In view of the low degree of discrimination of Mongolian speaker features, the problem of learning too much non-speaker information during the modeling process of Mongolian acoustic model, thereby reducing the recognition rate of the Mongolian speaker adaptive speech recognition system. the speaker adaptive model learns too much unnecessary speaker information, which interferes with the speaker adaptive modeling, thereby reducing the recognition rate of the Mongolian speaker adaptive speech recognition system. In this regard, this paper proposes a Discriminative Feature Transformation algorithm, constructs a Mongolian speaker feature transformation model, and gives a model training algorithm. The discriminative feature transformation method is used to perform discriminative feature transformation on the extracted Mongolian speaker features to ensure that the dispersion between humans is maximized and the dispersion within humans is minimized, and a more discriminatory Mongolian speaker feature is generated. In the design of the Mongolian speaker feature transformation model structure, the restriction on the internal variance of the speaking human is added, and the distinguishing information of the Mongolian speaker in the Mongolian speech data is concentrated in the feature extraction layer. This paper conducts experiments on the Mongolian corpus, and uses two evaluation indicators, Word Error Rate and Speaker Feature Discrimination, to evaluate the results of the experiment. The experiment shows that the proposed method can extract the distinguishing speaker features of Mongolian language to provide speaker information for Mongolian speaker adaptive speech recognition modeling.

As a hybrid modeling technology in speech recognition, DNN-HMM is composed of deep neural networks and hidden Markov models. In the process of using the Mongolian corpus to construct the DNN-HMM acoustic model, in order to study the influence of the DNN-HMM structure on the Mongolian acoustic modeling and the relationship between the size of the Mongolian corpus and the DNN-HMM acoustic model structure, the DNN-HMM acoustic model was designed For the structure of DNN in the model, four DNN-HMM acoustic models of Rectangle DNN-HMM, Trapezoid DNN-HMM, Polygon DNN-HMM and Hourglass DNN-HMM are proposed. Experiments are carried out on the basis of the Kaldi experimental platform, phonemes are selected as the modeling unit, and three-scale Mongolian corpora are used to construct four-structure DNN-HMM acoustic models. The experimental results of the depth structure and the width structure show that the Polygon DNN - HMM structure with a depth of 6 layers is suitable for Mongolian acoustic model modeling. As the corpus increases, the width of the acoustic model should be appropriately increased so that each layer of the acoustic model can be Learn more abundant features and improve the accuracy of speech recognition.