Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.vowen.ai/llms.txt

Use this file to discover all available pages before exploring further.

Low Accuracy / Many Errors

Try a Better Model

The #1 fix for accuracy issues is using a better model:
Current ModelUpgrade To
TinyBase.en or Small
Base.enSmall or Medium
Any local modelGroq Whisper Large v3 (free, cloud)
Any model (non-English)Large v3 or Groq Large v3

Set Language Explicitly

If you use Auto-detect, switch to a fixed language:
  1. Go to Settings > Language
  2. Select your primary language
  3. Auto-detect can occasionally pick the wrong language, reducing accuracy

Enable AI Enhancement

AI Enhancement catches and fixes most remaining transcription errors. Enable it in Settings > Recording > Enhance with AI.

Specific Issues

  • Slow down slightly — rapid speech is harder for all STT models
  • Switch to a cloud model (Groq, Deepgram) which handle fast speech better
  • Use a larger local model (Medium or Large v3 Turbo)
Add them to your Custom Vocabulary:
  1. Go to Settings > Dictionary
  2. Add the terms (product names, jargon, proper nouns)
  3. The model uses these as hints during transcription
For consistent misspellings, create a Thread replacement: Settings > Dictionary > Threads
  • Use a multilingual model (NOT .en variants)
  • Try Large v3 or Groq Whisper Large v3
  • Set the language explicitly (don’t use auto-detect)
  • For Indian languages, try Sarvam AI Saaras v3
  • Parakeet supports 25 European languages well
Repetitive output is a known model behavior that happens occasionally:
  • Re-record the phrase
  • Try a different model
  • Enable AI Enhancement (it detects and removes repetitions)
  • This is more common with smaller models — upgrade if it happens frequently
This is a Whisper model quirk when processing silence:
  • Start speaking immediately when you press the shortcut
  • Don’t hold the shortcut in silence before speaking
  • Update to the latest Vowen version (improved junk detection)
  • Enable cloud silence detection in Settings > Recording
The recording might be too short:
  • Hold the shortcut for the full duration of your speech
  • Don’t release too quickly
  • Check that your shortcut isn’t conflicting with another app
  • On Windows, check that the Command key isn’t part of your shortcut (known issue with left Command)
Wispr Flow uses cloud-based models by default. To match their quality:
  1. Use Groq Whisper Large v3 as your transcription model
  2. Enable AI Enhancement with Gemini Flash Lite or Groq Llama
  3. Add Custom Instructions matching your writing style
This combination produces results comparable to Wispr Flow’s output.

Comparing Your Setup

If you’re unsure whether your setup is optimal, here’s what most satisfied users use: macOS (fast + accurate):
  • Speech: Parakeet or Groq Whisper Turbo
  • AI: Gemini Flash Lite (free) or Groq Llama 3.3 70B
Windows (fast + accurate):
  • Speech: Groq Whisper Large v3 Turbo
  • AI: Gemini Flash Lite (free)
Maximum accuracy (any platform):
  • Speech: Groq Whisper Large v3
  • AI: GPT-4o or Claude Sonnet