Direct Cartesia SSE, add_timestamps on. Each word reveals off the audio's own playback clock. Watch + listen: the last word must still land on time (no drift).