Neil Zeghidour

Thanks so the above code should work with v1.0.0?

Hey Manuel, here are first answers for the model: > not clear if different encoders are used for separation and speaker stack. (from image in the paper it seems so)...

Hi @HuangZiliAndy what error are you referring to? There is indeed an averaging, AFTER computing the loss of the speaker stack. The speaker loss is applied at every time step...

That's correct, that is a bad typo! We will update the paper to correct this, thanks for spotting! We indeed wrote the log probability instead of the loss (which is...

The new version of the paper is now online: https://arxiv.org/pdf/2002.08933.pdf

Hey @popcornell , one-hot instead of embeddings should be fine but a bit worse than using the actual embeddings. Also for the SDR loss we use at each layer of...

https://github.com/mpariente/asteroid/blob/aa3485bc57ef1f21d37cff43fce69cc3c9c94b1c/egs/wham/WaveSplit/losses.py#L38 Initializing with one-hot works better.