search

ggerganov / whisper.cpp

Port of OpenAI's Whisper model in C/C++
MIT License
35.82k stars 3.65k forks source link

Webspher often produces hallucinations #1824

Open dfengpo opened 9 months ago

dfengpo commented 9 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.

bobqianic commented 9 months ago

Converting the same file multiple times can also cause hallucinations.

Since conversion doesn't require quantization, the weight should remain the same.

The model I am using is large-v3.

Try large-v2

dfengpo commented 9 months ago

Converting the same file multiple times can also cause hallucinations.

Since conversion doesn't require quantization, the weight should remain the same.

The model I am using is large-v3.

Try large-v2

Quantification is to reduce hardware and improve running speed. I heard that the distillation model is faster. May I know if it is supported？

hlevring commented 9 months ago

https://github.com/ggerganov/whisper.cpp/pull/1424