Thank you very much for the improvements to whisper, could you have more clarity on whether there is an "hallucination" problem, i.e. an issue with duplicate output in other languages

[isbn9877007]

And can you give the amount of video memory required to run the model? Does the windows system require additional configuration?

[Vaibhavs10]

It still suffers from hallucination a bit. However, we have a fix for that coming up shortly.
See here: huggingface/transformers#27492

The max GPU VRAM should be around 12GB.

It should work out of the box on Windows (as long as you have a GPU) - do check out the FAQs for more info: https://github.com/Vaibhavs10/insanely-fast-whisper#frequently-asked-questions

(I'm closing this issue for now, feel free to re-open)

[152334H]

the code (on whisper v2/v3, with chunk_length_s=30) seems to still produce repeated text on long transcriptions, even on the latest transformers commit (i'll try to add examples later). behaviour doesn't occur on faster-whisper

adding repetition penalty kind of fixes this, but also still has unwanted consequences on the general quality of the transcript. beam search doesn't fix it