This is the simplest AVSR recipe (including VSR, ASR, AVSR) for GRID with icefall.
For VSR, there are two models: Conv3d Map BiGRU CTC Model and Conv3d ResNet18 BiGRU CTC Model.
The WERs for them in TEST dataset are 15.68% and 13.63%.
For ASR, there is a model: Tdnn Lstm CTC Model. The WER for it in TEST dataset is 2.35%.
For AVSR, there is a model: CombineNet CTC Model. The WER for it in TEST dataset is 1.71%.