@withenoughcoffee.com loved the shout-out for George Washington Carver during the episode! I grew up in SW Missouri, visiting the national park and hearing about all the things he discovered. He is an excellent example of curiosity and persistent determination.

This is the full command I've used:
whisper \
--model large \
--language Turkish \
--task translate \
--compression_ratio_threshold 2.4 \
--logprob_threshold -1.0 \
--no_speech_threshold 0.6 \
--condition_on_previous_text False \
--verbose True \
--threads 6 \
--fp16 False \
input/video_file.mp4

I think most of my success eliminating the looping was from adjusting these options:
--compression_ratio_threshold 2.4 \
--no_speech_threshold 0.6 \

I found a bit of information about what some of the options do here: github.com/openai/whisp...

It seems to do this more on long form content I think where there are long periods of silence or something. I’ve had a bit of success by tweaking some of the command line arguments to produce transcripts that don’t get stuck in loops myself.

It turns out you can add a filter config setting to the host lvm config file so that the host will ignore the guest devices that show up as ZFS volumes on the host. This seems to have fixed my issues in some initial testing.

On the guest `lvm pvs` displays the attached disk in the guest VM (/dev/nvme0n1), but on the host `lvm pvs` displays the zfs volume (/dev/zd240) that was assigned to the VM. I'm still trying to make sense of what's going on here and why.

I found that this somehow relates to lvm, and in the guest machines there is a configuration task for bootstrapping a Ceph OSD with a command `ceph-volume lvm create` and this is applied to an entire attached disk. On the host if I remove the volume groups it lets me destroy the test cluster.