Chinese Dialects Identification Using Attention-Based Deep Neural Networks

With the progress of science and technology,intel igence has become the future of social development.Speech recognition is the core technology of intel igent information processing.After nearly a hundred years of development,it has made great progress,but the real effect of multi-language speech processing system is not satisfactory.The automatic identification of Chinese dialects are using computer to automatically identify class of the speech segment.It has significant application in information services,public safety,criminal investigation,language engineering and many other fields.The effective solution to the Chinese dialects identification is that extracting the generic features of the dialects using the machine learning algorithm firstly,then establishing the dialects class model and completing the dialect class decision by model matching final y.At present,there are still some problem,such as lack of database,traditional dialects features are fuzzy and unstable.In addition,the classification model has many problems such as lack of ability about complex signal learning,expression and classification.In this paper,the specific work of the above issues is as follows:(1)The expansion of Chinese dialects speech database.In the previous Chinese dialects database has been built.We continued to record the dialect voice,and actively seek the sharing of resources.We have added Wu,Min and Mandarin dialects nearly 100 hours.Secondly,the parameters of the dialect speech data are adjusted,the length of speech are automatically segmented,so that it can meet the experimental requirements and improve the reliability of the experimental results.(2)A method of Chinese dialects identification based on I-vector and rhythm feature fusion is proposed.We increased the global characterization of dialects from the syntactic structure,so as to improve the information of distinguishability.The experimental results show that compared with the traditional GMM-UBM model,the global feature fusion method can reduce the equal error rate of Chinese dialect by 56.32% and the effect is significant.(3)A method of Chinese dialect identification using attention-based deep neural network is proposed.The deep neural network can solve the problem of large input and complex nonlinear information.Compared with the common deep neural network,the deep neural network based on the attention mechanism reduced the relative EER by 28.3%,and has good stability.