Research On Semi-supervised Learning Of Lvcsr System Based On Adaptive

At present, it is still far away from perfect for the recognition speed, the correctness rate, system robustness in Large Vocabulary Continuous Speech Recognition(LVCSR). Especially in acoustic model training, the common way to training acoustic models in LVCSR is to use manually labelled corpus. Considering high cost and time-consuming of obtaining manually labelled corpus, a new method using semi-supervised learning paradigm to train acoustic models is proposed, which is named semi-supervised learning of dividing phases.Semi-supervised learning of dividing phases needs a small number of manually labelled corpus to initialize models and then loads models to recognize a great number of unlabelled corpus to get the N-best results; finally, a new data selection strategy is used to analyze and synthesize the N-best results to get the optimal results for iterately retraining the models. Depending on the method, author constructs the system named Automatically Labeling Corpus System (ALCS).Semi-supervised learning of dividing phases includes more advantages: first , it selects a small number of words frenquency distribution consistency and representativeness of manually labelled corpus as initial training corpus to avoid high cost of labelling corpus; second, it constructs adaptive models, which can spontaneously adjust models according to work environment and recognizing data; third, it can label corpus in phone level; finally, as long as it gets the unlabelled corpus endless, it can learn continuously and then performance may be better.Experimental results show that the new semi-supervised learning method can significantly improve the performance of the acoustic model in LVCSR. The recognition rate on the testing set may be further increased by about 4.5%.