Research And Implementation On Language Recognition System Over Telephone Channel

Process of globalization leads to the increasing of international exchanges, differentcountries has different languages, which results in a communication barrier. With the demand ofmulti-lingual services becomes greater and greater, language recognition over telephone networkhas opened its historical arena. As the front end portion over the telephone network ofcross-language information processing, performance of language recognition system is mainlyreflected in the recognition performance and real-time performance. Recognition performance isthe assurance for the system to be exerted, which represents the reliability of it. While themultiple and real-time ability is the key point for the system to be put into practicality. Therefore,on the basis of increasing the recognition performance, this dissertation put more emphases onthe implementation of the multiple real-time language recognition system.This dissertation relies on the key project of the National863Program, whose purpose is todevelop the language recognition system which can process multi-access voices simultaneously.Combining the real-time and multi-access needs over the telephone network, this dissertationputs more emphases on the arithmetic and implementation of the language recognition. Based onthe increasing of the recognition performance, the implementation and design of languagerecognition system are carried out. The main work and achievements of this dissertation can beoutlined as follows:1. The key technologies of the language recognition system based on GSV-SVM are studied.A division method of non-speech signal is introduced, which is dead against the characteristic ofthe speech over telephone channels. According to the experimental results, the languagerecognition baseline system of this dissertation is eventually established.2. A channel compensation algorithm based on feature transform is proposed. Aiming atresolving the problems of channels diversity and redundant information of GSV over telephonechannel, related research and mainstream algorithms are presented and analyzed. Based on thethinking of I-vector’s feature sub-space analysis and feature transform, feature vectors oflanguage recognition are treated by some kind of feature transform. This algorithm preserved andenhanced the discriminability between different languages, while restrained the influence ofnoise in channel. The experiment result shows that this method could greatly improve therecognition performance.3. An effective training algorithm of anchor models is proposed. As speaker differencesinfluence the performance of language recognition system, researches of serving this problem areintroduced, among these methods, this dissertation focus on the anchor model algorithm. Then,aiming at the lack of space constructed on the anchor super-vector, we select the support vectorof different languages to reconstructed anchor super-vector by the information of support vector.This algorithm effectively restrains the impact of speaker differences in language recognitionsystem, and experimental results verify the efficiency of it.4. According to the demand for real-time over telecommunications network, a multiplereal-time language recognition system based on DSP and FPGA architecture is proposed. This dissertation focuses on the FPGA implementation of the system. A memory accessingmechanism is introduced to meet the need of frequently accessing the large amount of template.Then, the rationality of this design is verified through the performance testing and resourceanalysis. This system is able to fulfill the238-access real-time recognition tasks, which canprovide reliable indemnifications for multi-access real-time processing and accurateidentification of the language recognition system.