[Bug]: All exclamation marks after a certain point with ggml-distil-large, nonsense with ggml-medium

[thigger]

What happened?

Windows 11, AMD 7800XT, Intel i7-13700k

ggml-medium delivers gibberish, with a few recognisable words ggml-distil-large works reasonably until about 55% (31 minute file) when it changes to all exclamation marks. Speaker diarisation continues

Reproducible with different files; though I only have one that I can make public (will upload if needed)

ggml-distil-large all.txt ggml-medium start.txt log_2024-09.txt

Steps to reproduce

Windows 11, AMD 7800XT, Intel i7-13700k

1. Run Vibe with speaker diarisation and ggml-medium -> results in gibberish (though a few words are recognisable)
2. Run Vibe with speaker diarisation and ggml-distil-large -> great results until approx 55% (31 minute file) when all subsequent text becomes an identical-length row of exclamation marks (speaker ID continues)

I have tried this on other files (wav) with similar results. I have a podcast file which demonstrates this which I can share, but unfortunately the other files are confidential. An online search suggests it might be an audio conversion issue?

Example of (1):

Example of (2):

Speaker 4: so the clinician is in a

Speaker 4: really difficult position when they disagree with AI. And we'd always assumed

Speaker 4: the disagreements with AI was going to be when the AI is doing something that's so obviously wrong, that of course we need to disagree with the AI. But actually, AI is generally pretty good. Most of the time the decision is going to be something like the AI would have been right, sort of given the correct treatment, 20% of the time, i. This may not be the ideal treatment, but sometimes it's going to be the right treatment, and the human wanted to pick the one that was going to be right 80% of the time.

Speaker 4: !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

Speaker 4: !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

Speaker 2: !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

What OS are you seeing the problem on?

Windows 11

Relevant log output

see log file