Thanks for the question.
I think @etzinis, @giorgiacantisani or @popcornell might have ideas about this?

I found the reason. MixIT is not effective for all 2+2 situations, and you need to enter a single source to ensure performance.

Samuele mentioned that he also had some problems with it.
What 2+2 situation didn't it work on please?

Should work also with 2+2 mixtures. Actually should work with arbitrary mixtures.
Okay with noise the result it's not so great but still works quite well. Check for example the model trained on YFC100 on Google separation team GitHub.

Yeah right but then it could be argued that it is not anymore unsupervised because you need single speaker speech.
You can try my implementation from here:
#320

I don't understand the one from Giorgia fully, it is much more flexible but maybe it has a bug IDK.

Single source mixtures are not needed. What probably might be needed and actually boosts the performance is adding a percentage of mixtures of mixtures containing a single mixture and zero out the second one. If the mixtures contain independent sources (MixIT assumption), that means that this would be identical to having a mixture of two mixtures where each contains a single source.

@etzinis Yes, adding a certain percentage of mixtures of mixtures containing a single mixture can improve performance. But I see no problem with asteroid's MixIT implementation.