PIT with nn.CrossEntropyLoss() from MIRNet

[Yang Wang]

🚀 Feature

This is to assign speaker ID of each separated signals with PIT.

Motivation

Authors from MIRNet considers all of the loss terms by calculating every possible permutation of candidate pairs. They computed the PIT loss on estimated speaker identity information. In addition, they use cross-entropy loss with a classifier for the speaker embeddings.

What you'd like

The entire training criterion is as follows:

[Pariente Manuel]

This sounds very interesting, do you want to give it a try?