​
Overview

Hallucinations in AI refer to instances where the system generates content that wasn’t provided by the user or produces unexpected results. In Superwhisper, these issues can occur in two distinct processing stages:

  1. Voice Transcription: When your spoken words are converted to text
  2. AI Processing: When the transcribed text is processed according to the selected mode

Check the History tab to identify exactly where hallucinations are occurring. You can analyze both the voice (transcription stage) and the AI-processed result to pinpoint the source of the issue.

​
Transcription Hallucinations

Hallucinations during the voice transcription stage typically stem from 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.

​
Silence Removal

1

Open Configuration Tab

Navigate to the Configuration tab in Superwhisper’s Settings.

2

Access Advanced Settings

Find the Advanced Settings section within Configuration.

3

Enable 'Remove Silence'

Toggle on the “Remove Silence” option. This will cut out silent periods from your audio before transcription, significantly reducing related hallucinations.

​
Vocabulary

If you continue to experience hallucinations in the transcription stage after activating silence removal, check your vocabulary tab and make sure it only includes essential terms. The transcription model performs best with a focused vocabulary list.

Vocabulary Best Practices

Learn about how Vocabulary can improve your transcription accuracy.

Language Detection Issues

If your dictation is being incorrectly transcribed in the wrong language, this may also be due to Vocabulary.

​
LLM Hallucinations

Hallucinations during the AI processing stage can manifest as:

  • AI answering questions you didn’t ask
  • Incorrect formatting of your dictated text
  • Additional comments or content which was not part of your dictation

These issues typically arise from the following:

​
Custom Mode Prompts

The built-in modes in Superwhisper are optimized for reliable results. When creating custom modes, proper prompting becomes essential.

If you’re using custom modes, hallucinations may occur due to:

  • Imprecise prompt instructions
  • Lack of clear examples
  • Conflicting directives in your prompt

Prompting Tips

Learn effective techniques for crafting custom mode prompts that produce reliable, consistent results. Discover how to structure instructions, provide examples, and avoid common pitfalls that lead to unexpected outputs.

​
Model Capability Limitations

Not all AI models have the same capabilities. They vary in performance and reliability.

1

Cloud Models

Cloud models generally provide the best performance and are less prone to hallucinations with complex requests.

2

Local Model

Local models may have more limited capabilities and require more specific prompting to avoid hallucinations.

Superwhisper can use different types of context to enhance AI processing, but these can sometimes lead to hallucinations:

1

Application Context

When enabled, Superwhisper captures text from your active input field. If there’s too much text or irrelevant content, it may confuse the AI.

2

Clipboard Context

If you accidentally copy text during dictation with clipboard context enabled, this unrelated content may confuse the AI model.

Best practices for context usage:

  • Only enable context features in modes where they’re specifically needed
  • Use application context selectively in text-heavy applications
  • Be mindful of clipboard operations during dictation

Context Awareness

Learn how context awareness features work in Superwhisper, when to use them, and how to avoid related issues.