Research Of Speech Recognition Based On Mixture Feature Extraction And Improved Continuous Hidden Markov Model

With the development of science and technology，speech recognition products haswalked away from the lab, and come into every aspect of people’s daily life. Someproblems in speech recognition has not been completely resolved, especially the studyon the acoustic model. Input data for acoustic model is the characteristic parameters ofspeech, but Mel frequency Cepstrum coefficient can not accurately and completelyrepresent all the useful information in speech signal, especially in Chinese, thus affectsthe accuracy of the acoustic model, At the same time in the acoustic model the problemof local optimization is not be solved, thus causes the product performance are difficultto meet the requirements of ideal for use in speech recognition.This paper introduces the research status of feature extraction and acousticmodeling in the speech recognition, then analyses and compares kinds of algorithmsthrough advantages and disadvantages, Aiming at the problem of the mainstreamalgorithms, new methods are proposed. In the feature extraction, the mixed parameterbased on Fisher ratio is proposed. In the acoustic model training, the paper proposes thenew method of models parameter initialization based on distance and density. Theprimary research includes:①In the feature extraction,Aiming at the low identification precision of MFCCparameters in high frequency signals the method of extracting features based onMFCC、IMFCC and MidMFCC, combined with Fisher criterion was adopted.②In the acoustic models, the traditional approach of parameter initialization ofhidden Markov model can lead to the problem of local optimization, the paper proposesa new approach of models parameter initialization based on characteristics of speechtraining data, thus to optimize the final model.