/jp/ - Otaku Culture

Hey guys /g/ here, been working on artificial intelligence for my master's thesis and been using Kizuna Ai as the standard for what I want to achieve. I picked Kizuna specifically because she has such a ridiculously large range of activities and comments she makes and she seems to be extremely versatile. Compared to the other youtuber AI personalities like hololive which seems to be popular recently, she just seems more natural while hololive characters tend to get into basic looping errors where they tend to repeat the same dialogue or comments repeatedly and often have overlap with one another which leads me to believe that their datasets are lacking and they weren't given sufficiently large examples to learn from. Kizuna seems to be a little more gifted in that sense and I want to work off that.

Does anyone know the specifications or what datasets they used to train her? Its ridiculous trying to get that level of natural speaking ability in an AI. I've been using datasets using around 200TB of youtuber let's play videos and host commentary, but its been a slog to say the least, and even with that training it tends to be lacking the same versatility Kizuna is.

Anyone might know where I can find her technical specifications? Thanks.
Keep in mind I'm only looking at the voice, I'm sure they have additional editors for her videos and animations since it seems to be more choreographed, especially the music videos.

>>25047572
Black shorts with white belt and white top. Not complaining, looks kinda hot.

>>20493077
What about big fat ponytails?