Current mainstream speaker diarization models (such as PyAnnote and the relevant modules in NVIDIA NeMo) can already achieve high accuracy under clean recording conditions, but there is still room for improvement in complex scenarios such as overlapping speech, background noise and accent variation.