Recognition, Support Vector Machine-based Languages

Language is the main way of people to get information. It is also the most natural way to communicate with others. And it is the most convenient and effective communion tool. Now, we have step in an information society. With the computer science and technology development and the increasing global cooperation, language recognition has increasing importance in speech processing applications and can be used in multi-linguistic information services and security applications.From the seventies of the last century up to now,though it is just several decades,many kinds of ways of Language identification with their own characteristics have already come into being,most of which are not mature.This thesis is focus on the test-independent and speaker-independent language identification method. In order to improve the recognition effect we use support vector machine as a classifier to establish the appropriate language recognition system. The main works of the thesis are as follow:(1) In order to improve the recognition effect Language Recognition need to eliminate the individual differences of pronunciation in the same language, and try to find the different acoustic characteristics among the different languages. In the thesis, we extract six features subsets including MFCC, LPCC, fundamental frequency (F0), short-time energy (En), cadence and the first format (F1) for Language identification. And to find the most effective language recognition feature parameters.(2) Using the SVM as a classifier to design a single-feature and single-classifier language recognition system and a multi-feature and single-classifier system. In this experiment, the two systems identified five languages (Chinese, English, Japanese, Naxi and Bai) from the two different databases. We use five single-features and four multi-features as the input features of SVM. in three experiments, including male experiment, female experiment and male and female experiment. And we detailed analyze of the experimental results.(3) Based on the above work, we use multi-feature multi-classifier fusion concept to design and implement a new language recognition system. Use of four characteristic parameters and two support vector machined to structure eight classifiers, then use majority vote, weighted average method and decision-making template three methods for classifier fusion, given data set based on the results, and complete the experimental results.Experimental results show that discussed in this thesis six characteristic parameters, the pitch frequency has the best recognition performance; for mixing sound, a single feature the best single classifier recognition rate is 60.20%; multi-feature single-classifier of the most good recognition rate is 80.37%; and multi-feature multi-classifier fusion of the best identification is up to 90.27%.