As mentioned in the section 3.1 of the submission, we incorporate three separation backbones (ConvTasNet, DPRNN, and Sepformer) to evaluate the effectiveness of our proposed AC-SIM and multi-loss training paradigms. Since three backbones are trained with different hyperparameters and configurations in prior and original works, we attach this information below:

1. ConvTasNet: we set N=512, L=32, B=128, H=512, Sc=128, P=3, X=8, R=3; the learning rate is 0.001
2. Dual Path RNN: we set N=256, L=16, B=64, H=256, K (chunksize)= 250, LSTM Hidden Dimension=128; the learning rate is 0.001
3. Sepformer: we set N=256, L=16, IntraT-InterT-N=2, Nintra=8, Ninter=8, Nhead=8, dffn=1024; the learning rate is 0.00001

For all models, we use the Adam optmizer with &beta=(0.9,0.99) and a scheduler, which halves the learning rate if the average SI-SDR performance over (D-All, D-NE, D-NR, D-ON) four vadlidation sets is not improved after three successive epochs. We also set the gradient clipping to limit the L2 norm of the graidens to 5.

As mentioned in the section 3.1 of the submission, we collected a real-world speech test set with 16 cases. Some of them are from the Look-to-Listen Real-World Material (paper link) We attach the source link (Youtube) of these cases below: