We use PyTorch to build the LipNet model with minor changes. This reproduction achieves 13.3%/4.6% WER in unseen/overlapped testing, which exceeds all evaluation metrics in the original paper and ...