Context-Dependent Pre-trained Deep Neural Networks for Large Vocabulary Speech Recognition