Ask HN: Talking to the Dead with Large Language Models

I was thinking about a idea for a company, and I was wondering how much data it would take to actually make feasible.

With Large Language models today, would it be possible to record all the audio from a person for say ~2 years, autotranscribe the recordings and use it to either train or heavily prompt inject an LLM?

Perhaps this would be a way for those near death to persist a version of themselves beyond the grave?