Back to docs

Accuracy Tips

Use a quality microphone

The single biggest factor in transcription accuracy is audio quality. A headset or dedicated USB microphone will outperform a built-in laptop mic, especially in noisy environments.

  • Position the microphone 15–20 cm from your mouth.
  • Avoid microphones that pick up keyboard or fan noise.
  • Check your input levels in your OS sound settings before recording.

Reduce background noise

Whisper handles noise well, but excessive background noise reduces accuracy. Where possible:

  • Record in a quiet room.
  • Mute notifications and system sounds.
  • Close windows and doors if there is ambient noise outside.
  • Use a noise-cancelling microphone or enable noise suppression in your OS audio settings.

Speak clearly and at a consistent pace

  • Enunciate words rather than rushing.
  • Avoid trailing off at the end of sentences.
  • Pause briefly between sentences — this helps Whisper detect sentence boundaries.
  • Avoid speaking over audio playing from other sources.

Set the correct language

If you speak a language other than English, set the Transcription Language in Settings to match. Providing the correct language hint significantly improves accuracy and reduces hallucinations (invented words or phrases).

See Supported Languages for the full list.

Use a prompt for domain-specific vocabulary

The Whisper API accepts an optional prompt string that can prime the model with domain-specific vocabulary or context. This is particularly useful for:

  • Technical jargon or product names that are not common words
  • Names of people, companies, or projects
  • Acronyms that should be preserved as-is

Prompt configuration is a developer-level setting and is not currently exposed in the Settings UI.

Keep recordings reasonably short

Very long recordings (over 30 minutes) increase the risk of API timeouts and may result in a single large audio file being sent to the API. For best results with long meetings, use the pause/resume feature to break content into natural segments.

Check your API key quota

If transcription fails or returns incomplete results, your OpenAI account may have hit its usage limit or the API key may be invalid. Visit your OpenAI account dashboard to check usage and billing status.

Review auto-editing output

The auto-editing feature (enabled by default for dictation) uses GPT-4 to clean up the raw transcript. If auto-editing is making unwanted changes, you can disable it in Settings → Auto-Edit.