Skip to main content

Table 6 SDRi/SI-SNRi(dB) performance on WHAMR! test set with SCT-2

From: Heterogeneous separation consistency training for adaptation of unsupervised speech separation

SCT

System

#Iter

CPS-2

Oracle

SCT-2

Conv-TasNet

1

8.28/7.85

8.64/8.28

  

2

8.48/8.06

8.68/8.27

 

DPCCN

1

9.26/8.81

9.31/8.86

  

2

8.84/8.40

8.95/8.52

  1. “Oracle” and \(\eta\) have the same meaning as in Table 5. All source models are well pre-trained on Libri2Mix. The best setup of \(\{\alpha ,\beta ,\eta \}\) are \(\{8,5,8\}\) and \(\{12,5,8\}\) in the 1st and 2nd SCT iteration, respectively