The Establishment And Application Of Uyghur Speech Corpus Based On Online

The establishment of a rich speech corpus is the basis of the study of voice technology,from the engineering perspective,the speech corpus is one of the important links to sophistication voice technology.The study of speech recognition is inseparable from the support of large-scale high-quality corpus.In view of the existing Uighur speech corpus,its research status still based on the pronunciation corpus of young people.In the face to the need of in-depth research and development,the Uighur speech corpus needs to be expanded,especially the diversification of pronunciation data needs to be improv to the practical application.Therefore,this paper studies the establishment of Uyghur speech corpus and the application of corpus in speech recognition.The main contents of this paper are as follows:1)The traditional speech data collection and annotation of speech corpus has been improved.collection of speech corpus is requiring a lot of manpower and time.In order to solve this problem,This paper researched the speech data collection software on mobile devices,which makes the collection of speech data speed up several times.Anyone can use this platform for speech data collection.In order to predict the validity of speech annotation,this paper also implemented the annotation platform.These new methods have achieved very good efficiency in practical application.2)According to the speech characteristics of Uygur language,Construct a diverse and large-scale speech corpus.Each dialect,according to the different regions also contains some native language.Some of the same words are different in different dialects.In addition,the pronunciation characteristics and rhythm characteristics of different age have certain difference.However,the collection of dialect speech in various regions and various of natural persons who with different ages and education degree,Has a certain research significance in optimization of Characteristic parameters and acoustic models for speech recognition.The key factor affecting the improvement of recognition rate is the variability of speech.establishment of corpus containing the most linguistic phenomena is essential for the analysis and identification of speech.to improving the quality of speech corpus is,select text data which covers more language phenomenon.In this paper we use Two screening methods which Common words include degree and triphone include degree used to Screening work.The usual random screening method was done for comparative experiments and screening by triphone method coverage reached 91%.3)Finally,The HMM and DNN,which are widely used in speech recognition technology has been used for acoustic features,acoustic model training and continuous speech recognition experiment are implemented for this speech corpus.the N-gram speech model is used in the experiment.In the Linux environment with Kaldi Toolbox experiments were carried out,the experimental results show that for large-scale speech data,DNN acoustic model for speech recognition results are even better.In this paper,the recognition rate based on DNN model is 84.49 %%.Compared with the traditional model,the recognition rate of the system is improved about 1.77%.