Fine-tune a text model on /a/ threads, pair it with a voice model once they make the jump from 'clear and unobjectionable' to 'can convey emotion reliably', and feed some of the terabytes of publicly available transcript to pose data from various vtuber streams into a multimodal model, and this is 100 percent viable within like a year or two. Wan cooking tutorials galore.
At the rate open-source models are improving, you'll even be able to make her racist.