Incorporating label dependency for ASR error detection via RNN

Abstract

Recent advances in Recurrent Neural Networks (RNNs) have been successfully used in several natural language processing tasks. This paper addresses the problem of Automatic Speech Recognition (ASR) error detection and proposes a new features-based approach using a variant RNN, where words labels are re-injected as input into the network. Therefore, this study makes a major contribution to research on ASR error detection by demonstrating the utility of considering label dependency in training RNNs. Experiments are made on automatic transcriptions generated by the 2015 Sheffield ASR system applied on the MGB data. The results have shown that the proposed V-RNN greatly outperforms both unidirectional and bidirectional LSTM-RNNs.