sandrohanea / whisper.net

Whisper.net. Speech to text made simple using Whisper Models
MIT License
505 stars 77 forks source link

Webspher.net often produces hallucinations #162

Closed dfengpo closed 5 months ago

dfengpo commented 5 months ago

How to solve the problem of hallucinations caused by using wispher.net? The model I am using is large-v3. Converting the same file multiple times can also cause hallucinations.

sandrohanea commented 5 months ago

This is a normal limitation of Whisper model, inherited in Whisper.cpp and Whisper.net. However, you can try to fine-tune the settings for your use-cases (check Temperature for example) https://github.com/sandrohanea/whisper.net/blob/main/Whisper.net/WhisperProcessorBuilder.cs#L333C3-L346C6

dfengpo commented 4 months ago

I adjusted the temperature to 0, but it had no effect and I still frequently experienced hallucinations. Whether using large v2 or large v3, it's the same