How to make transcribe long file

[vshlykov]

I try to transcribe audio about 30 minutes long. And after 10-12 minutes of transcribe Whisper starts to make a lot of errors in text. Why?

I use model large-v3 and cuda.

And this is my code:

result = model.transcribe(audio, language='ru', beam_size=5, best_of=None, verbose=False, suppress_tokens="")

[vshlykov]

I tried beam_size up to 60 and stil have a lot of errors after 10-15 minutes.