windows or Linux.
  3. Configure the local model and endpoint in Cleanup settings and use Test local LLM.
  4. Try a passage longer than eight words, then compare the original meaning and the final text. Short utterances skip the LLM in the documented pipeline.

Questions

Do I need an LLM for punctuation cleanup?

No. Light uses local rules and is the default. The extra local model is for the optional Medium and High polishing stages.

Does local cleanup imply that all speech stays local?

No. Transcription and cleanup are separate stages. Select a local speech engine as well, and review optional sync and the destination app's behavior independently.

Sources and related guides

Product behavior is based on the documentation below, reviewed 2026-09-09. Examples are authored illustrations; measurement worksheets are for your own observations.

Keep going