Methods to Improve Gaussian Mixture Model for Language Identification
Yonghua Xu, Jian Yang, Chen Jiang · 2010
Multiple-language phone recognition and n-gram language modeling produce the best performance in formal language identification (LID) evaluations, but this method needs many kinds of data that were phonetically transcribed utterances. This phone-based system isn't easily to apply for dialects or minority languages identification, because it is difficulty to obtain kinds of utterances which were orthographically or phonetically transcribed. Gaussian mixture model with the Universal Background Model (GMM-UBM), which has been successfully employed in speaker verification, is an effective approach to solve this problem. The GMM-UBM LID system, which didn't obtain utterances that were orthographically or phonetically transcribed, reduces the time requirement for both training and testing. In this paper we proposed a new combination method utilizing Output Scores Fusion techniques for acoustic scores and language model scores to improve the performance of the GMM-UBM based LID system. Experiment results show that the combination method as described above is another efficient method for language identification problems.