Audio-visual speech recognition based on DCM

this repo is implementing AVSR task in Fairseq==0.8.0 toolkit.

1. The dependencies are noticed in conda_env.yml file.
2. Arguments about train or inference same with speech_recognition example in the original Fairseq toolkit.
3. The model is composed about three blocks. 1) self-attention transformer based modality encoder, 2) dual-cross modality attention layer and 3) transformer based attention decoder.
4. The mel-filterbank audio features and pre-trained CNN video features are fed in the model, then the model creates character-based sentence.
5. WER and CER calculated by Sclite package using prediction and ground-truth sentences.