Listen and Fill in the Missing Letters: Non-Autoregressive Transformer for Speech Recognition