Based On Multiple Classifier Of Minority Language Recognition Research

China is a multi-ethnic of national unity, with science, technology and the society are progressing, communication technology in ethnic minority areas has gained popularity application. National language system, more dialogue national language information query system, many national language speech recognition system and national language spoken translation the application of the system demand also into rapid growth trend, and minority language recognition is part of these systems. Therefore, it is a complex, challenging and meaningful work to study, and well worth studying. This paper, based on the laboratory design telephone voice minority language recognition database, explore improve the method of minority language recognition system, emphatically resolve minority language recognition with Chinese loanwords influence of recognition rate. The paper mainly work include:1. Building the model based on GMM-UBM language recognition system frame, this as baseline system. In the feature extraction module, using a RASTA filtering, VAD and CMS techniques to improve robustness of characteristic parameters to noise and channel, and based on this, the SDC extracted as acoustic characteristic parameters. In the training modules, train out UBM based on the training method of MLE standards, and adaptive acoustic model of different languages from UBM. In the test module, at first, normalized the scores, with maximum likelihood as decision rules.2. In order to further improve the minority language recognition system performance, apply MMI standards to baseline system framework. The traditional MLE standards focuses on adjustment model parameters. MMI standards more emphasis on adjustment model of surface between classifications, the can better trained data classification. The test results show that the improved GMM-MMI system is better than the baseline system.3. Considering the minority language have the characteristics of the accent, according to the principle of classifier fusion, in characteristic layer extraction super sound segments of information (pitch frequency FO) as SDC characteristics complementary for better depict national language speech characteristics that contains Chinese loanwords. Pitch frequency FO as a support vector machine (SVM) input to training language model. In decision-making, we adopt GMM-MMI model and the SVM model to classification and calculate the output scores, then fusing the scores with linear score weighting.The experimental results show that the recognition rate of fusion systems is better than single use GMM-MMI system and SVM system. The rate up about18.00%and2.48%, and3s (including Chinese loanwords) statement recognition rate is up25.83%and3.7%respectively, and the influence of Chinese loanwords almost negligible. This shows that merged MMI-SVM system to reduce the Chinese loanwords on minority language recognition effects.