Fascinating! This is relative time in a continuously processing voice model, here, they're using an LLM with absolute time.