Offset delays in start of the speaker

[Tortoise17]

I have tried to use the dafault values of config for offline speaker diarization. There is consistent issue in the returned diarization results. There is approximately 2 seconds of delays in the start of detected speaker in most of the cases. I can say almost all cases. Is there any reason for that ? or any specific advise to handle this? I am using the pyannote audio recent available branch installed, and the offline mode of the speaker segmentation and speaker diarization. I tried both on CPU and CUDA. Result is the same.