A novel RNN cell named VRNN, which consumes less parameters and training time but gain comparable performance as LSTM/GRU