Strange and chaotic results on LRS2

[wcycqjy]

Hello. I trained on LRS2 by following the recipe of auto_avsr and using the pretrained LRS3 AVSR model avsr_trlrwlrs2lrs3vox2avsp_base.pth. The LRS2 dataset is processed by following your instructions. But the results is strange:WER is above 1. Even if I use avsr_trlrwlrs2lrs3vox2avsp_base.pth, the WER is still above 1. And the predicted sentence is not correct, like the truth is "And for me the surprise was", but the predicted one is "It's time for the final round which".

Do you have any idea about this? Thank you in advance.