I'm running an i9 12900 and an RTX 3070ti so I might push it a little more. It's very intriguing. I'm starting to think of albums that I wish had different vocalists and what could be done to fix that. It has the potential to be quite amazing.
It does, although honestly Anders is not the best vocalist to train an AI model from, owing to the fact his vocals and the production of albums changes so much. You're limited to one album, maybe two at most as far as consistent training material is concerned.
For growls you might get away with TJR/Whoracle combined, although Anders pronounciation dramatically improves on Whoracle, and the production is noticeably different. Colony and Clayman you again may get away with, but Clayman's vocals are weaker than Colony's overall. Reroute's vocal production is so different that I'm not sure it fits in with anything else. STYE and CC are another two you may be able to pair up, although again production differences could make it awkward. After that you're kind of fucked, due to Anders' change to the yell/scream combination which an AI isn't going to be able to easily replicate or clone. Truthfully though, I'd question why you'd even want to try with his vocals after CC. SOAPF and Foregone probably the pick of the bunch post-2006, but both also wildly different in terms of approach.
When it comes to cleans, all I can say is good fucking luck. I don't think there's any two albums where they sound the same, with the usage of ProTools in particular ramping up massively after STYE. I'd say the best two to combine for cleans would be Reroute and STYE, but even then you have quite big production differences.
Compare to somebody like Stanne, who I think you'd have a lot more material to work with in terms of training a model. You could basically take Enter Suicidal Angels all the way up to Haven and not really encounter any significant differences in growling style or really glaring production changes. Damage Done is very different in terms of vocal style, mainly because Stanne's voice was failing by that point, so you'd probably have to exclude that. Character and Fiction would go fine together, and then from WATV to the present you have a pretty clear run of training material if, for some reason, you wanted to make a model of Robo-Stanne.
In terms of cleans also plenty of consistent output to work with, as Stanne's cleans have remained pretty consistent since 1999. Shouldn't be much of an issue in terms of building a model with lots of audio to work with.