Research On Language Identification For Telephone Speech

Language identification (or LID, for short) is a technology to identify which language the given speech belongs to. With the rapid development of mobiles, LID for telephone speech is becoming more and more important, especially in multilingual information service and military security.This paper focuses on LID for telephone speech. The main work can be summarized below:1.Applying some features using in speaker recognition to LID.To solve the problem of noise,making a study of features which outperform in speaker recognition, such as MVDR and GFCC and applying them to UBM-GMM.2.Building three baseline systems.The approaches mainly used in LID can be divided into three parts: UBM-GMM, PPRLM and SVM. This paper makes a study of them.In UBM-GMM, selecting boundary samples to make the model more accuracy and improving the score approach. Both of them can improve the performance of the system.In PPRLM, this paper builds a simple PRLM based on Chinese phoneme recognizer lack of labeled samples and analyze the interpolation of different background language models.In SVM,this paper mainly focuses on SVM based on GMM. The results show that GSV outperform UBM-GMM and the use of Gaussian scores.3.Researches on combination of different LID systems.To improve the accuracy of LID system further, do some research on combination of different LID systems to make full use of them and their complementary. The results show that the combined systems outperform the single system. What’s more, the combination of baselines perform better that shows the baselines has a strong complementary.