hitachi-speech/EEND

Confusion about the calculation of DER

875441459 opened this issue · 2 comments

In general, diarization tasks tolerate 250ms diar-error at the start and the end of each segment, but the function "calc_diarization_error()" in "model.py" seems do not take it into account. Did I miss something that concerns the "tolerance principle"?

calc_diarization_error() is a function to monitor DERs during training, and as you mentioned it doesn't consider the collar torelance.
Our final evaluation is conducted, e.g., here, which allows 250 ms of collar torelance.

OK, I got it. Thank you~