Keras: How to predict lstm with one time step?

Question

I use audio specgram as feature to fit into model. How to predict with one timestep? For example, the model is like:

def getModel(input_shape):
    X_input = Inpu