From text to performance
G.E.N.E.S.I.S. gives characters something to say. Genesis Voice is concerned with how it is said, which is where most of the believability actually lives.
Flat delivery of well-written lines reads worse than adequate lines delivered with the right emotion, so the research is on prosody and appropriateness rather than raw audio quality.
Memory across encounters
A character who forgets you between conversations is not a character. Persistent conversational memory means recalling that you have met, what was discussed, and how it went.
The difficulty is selective recall: perfect memory of every interaction is both expensive and uncanny. Deciding what a character should retain and what it should plausibly forget is the interesting problem.
Synchronisation and language
Automatic lip synchronisation from generated audio is required for this to be usable, since dialogue is produced at runtime and cannot be authored against animation.
Multi-language support is a first-class goal rather than a later addition, which constrains the architecture: the pipeline must be language-agnostic from the start.