[joegibbs]

You could, the further into the future that it predicts the less accurate the prediction will be, but you could probably predict a few sentences of the speech correctly. The LLM can only guess from historical context and grammar though, and the speaker is unlikely to make the most likely next word choice for all words (there are some sentences where the next word is extremely likely, mostly aphorisms, but a lot of the time you don't know where it's going).

You would want to use a raw LLM which just does next token completion rather than a model like Claude or ChatGPT, and you could probably use it to display branching chains of next most likely strings.