An Unsuper vised Adaptation Method for Deep Neural Network-based Large Vocabulary Continuous Speech Recognition

摘要

Recently, the Deep Neural Network (DNN) to acoustic modeling has achieved great success on Large Vocabulary Continuous Speech Recognition (LVCSR) tasks. However, the performance of the DNN-based Automatic Speech Recognition (ASR) systems still surfer greatly from the mismatch between training and testing in real applications. In the past, commonly used methods for DNN-based acoustic model adaptation have focused on supervised methods such as affine transformation, regularize training. There has been little work on unsupervised adaptation methods for DNN. However, in many cases, it is very expensive to get the matched data with transcription, but there are tremendous unlabelled data available. In this paper, a novel unsupervised adaptation approach is proposed to mitigate the effects of the mismatch. To be specifically, the original DNN is adapted with these acoustic observations of the unlabelled data, and the boosted posterior probabilities generated with the original DNN are used as training targets. With around 1000 hour unlabelled data used for adaptation, experiments results on a Mandarin voice search recognition task demonstrate the effectiveness of the proposed adaptation technique. Compared to the baseline, the adapted DNN achieve a 10% relative Character Error Rate (CER) reduction.