How to use another whisper models? Please, help.

[ShaneTracey]

I have encountered that bug which makes subtitles repeat one phrase many times in a row. So the problem is in the whisper model itself, but when I download another whisper model from hugging face, player shows an error. So how to use another whisper models and which ones exactly should I use? Because in hugging face whisper models are not just one .bin file like player provides me with. Please, help. Player is so helpful!

[umlx5h]

Hi @ShaneTracey
LargeV3 and LargeV3Turbo are known to be prone to hallucination bugs (repeats).
I use LargeV2 + CUDA and have very few problems.

There are some other workarounds that can be taken, but they are difficult to solve fundamentally, so it is necessary to wait for the whisper library to fix it.

#2 (comment)

Is the downloadable model in the player not sufficient?
It appears that the models need to be converted to ggml format. If this is placed in the whispermodels folder with a specific name like ggml-largev2.bin, it will probably work.
(After placing a model with specified name, you need to select model from Subtitles -> ASR in settings)

It seems to be possible to convert it with a python script, please check the second link.

https://github.com/sandrohanea/whisper.net?tab=readme-ov-file#ggml-models
https://github.com/ggerganov/whisper.cpp/tree/master/models

[ShaneTracey]

Thank you a lot! I will try the ways you propose.