Text this: Towards efficient recurrent architectures: a deep LSTM neural network applied to speech enhancement and recognition