Ask HN: Could an LLM predict someone's speech 5 seconds ahead in realtime?

LLMs are in many ways text predictors. Could you have an LLM that runs in realtime, watching a speech on TV for example, and predicts what the person is going to say 5 seconds in advance based on what they are currently saying? My friend was thinking about this as a test for State of the Union but then we realized that the transcript is published ahead of time so it would not be that impressive. But maybe something like the inauguration speech? Especially if tuned on previous speeches from that person.