CoreML not usable for large files anymore?

[mirozahorak]

I have upgraded whisper to latest version and downloaded large-v3 model.
whisper.cpp was working wonderfully for me before.
I am processing the same kind of files as before 30-80 minutes.
Now it is very hard to achieve the same quality and even finished transcription of files.
Many times, after 30-40 minutes, it just repeats the sentences.

But most of the time in just crashes (sometimes after a few minutes, sometimes after 20 minutes) with error:

whisper_full_with_state: failed to decode
/Volumes/DEVEL/TRANSCRIPTION/w2/whisper.cpp/main: failed to process audio

I have tested all kinds of settings, and they have an influence, but the previous quality and simplicity is impossible to achieve. With default settings it just does not work anymore.

I have recompiled and double tested, but either something is broken or i am doing something wrong.
Let me know what info i can provide to help solve this problem.

i have tested changing parameters:

-bo 8 -mc 64 -bs 8 -et 2.9 
-bo 8  -mc 56  -lpt -0.9  -wt 0.005 -sow

and other combinations of above, but while they change behaviour, i was not able to achieve quality of whisper.cpp version i installed in august with large-v2 model

[bobqianic]

I have upgraded whisper to latest version and downloaded large-v3 model.

Now it is very hard to achieve the same quality

It's better to use large-v2 instead, as the problem lies with large-v3 itself, which has experienced a significant decline in quality compared to large-v2.

openai/whisper#1762

and even finished transcription of files.

whisper_full_with_state: failed to decode
/Volumes/DEVEL/TRANSCRIPTION/w2/whisper.cpp/main: failed to process audio

Give large-v2 a try and check if the issue persists.

[ggerganov]

Can you test: #1633

[simicvm]

Can you test: #1633

seems like this is still a problem. on big files, large-v3 at some point just starts repeating sentences. large-v2 transcribes them without issues.