Research And Implementation Of Uyghur Speech Corpus Management Platform

With the development of Natural Language Processing technology,speech synthesis,speech recognition,speech translation and speaker recognition have also developed rapidly.With the development of Natural Language Processing technology,speech recognition,speech synthesis and speaker recognition have also developed rapidly.These techniques can not be separated from the speech corpus,and they all need large-scale,high quality the speech corpus for training and testing.In order to get high quality corpus,it is necessary to manage the corpus well,which requires the design of a uyghur speech corpus management platform.It's a speech corpus platform that can record,annotate and manage.It can enable uyghur related technology scholars to play,see,query,count related speech corpus,and download uyghur application tools.In order to solve the design gap of uyghur speech corpus management platform and the problem of traditional application software C/S architecture,an online uyghur speech corpus management platform is proposed.The main research content and results are as follows:1?With the knowledge of phonetics and acoustics,the phoneme,coding,phonological structure,syllable structure,rhythmic features and synergetic pronunciation of uyghur language are studied.A total of 32 phonemes in uyghur language are coded by Unicode,and the rules of phonological structure,syllable structure,accent rhythm,length rhythm and coarticulation are obtained.2?The corpus is standardized from 6 aspects,namely,speaker specification,data acquisition specification,data storage specification,corpus selection specification,corpus annotation specification and legal declaration.To pronounce the text for the design,including selection of text text,language acquisition,conversion,data conversion.The speech recording is designed,including the determination of the speaker and the voice acquisition.Praat software is used for the annotation of the speech library.Among them,crawler technology is used to get text.3?Through the Microsoft Visual Studio 2012 development tools,Asp.net web development,C # language,Microsoft SQL Server 2012 database,and the GridView control and SqlDataSource data binding to display the corpus on the page,the chart control to display statistical analysis and analysis diagram,config configuration Data connection and audio control to play,these technologies to achieve speech corpus management platform to add,delete,edit,query,play,export Excel table,view,download,upload,user permissions,statistical analysis and other functions.The management platform has been applied to the multilingual laboratory in Xinjiang.The speech syntheses 12000 speech corpus,and syntheses the annotation 6000.There are15000 speech recognition languages,including 3000 telephone speech corpus,6000 emotional speech corpus,3000 dialect speech corpus,3000 other ASR speech corpus,and 3000 identification marks.4?The uyghur language annotation platforms,the main function of task assignment,upload tasks,batch upload task,message management,audit content,conversion,and Latin uyghur broadcast speech corpus.The main technologies are JetBrains PhpStorm 10.0.1 editor,XAMPP server software,PHP language,MySQL database,and CI framework.The password is encrypted by the MD5 algorithm.The platform has been applied to the Xinjiang multilingual laboratory.18000 sentences are labeled online,of which 60(30 men and 30 women)and 300 per person.5?Uyghur recording software is implemented.The main functions are recording,audio files viewing,playing,renaming,deleting,querying the number of recordings,viewing help,and downloading tasks.The main technologies are Eclipse development software,Java language,SQLite database and Android platform.The recording software has been applied to the Xinjiang multilingual laboratory.20 people(10 men and 10 women)have recorded 105 declarative sentences,137 exclamations and 100 questions.The uyghur speech corpus management platform respectively from function,performance,and security of the pages were tested in performance testing using Google browser developer mode 5 features of requests,transferred,Finish,access platform of DOMContentLoaded and Load were tested.Compared with the traditionalC/S architecture,the platform has a friendly interface,complete functions,and a great improvement in the quality of speech and speech materials.The results of the platform test and operation show that the platform is more effective.The speech recognition and speech corpus collected are trained and tested on Kaldi with various models.The WER of the DNN model is 8.24%,and the speech recognition results are the best.