Research On Dialect Accent Speech Recognition Based On Articulatory Information

The rise of deep learning has led to the development of modern Automatic Speech Recognition,Deep neural network's advantage over traditional machine learning are mainly in several aspects.First,Deep neural network integration feature extraction in the process of training,but the feature extraction of the traditional machine learning model and the training of the model is independent of each other.Secondly,deep neural network is good at end-to-end learning and is strongly able to represent nonlinear feature.The wide application of speech recognition was standing in the basis of constantly improving the performance and robustness,which are the eternal themes on speech recognition research.These themes are based on different start points on the subject.On the basis of previous studies,this paper is standing in the perspective of phonetics combining with the characteristics of speech recognition and studying how to utilize deep neural network model to extract articulatory information and considering articulatory information on how to improve the performance of speech recognition,so as to better use of articulatory information in dialect accent speech recognition.Deep neural network has a powerful ability of representations of function,which is very suitable to serve as the model of feature extraction.In this respect,this paper utilize the acoustic-to-articulatory inversion method by adopting the idea of a joint training,and put forward an improved training method based on joint training of features extraction model.In this paper,a method of deep learning optimization is studied,and the knowledge distillation is applied in the acoustic model to make use of the articulatory information.The knowledge of the articulatory features learned on the large model is migrated to the smaller model.By using distillation method to train a few smaller models on the accent difference data set,and on the model of teacher training on the application of an improved articulatory information extraction method,when comparing the model the difference on the performance of speech recognition,we could abtain a good model of a better performance.