Do you send the model mixed audio or s separate track for everyone ? When people talk at the same time, how to speaker diarization ？