Research On Language IDE NT Ification

Language identification is a technology of recognizing the spoken language and it can be widely used in multilingual information services and military security. Most of language identification system is based on continuous speech recognition and modeled on the phoneme information. It requires professional knowledge of language and a large number of prior knowledge, so it usually has a poor scalability.This paper studies the problem of language identification with text-independent, multi-language and telephone speech. We focuses on the modeling method of Gaussian mixture model and support vector machine. Besides, we have a deeply research on the feature parameters and new improved methods on GMM/UBM modeling and sequence kernel based SVM model. The main research work above is as follows:1. Built a GMM/UBM based language identification system and provided a new log-likelihood ratio calculation method to reduce the noise in the scoring process. The experiment showed that using likelihood ratio on the level of frame and mixture components had a better performance than traditional methods2. Provided GMM-UBM based sequence kernel transform method, since it was hard to directly applied SVM in language identification system because of such large sample and hard chosen of background language. The approach method has a better performance in the experiment. Besides, this paper promoted a new feature fusion method:according to Luradour sequence kernel, we can generate a new kind of high dimension feature as the input to SVM system, which can combine both the long-term and short-term feature.3. In addition, on the basis of the above two types of models This paper had a research on the relationship between the number of mixture components and the recognition precision.