Research Of Speech Recognition Technology Based On HMM

Speech recognition technology is an important part of machine learning, which involves multiple disciplines content of the information processing, artificial intelligence and pattern recognition and so on. It has been widely used in the social life. Speech recognition allows machines to understand human language and the intent of the person and react accordingly, ultimately realizes the interactive communication between man and machine. This paper sets up a speech recognition simulation system based on Hidden Markov Models.This paper first analyzes the process of speech signal pretreatment, proposes dual-threshold method, which is the combination of short-time energy and short-time zero-crossing rate, to make speech signal endpoint detection. Compared to the single method endpoint detection, it can get more accurate speech segments, and establishes the foundation for the subsequent processing of the speech signal.Then two feature extraction method are discussed in details in this dissertation, including linear prediction Cepstral Coefficients (LPCC) and Mel Frequency Cepstral Coefficients (MFCC), and the paper analysis the first-order differential coefficient of MFCC, obtained that cepstral parameters combined with the first-order differential coefficient can make speech recognition rate be increased by about4%.Finally, this paper focuses on the HMM-based speech recognition algorithm, and a simulation of the complete speech recognition process is implemented in Matlab platform, which include creating the voice library, pre-process, training (parameter selection) and recognition. This paper creates a database which contains700pronunciations from10people, including Chinese digits0to9,"Zheng","Zhou","Da","Xue", a total of14samples elements, to provide data for the simulation. Then optimizes the practical application of HMM, including the selection of initial model and the logarithmic processing of the Viterbi algorithm, as well as the revaluation of parameters. By analyzing the test results, we conclude that the number of CHMM state is4, the training times is20, and selecting a mixture of48-dimensional LPCC and MFCC parameters as the characteristic parameters, the speech recognition system for the Chinese isolated word recognition rate can be up to90%. In the process of testing, the paper also analyzes the misrecognition problems to some sample elements caused by the complexity and particularity of the Chinese pronunciation, and proposes future research directions to solve this problem.