Efficient speaker diarization and low-latency speaker spotting