I guess you still need a bit of audio context (the most recent) to capture some audio features not present in the text (like intonation)?