| In recent years,domestic cities have successively opened government service hotlines,which have now become the most important platform for interaction between the government and the people.Faced with the increasing telephone traffic,the system has some problems,such as the drop of connection rate and the untimely response to demands.With the development of artificial intelligence and the maturity of speech recognition technology,dialect recognition has achieved some research results.Taking Handan City of Hebei Province as an example,this thesis designs and implements a new generation of government service hotline system based on speech recognition technology.The main research contents of this thesis are as follows:(1)Design and build Handan dialect phonetic corpus.This thesis analyzes the structure and pronunciation rules of Handan dialect,designs a text corpus,collects 80 permanent residents of Handan to record audio,cleans,screens,checks and labels the audio,and finally forms a 31.4-hour Handan dialect speech corpus,which consists of a training set,a test set and a verification set according to the ratio of 10:1:1 for model training.(2)Aiming at Handan dialect recognition,the end-to-end speech recognition method is used to build an acoustic model based on convolution neural network(Convolutional Neural Network)operation criterion of connection temporal classification(CTC).Overlapping layers of one-dimensional convolution networks,training to get the benchmark model and optimizing it.In order to reduce gradient dispersion,the network GLU activation with fused gated convolution unit is used to accelerate convergence.To prevent over-fitting of low-resource language training model,Dropout layer is added to convolution module to optimize acoustic model.Aiming at the problem of poor performance of low-resource model,data enhancement,adding RNN-based external language model,and enhancing the recognition rate and generalization of the model are proposed,which are verified by experiments.(3)And design and implement a new generation of government service hotline system based on speech recognition.The speech recognition module adopts B/S structure,which is realized by Python language and Vue framework.The My SQL storage system is deployed,and the recognized words are saved to the database to automatically generate work orders.The background of the government hotline system is implemented by using Java language and Spring Boot framework.The system mainly realizes two working modes: first,based on voice recognition,real-time online recognition of citizens’ voices as characters is realized,and work orders are automatically generated;Second,the operator manually generates the work order,and the appeal desk sends the order to the organizer. |
