On training the recurrent neural network encoder-decoder for large vocabulary end-to-end speech recognition.