Hengyang Dialect Speech Recognition Research Based On HTK

Speech recognition is the simplest and straightforward method of human-computer interaction.It is a comprehensive discipline,which involves a series of disciplines such as linguistics,pattern recognition and artificial intelligence,and has a very wide application prospect.In recent years,with the development of information technology,the research of Chinese speech recognition has achieved some scientific results,and gradually applied to the actual product.However,in order to make speech recognition technology into real people's lives,there are still lots of problems.Due to Chinese culture,different regional has different speech within ten miles,so the research on speech recognition of regional dialect has become particularly important.This paper studies the Hengyang dialect in Hengyang,Hunan Province,and it establishes a continuous speech recognition system based on Hidden Markov Model(HMM).Briefly,this paper introduces the system structure and the method of speech recognition,deeply analyzed the acoustic foundation of speech,and studied the characteristics of Hengyang dialect.It is emphatically studied the basic principles and methods of feature parameter extraction,and deeply studied the hidden Markov model.First of all,through the study of the characteristics of Hengyang dialect,we can found that there is a big difference in pronunciation between Hengyang dialect and Mandarin,in order to establish a high-performance and high-quality voice recognition system,you need to carry out in-depth study of Hengyang dialect.In this paper,using the hidden Markov model toolbox(HTK3.4.1)to the basic phoneme recognition unit,using the linear predictive cepstral coefficients(LPCC)and Mel frequency cepstral coefficients(MFCC)feature extraction based on 5 states HMM model,build Hengyang dialect continuous speech recognition system.Design experiments we should compare the recognition performance of the system under different phonemes models,different characteristic parameters,and different Gaussian mixing numbers(Mix).The experimental results show that the system recognition performance to achieve the best when this experiment be combined with the tri-phones model,39-dimensional MFCC and 6 Gaussian mixed model numbers and HMM model.According to the basis and the test of system,and test results show that the system has a certain degree of adaptability in the actual environment,the recognition rate is better.