Improved Noisy Student Training for Automatic Speech Recognition

Noisy Student training is applied to ASR using bidirectional LSTM networks. A language model is used to filter out unlikely recognized sequences by the teacher model.