Hybrid Input-type Recurrent Neural Network Language Modeling for End-to-end Speech Recognition

The out-of-vocabulary (OOV) words is a problem that impacts recognition accuracy, whether it is the HMM model, DNN model, or end-to-end speech recognition. This paper proposes a hybrid input-type recurrent neural network language model (RNNLM) for end-to-end speech recognition, which uses word and pseudo-morpheme (PM) as a sub-lexical unit during training. The advantage of PM is a new vocabulary, or unseen vocabulary can be reconstructed from sub-lexical units. The results showed that the accuracy of using the proposed method could reduce the error rate by 1.28% compared to the conventional end-to-end technique.