Voxtral 3's realtime transcription model identifies the end of a speaker's turn using acoustic and semantic cues. This helps developers create voice interactions that respond promptly without cutting users off during pauses.