A new method for language recognition based on improved GMM

Abstract

The automatic recognition of speech language is called diagnosis of language by signals. These systems often make decision by comparing the privilege of speech signal dependence to various languages. A new method has been applied to improve the results of the automatic recognition of language in this article that acts based on the improved Gaussian Mixture Model (GMM) model. A GMM model is trained by non-overlapping data through the method by applying SDC selective feature vectors. By comparing between this method and the usual GMM fulfilled to diagnose the 4 languages, we have achieved an average improvement of 4.4% in determining language recognition accuracy. At the end of this article the neural network (NN) method has been used as a back-end processing (BEP) method. By applying BEP, recognition error has reduced to 10% in the usual GMM, and to 2.5% in the improved GMM.