Skip to main content
AI Enhancement is a cleanup pass that runs after transcription and before the text is inserted. It fixes spelling, grammar and punctuation, applies self-corrections, and formats lists and emails. It does not rewrite your sentences. It needs a connected language model. See Set up a language model.

How to turn it on

1

Connect a provider

Click Language models in the sidebar and connect a provider on the Configuration tab.
2

Turn on Enhance transcription with AI

The toggle sits at the top of the Configuration tab, above the provider grid. Its description reads “Automatically clean up and improve transcription using dictionary, threads and AI”.
Language models › Configuration with Enhance transcription with AI on, Pass screen as context (Beta), Custom Instructions and the provider grid The info button next to the toggle opens How AI Enhancement Works, with before-and-after examples for filler words and typos, self-corrections, lists and email formatting.

What it fixes

AI Enhancement also turns spoken email addresses into addresses: “alex at example dot com” becomes [email protected].

Filler words

Two separate steps remove fillers:
  • Settings › General › Remove filler words runs on every dictation, with or without a provider. It removes a fixed list of sounds: uh, um, uhm, umm, uhh, uhhh, hmm, mmm, mh, ehh, hm, äh, ähm, öhm. When a transcript looks German, “um” is kept, because it is a real German word.
  • AI Enhancement removes fillers more broadly, including “like” and “you know”, while keeping phrases like “I think” and “I mean”.

Spoken formatting commands

With enhancement on, say these and Vowen inserts the symbol instead of the word:

Where it runs in the pipeline

  1. Your speech model transcribes the recording. The filler filter, your Dictionary misspellings and any Threads set to run before AI are applied.
  2. AI Enhancement sends the text to your provider with your Dictionary terms, your Custom Instructions and the active Tone.
  3. Threads set to run after AI are applied, and the text is inserted into the focused app.

Turn it off for one shortcut

Alternative shortcuts for Transcription Shortcut and Hands-Free Mode each have a two-sided pill in Settings › Shortcuts: Enhance with AI or Off. A new alternative starts at Off. Use one shortcut for polished writing and another for raw text such as code or prompts. The primary shortcuts follow the global toggle. Command Mode shortcuts have no pill, because Command Mode always uses your provider. Adding alternative shortcuts is a Pro feature. See Shortcuts and Shortcut patterns.

English spelling

Settings › Language › English Spelling offers American English and British English. With British English selected and the dictation language set to English or auto-detect, enhancement keeps British spellings such as “colour” and “organise”.

Pass screen as context

Speech models often mishear names. With Pass screen as context on, Vowen reads the names visible in the window in front of you when you start dictating and offers them to the model as likely corrections. A word is replaced only when it clearly sounds like one of those names.
  • It appears under Enhance transcription with AI once enhancement is on and a provider is connected. It is marked Beta.
  • macOS only.
  • Off by default. It sends words from the front window to your AI provider.
  • It uses the Accessibility permission. No screenshot is taken.

When enhancement fails

Your words are kept. If the provider is down, the key is rejected, the request is rate-limited or the reply is unusable, Vowen inserts the unenhanced transcript and shows an AI Enhancement Failed banner ending in “Processed as-is, no AI enhancement was applied.” The AI Enhancement Failed banner with Switch to provider chips and Hide for 1hr
  • Switch to: lists your other configured providers, without the one that failed or the active one. Click one to make it your default and re-enhance the same dictation. When it finishes you see “Switched default provider. Enhanced text on clipboard — paste to replace.”
  • The chips are hidden when a Tone chose the provider.
  • Hide for 1hr stops this error from showing for an hour.
Vowen also treats two kinds of “successful” replies as failures, so they never reach your document:
  • An empty reply, usually from a reasoning model that never produced an answer.
  • A reply addressed to you instead of a cleaned transcript, such as “Please provide the transcript”.
Leftover reasoning blocks and wrapper tags around the reply are removed before the text is inserted.

Speed

Vowen asks each model for the least reasoning it accepts, because cleanup is not a reasoning task. There is no setting for this. If enhancement feels slow, choose a smaller or faster model. Groq and Cerebras are built for low latency.

Languages

Enhancement runs in the language of the transcript. For non-English dictation, pair a multilingual speech model with a capable model such as Claude Sonnet 4.6, GPT-5.4 or Gemini 3.8 Flash, and set your language in Settings › Language rather than relying on auto-detect. See Languages.