Research On Chinese Dialect Recognition Based On Attention And Transfer Learning

With the increasing popularity of Putonghua,Chinese dialects are gradually assimilated.However,as one style of intangible cultural heritage,the importance of Chinese dialects can not be obliterated.Dialect plays a role of living fossil in culture,which has precious historical and linguistic research value.As a special language variety,it has always been a hot topic in sociolinguistics.It is not only a linguistic phenomenon,but also a social and cultural phenomenon.Therefore,the protection and research of dialects has important practical significance.Dialect discrimination is a key step in the intelligent processing of dialects.As deep learning has achieved good performance in various tasks of natural language processing,deep learning technology is also widely used in dialect discrimination.However,the existing representative methods pay more attention to the automatic extraction of all kinds of speech level or text level features,pay attention to the underlying acoustic features,and do not consider the meaning of the pronunciation itself.Moreover,it is rare to use the transfer learning model for low resource dialect recognition.At the same time,the effective integration of attention mechanism is also insufficient.In view of the shortcomings of the existing methods,the following work is carried out in this paper:(1)Proposes an end-to-end dialect discrimination model based on multi head self attention mechanism.The model consists of two parts.One is to extract the unique phoneme sequence information of different dialects to form the pronunciation features,usinga residual CNN and multi head self attention mechanism;the other is to recognize the dialect types,using self attention mechanism and bidirectional LSTM based on the extracted pronunciation features.The experimental results on i FLYTEK dialect benchmark corpus show that the performance of dialect type recognition has been greatly improved.(2)Presents a model of dialect discrimination based on transfer learning and data augmentation.In this model,a relatively large-scale dialect corpus is used to train a dialect speech recognition model at the source end to obtain global knowledge,and the generated intermediate semantic representation is used to identify low resource dialects at the target end.At the same time,the data enhancement strategy is used to expand the low resource dialect data,and the self attention mechanism is integrated effectively.The experimental results on the international benchmark dialect corpus verify the effectiveness of the model.(3)Designs a dialect intelligent speech dialogue platform,which effectively integrates the dialect speech recognition model and dialect language recognition model studied in this paper,and integrates the existing multi round dialogue model and speech synthesis technology.