We retrained the model after fixing the bug in loss funtion and upload the pre-trained model again, trained on VB with 0.5 as the dropout ratio.
Using this pre-trained checkpoint, we got:
csig:3.8264 cbak:3.2791 covl:3.1965 pesq:2.5878 ssnr:9.3684 stoi:0.9336 on VB
csig:2.7520 cbak:2.1276 covl:2.0501 pesq:1.4693 ssnr:1.8087 stoi:0.8304 on CHIME-4