typically 0.8.94.

De Moon versus Reza, 0.3.58, close quote.

Then persistent automatic identification becomes easy.

If instead everything overlaps around 0.65.75,

we know the problem is harder.

And critically, I'd identify the recording-level diarized speaker,

not every utterance.

Feed Titanet a decent amount of clean speech

from each Pionote speaker cluster.

Given that your RTX has only 8GB VRAM,

this is actually one of those cases

where switching models may make things easier rather than harder.

Titanet is tiny compared with the models

that are currently causing your GPU juggling.