The Research And Implementation Of Deep Learning Based Spoken Language Identification

Language is one of the most common ways people communicate in daily,and are an indispensable skill.In the process of globalization,people's barriers to language communication have become increasingly prominent.In this context,researchers need to achieve language recognition urgently,therefore,language recognition is one of the most important research topics in recent years.The existing language recognition system still has many problems,such as extracting the pure voice information in the complex background,having the language attribute information stripped out from the confusing language,and so on.therefore,language recognition is still have a long way to go.Language identification(LID)is based on the voice of the speaker's language to automatically distinguish,to find the type of language is a kind of biometric identification technology.Based on phoneme features or acoustic features have been shown to be very effective in representing language information.Although the machine learning can effectively improve the language recognition performance,the recognition rate is still not meet the requirements,especially for the tasks of short-term speech segment,the performance is still to be improved.In recent years,language recognition technology based on DNN has been a hot research in academia and industry due to DNN's widely used and well recognition effect.This topic is language recognition based on DNN,is committed to complete a well performance language recognition system.Mainly do the following work:1.Achiving language recognition system based on DNN.2.In this thesis,we propose a phoneme feature vector based on the acoustic features,that called the DBF feature,which is more effective than acoustic features or phoneme features represent a language.3.The traditional feature domain model is improved by using a method to extracting I-Vector by DBN training DNN.DBF is used instead of UBM in GMM model to obtain more accurate statistics,and improving recognition effectiveness.4.Test and analyze the system.Firstly,the DBF feature is compared with the SDC feature in language feature performance.The results show that the DBF feature has stronger ability to express the language feature,and the performance of the DBF featureis improved significantly in the short-term voice task and the long-term voice task and the confusing and dialect recognition task.