Skip to main content
Most quality problems come down to three things: the speech model, whether it receives your Dictionary, and the language setting. Check them in that order.

Check your model first

Changing the model has the largest effect on accuracy. Models are on the Speech models page in the sidebar. See Speech models.

Dictionary words have no effect

Whether your Dictionary changes what the model hears depends on the engine. These models receive no vocabulary at all:
  • Whisper Tiny and Base, including the English versions
  • Parakeet, every version, on macOS and Windows
  • Nemotron EN and Nemotron Multilingual
  • Cohere Transcribe
On these, Dictionary words never reach the speech model. Nothing is broken. Parakeet is the default model, so this affects many new users. The full per-engine table, with limits, is in How vocabulary reaches each model. These still work on every model:
1

Add misspellings to the Dictionary word

A misspelling is replaced with the correct spelling after transcription. See Dictionary.
2

Add a Thread

A Thread replaces a phrase with fixed text every time. Triggers match whole phrases, so a sig thread does not rewrite the middle of “design”. See Threads.
3

Turn on AI enhancement

AI enhancement receives your Dictionary as context and fixes words the model got wrong. This is the most effective fix on Parakeet and Nemotron. Turn it on in Language models › Configuration › Enhance transcription with AI.
4

Switch to an engine that uses your Dictionary

Moving from Whisper Tiny or Base to Whisper Small is the smallest change that turns vocabulary on. For the strongest effect, use Deepgram Nova 3, AssemblyAI, ElevenLabs Scribe v2, Speechmatics or Gemini gemini-3.5-transcribe.
When your list is longer than a provider allows, Vowen sends the terms you say most often first, then the most recent.

Set your language

Auto-detect sometimes picks the wrong language, and a wrong-language transcript is far worse than a weaker model.
  1. Open Settings › Language.
  2. Next to Dictation Language, click Change and pick your language.
The list changes with your model. Deepgram Nova 2 offers 33 languages, Nova 3 offers 50, AssemblyAI streaming 18, xAI 25, Mistral 13, Sarvam 23, Speechmatics 39 and Cohere Transcribe 14. If you switch to a model that does not support your saved language, the setting resets to auto-detect. English-only models such as Parakeet V2, Phonon-2 and Nemotron EN lock it to English. See Languages.

Turn on AI enhancement

AI enhancement fixes most remaining errors and receives your Dictionary as context. Turn it on in Language models › Configuration › Enhance transcription with AI. If no AI provider is connected, you see a locked card with Connect an AI provider instead of the toggle. See AI enhancement.

Specific issues

  • Slow down slightly. Fast speech is harder for every model.
  • Switch to a cloud model. Deepgram Nova 3 and ElevenLabs Scribe v2 handle fast speech well.
  • On local Whisper, move up a size.
Check your engine first. See Dictionary words have no effect.
  • Whisper Tiny or Base, Parakeet, Nemotron or Cohere: use misspellings, a Thread or AI enhancement, or switch engines.
  • An engine that uses your Dictionary well: add the terms in Dictionary.
  • Local Whisper Small or larger, or Groq: add them to the Dictionary. They are sent as a hint, which helps less.
  • Use a multilingual model. Whisper English sizes, Parakeet V2, Phonon-2 and Nemotron EN are English only.
  • Set the language instead of relying on auto-detect.
  • Parakeet V3 and Parakeet Ultra cover 25 European languages. Cohere Transcribe covers 14 languages on a Mac.
  • Parakeet Japanese and Parakeet Mandarin are dedicated Mac models.
  • For Indian languages, use Saaras v3 (Sarvam AI).
  • Whisper Large v3 covers 99 languages.
Repetition happens most on smaller models.
  • Use Retry transcript on the History entry to run the saved audio through a different model.
  • Move up a model size, or to a cloud model.
  • AI enhancement removes repetitions.
Whisper models can invent text over silence, most often “thank you”.
  • Start speaking as soon as you press the shortcut.
  • Vowen already drops local results that are exactly “thank you”, “thank you.”, ”.” or a lone period with a line break.
  • For cloud models, turn on Settings › Experimental › Enhanced silence detection. A Sensitivity slider from 1 to 5 appears, set to 2 by default. Settings › Experimental with Enable Connectors, Capture screenshots while dictating and Enhanced silence detection
The recording was probably too short.
  • Hold the shortcut for the whole time you speak, or set the shortcut to Tap in Settings › Shortcuts.
  • Parakeet needs at least 300 ms of audio. Anything shorter is discarded with no error and no History entry.
  • Check that the app you are typing in is not taking the shortcut. See Shortcut issues.
See First or last word cut off. Use Retry transcript on the History entry to check: it always re-reads your original recording.
Smart spacing and capitalization matches your dictation to the text around the cursor. When Vowen can’t read that text reliably, it inserts the dictation as spoken. Check that Settings › General › Smart Spacing and Capitalization is on. It is off by default. See Smart spacing and capitalization.
Set Settings › Language › English Spelling to British English. You don’t need an AI instruction for it.
Mac, fully local transcription:
  • Speech: Parakeet Ultra, or Nemotron EN for a punctuated live preview.
  • AI enhancement: any connected provider. Gemini 3.8 Flash is Google’s recommended model.
Windows:
  • Speech: Parakeet V2 or V3 locally, or Whisper Large v3 Turbo on Groq.
  • AI enhancement: Gemini 3.8 Flash.
Best accuracy, any platform:
  • Speech: ElevenLabs Scribe v2, AssemblyAI Universal-3.5 Pro, Deepgram Nova 3 or Gemini gemini-3.5-transcribe. All of them use your Dictionary well.
  • AI enhancement: Claude Sonnet 4.6 or GPT-5.4.
The AI model list changes over time. Open Language models › Configuration for the current models of each provider.