Research And Implementation On Classification Algorithm Of Language Recognition System Based On Anchor Model

With the arrival of the globalization era, the information exchanges between various languages in the telephone network become more and more frequently. As the front-end processing subsystem of the cross-language telephone network, the performance of the language recognition system will directly affect the efficiency of the telephone network's communications. In which, how to solve the real-time processing problem when multi-access voices arrive at the same time is the key for it to be put into practicality.This dissertation relies on a key project of the National 863 Program, whose purpose is to develop the language recognition system which can process multi-access voices real-timely, combining with the practical application demands and based on the analysis and comparison results of current mainstream language recognition systems, the Gaussian Mixture Model Super Vector-Support Vector Machine(GSV-SVM) language recognition system based on anchor model is adopted as the research target, the research of its back-end classification algorithm is carried out and this algorithm is put into FPGA implementation, which provide a reliable back-end classification indemnification for real-time processing multi-access voice. The main work and achievements of this dissertation are outlined as follows:1. The key technologies of the language recognition system based on GSV-SVM are studied, and an equivalent language recognition system is proposed, which is based on the characteristics of the SVM's kernel function in it. We serve this equivalent system as our baseline language recognition system, and the output format and the two classification strategies of the SVM is analyzed, we adopt the one against one classification strategy of the SVM's probability modeling algorithm-Pairwise Coupling is treated as the baseline system's back-end classification algorithm, which is based on the results of the experiments and the conclusion of relative papers;2. A discriminative language training algorithm of anchor model is proposed. Based on the introduction of the anchor model into the baseline system and combining with the idea of the fast discriminative training algorithm, this algorithm build a more discriminative anchor space, which not only suppress the interference of the speaker information, but also reduce the GSV's dimension, which can reduce the SVM training time. The experiment results show that the introduction of this training algorithm can effectively improve the recognition performance of the GSV-SVM language recognition system based on anchor model;3. The two commonly used SVM probability modeling algorithm based on one against one classification strategy is analyzed and compared, the classification performance and the degree of the difficult of the FPGA implementation is considered compromisely, a fast probability modeling algorithm of the SVM is introduced into the GSV-SVM language recognition system based on anchor model. The experiment results indicate that this algorithm insure the recognition performance of the system and can simplify the operations of the posterior probability output at the same time, which laid the theoretical foundation for the FPGA implementation of the back-end classification algorithm of the language recognition system;4. Combined with the two proposed algorithm above, a normalized anchor space projection optimization algorithm is proposed for reducing the computation, storage and the difficult of the FPGA implementation. Based on this optimization algorithm, the back-end classifier is designed and implemented based on FPGA, which is tested and verified via four aspects of real-time, resource utilization, accuracy and identification performance. The results indicate that the back-end classifier is able to fulfill the 842-access real-time recognition tasks and the performance between the FPGA platform and the VC++6.0 platform is almost the same, which can provide reliable indemnifications for multi-access real-time processing and accurate identification of the language recognition system.