The Research Of Robust And Feature Extraction On Uyghur Speech Recognition

Through the study on Uyghur speech recognition system processes,from speech corpus resource three tone sub degree,improved speech enhancement spectrum subtraction algorithm and improved MFCC feature extraction method based on ICA three aspects to improve,together play an important role in the speech recognition system.The purpose of the experiment is to improve the performance of the system in different circumstances.(1)In the aspect of the binding of the corpus three,a large amount of spoken text is acquired from the common social forum of Uygur language, and the text is pre processed and screened.And then combined with the Uyghur own adhesion characteristics,an efficient method is proposed,to the selected corpus covering more tone phenomenon,greatly reducing the traditional methods of work and improve the recognition of the corpus finesse.(2)The disadvantages of the traditional speech enhancement spectrum subtraction and the improvement are pointed out.In the formula,the algebraic formula neglected in the traditional spectral subtraction is added,the enhancement effect of the improved spectral subtraction is more obvious.(3)In the feature extraction method is introduced in the ICA model and use efficient Fast ICA algorithm to estimate the unknown vectors in the ICA model,using first difference and second order difference of characteristics of final feature parameters are extracted.Finally,the three steps of the speech recognition system are combined.The average recognition rate under different SNR increased more than five percentage. These methods improve the robustness of the feature extraction parameters,and make a solid foundation for the further improvement and research of the acoustic model and.the language model.