Research On Robust Processing Technologies In GSV-SVM Based Language Identification System

Language Identification is the process of recognizing the language spoken from a sample ofspeech with a computer, and it has a wide range of applications. This paper studies the keytechniques of the Robust Language Identification. In this dissertation, we explore the possibilitiesto obtain high identification rate while enhancing the robustness of the LID system. They areexplored in model parameter estimation, GSV estimation and back-end score process. Severalnovel and efficient algorithms are proposed. The main work and innovations obtained in thispaper can be summarized as follows.(1) Joint factor analysis is a popular technique for coping with the training and testingenvironments mismatch in language recognition. But the language diagonal matrix and thesession variability need to be estimated jointly, consequently, a large amount of data of differentlanguages is needed, and its complexity is high. A simplified factor analysis model method isproposed. In this method, only the session variability is estimated, and the noise is removed withthe estimation obtained. After which MAP is used to estimate the variety between the languagemodel and the UBM model. Finally, the variety estimated is used for the training and testing ofthe SVM. Experiments show that the proposed method can effectively eliminate the sessionvariability and the efficiency is greatly improved while the performance is slightly declined.(2) Factor analysis is an effective technique for coping with the training and testingenvironments mismatch. But when the amount of training data is large, the update step sizevaries with the data length, which results in the imprecision of the description of the sessionvariability and language discriminative information. A speech length adaptation method forlanguage supervector estimation is presented. Firstly, the language diagonal matrix is estimated,with which the adaptive relevance factor is derived, which can represents the length information.Then the language supervector and session variability are jointly estimated. Finally, the languagesupervector is used for the training and testing of the SVM. Experiments show that the proposedmethod effectively eliminates the session variability and is robust to the data length variety, andthe system performance is improved.(3) Current i-vector factor analysis method only model the total variability subspace, so thei-vector for the representative of the language information contains information about thechannel. A score processing method based on subspace mapping is given. Firstly, the i-vector isextracted from the data, with which the NAP model and multiple SVM model are estimated.Then the channel information is removed with NAP on multiple SVM model. Finally, the scorevector is derived to train the SVM. Experiments show that the proposed method effectivelyeliminates the noise, and the system performance is improved.