Repeated Text or Hallucinations
Dealing with repeated or imaginary text in your transcriptions.
Issue: Repeated or Imaginary Text
You might notice that sometimes your transcription contains words or phrases that you didn't actually speak, or words and phrases that get repeated. This is a phenomenon known as hallucination, where the AI model generates content that wasn't present in the original audio.
Different AI models have different tendencies for this. For example, larger models might be more prone to hallucination in some cases, while smaller models might be less prone.
Solution to Hallucination Issue
Turn on VAD
Make sure that Voice Activity Detection is turned on. To do this, go to the AI Models tab, click the gear icon next to it AI Models filter. In the settings menu that appears, ensure that Voice Activity Detection is enabled. This helps to reduce repeated or hallucinated text in your transcriptions.
If this doesn't help, the underlying model doesn't understand the audio correctly, so please use a different model for that case.
Change Your AI Model
If you are experiencing hallucinations, the best solution is to try a different transcription model.
- Navigate to the AI Models tab in the VoiceInk application.
- Select a different AI model from the list. I recommend trying a smaller model like Base or Tiny if you are experiencing issues with a larger one.
- Go to the Transcription History view, find the recording you want to fix, and re-transcribe it with the newly selected model.