Skip to main content
Words you never said, repeated phrases, or content the AI added on its own are called hallucinations. They come from one of two stages:
  1. Voice transcription: your spoken words become text
  2. AI processing: the text is processed according to the mode
Open the dictation in History and compare the raw transcription with the AI result to see which stage produced the problem.

​
Transcription hallucinations

Hallucinations at the transcription stage have two main causes:

Silence

Periods of silence can confuse the transcription model, causing it to generate words or sentences that weren’t spoken.

Vocabulary

The vocabulary prompt sent to the transcription model can sometimes lead to confusion, especially when too many custom words are added.

​
Fix: silence removal

Superwhisper's Silence Removal
1

Open the Sound tab

Go to the Sound tab in Superwhisper’s settings.
2

Enable Remove Silence

Toggle on Remove Silence. It cuts silent periods from your audio before transcription, which removes most silence-related hallucinations.

​
Fix: trim your vocabulary

If hallucinations continue after enabling silence removal, check your Vocabulary tab and keep only essential terms. The transcription model performs best with a short, focused list.

Vocabulary Best Practices

How vocabulary and replacements improve transcription accuracy.

Language Detection Issues

Dictation transcribed in the wrong language can also come from vocabulary.

​
AI processing hallucinations

At the AI processing stage, hallucinations look like:
  • The AI answers a question you didn’t ask
  • Your dictated text is formatted incorrectly
  • Comments or content appear that weren’t part of your dictation
These have three usual causes:

​
Custom mode prompts

Built-in modes ship with tuned instructions. In custom modes, the prompt is yours, and hallucinations usually come from:
  • Imprecise instructions
  • Missing examples
  • Conflicting directives

Prompting Tips

How to structure instructions, provide examples, and avoid the common causes of unexpected output.

​
Model capability

Models vary in capability. Larger cloud models handle complex requests with fewer hallucinations. Smaller local models need more specific prompting. If a mode hallucinates with one model, try a more capable one.

​
Context problems

Context features improve results, but they can also mislead the AI:
  • Application context: if the active input field holds a lot of text or irrelevant content, it can confuse the AI
  • Clipboard context: text copied accidentally during dictation gets sent as context
To limit this, enable context only in modes that need it, and avoid copying unrelated text while dictating.

Context Awareness

How each context type works, when it’s captured, and how to fix related issues.