Using MMI criterion to realize language identification of minority languages

In order to further improve the performance of our system of minority language identification, this paper proposes an approach to language identification of minority language with discriminative training of GMM models with Maximum Mutual Information (MMI) criterion to reduce the overlaps between target languages, increase the discrimination, and reduce the effect of Chinese loanwords to performance of the system. Firstly, we extracted Shifted Delta Cepstra (SDC) feature parameters, and construct the GMM-UBM models as our baseline. Then, GMM models of each language are training by MMI criterion. In the testing stage, three kinds of utterances from 7 languages with different durations are tested in two kinds of system. The experimental results show that discriminative training improves the system performance.