An integrated top-down/bottom-up approach to speaker diarization