search

FrenchKrab / IS2023-powerset-diarization

Official repository for the "Powerset multi-class cross entropy loss for neural speaker diarization" paper published in Interspeech 2023.
68 stars 4 forks source link

multilabel usage. #7

Closed tujie-jiangye closed 4 months ago

tujie-jiangye commented 4 months ago

Can you provide a demo for using multilabel model for inference.

FrenchKrab commented 4 months ago

It should be exactly the same as that part of the readme, but you need to uncomment the 'threshold' parameter in the params (because multilabel model outputs a 'probability' of speaker activity (in [0,1]) that needs to be discretized with a threshold to obtain binary activity).

And of course you'll need to swap the model path to a multilabel model.