Diarization result for "End-to-end speaker segmentation for overlap-aware resegmentation"

[ChokJohn]

I use the pre-trained model directly for the speaker diarization task. But I got a bad result on AMI dataset, is this correct ？
AMI Mix-Headset ：
FA = 0.06082287973986739
Miss. = 0.08262689808073657
Conf. = 0.46741261779675336

[hbredin]

Yes, that is probably correct.

The sum of FA and Miss. does match the numbers reported in the paper.
The high Conf. is due to the fact that the segmentation model is not capable of tracking speakers over time (it only works on small 5s chunks).

You'd have to combine the segmentation model with speaker embedding to perform proper diarization.
See this paper and code for a way to do that.

[385614027]

你好，我想问下，例如，HIT,MISS MH,FA他们是怎样计算的，根据的是什么？？？

[Yagna24]

Hi @ChokJohn ,
How did you fetch these results, how to perform evaluation here ?