Interesting point, although I don't see how you'd manage to run modern TTS (the models can get very large, and that's per voice; as an example Parler-TTS's mini model is 800Mb, the HQ model is 2.3Gb - for one voice) + a LLM for content synthesis on any personal hardware, console or not. The storage requirements alone would make that grossly infeasible.

Thing is, I really don't see how that'd work for any useful scenario. Even in this very limited case, natural language processing and response generation is definitely off-loaded to a remote server farm. It certainly isn't running on the console. TTS for a single characters voice could - but likely isn't - be done on the hardware, but as I mentioned elsewhere, the voice models for modern text-to-speech synthesis are very, very large. Typically gigabytes of data - per voice. Completely unrealistic for any meaningful game.

Even if all those trained models existed (they don't - and making one is considerably more work and expense than having a voice actor just deliver a fixed set of lines), I doubt any consumer would be amused by a multi-terabyte download.

Even as it is, the response latency is hilariously bad. I imagine players having to wait seconds for each NPC response would fly as well as a lead brick.