Number of speakers

[subhanjansaha]

During inference, if I need to test for a track with 3 speakers, do I need to retrain the model? Because from the given checkpoint best.pt, there is a mismatch in the shape of separation.conv2d.bias and separation.conv2d.weight tensors, when tried with num_spks = 3.

[JusperLee]

yes, you need to train the new model for three speakers

[subhanjansaha]

Since the wsj0 dataset by LDC is not open-source, would you recommend training dprnn on the LibriMix dataset with different number of sources?

[JusperLee]

I think this depends on the mixed number of methods you compare.