The Research On Speech Recognition Based On Hidden Markov Model In Noisy Environment

Using speech signal as research object, speech recognition is an important research direction of speech signal processing and it is an embranchment of pattern recognition, too. It involves linguistics, computer science, signal processing, physiology and psychology etc, and even relates to body language. The final goal of speech recognition is to realize the natural communication between human and machine. Speech recognition has a wide application future. It has made a full application in dictation machine, telephone inquiry system and home application control etc. However in the progress of application and industrialization, many difficult problems has come out. Especially the problem of background noise is almost unavoidable. The problem has to solve is how to separate original speech signal from background noise, which can improve the adaptability of recognition system. Therefore the research on speech recognition is significant.Aiming at the difficult problem of background noise, the thesis is just about some aspects of the recognition system and some relevant algorithms are put forward which are also validated. First we introduce some basic concept and theory about speech recognition system. Then according to the algorithms of the endpoint detection for pretreatment, feature extraction of speech recognition , pattern matching and model training,we have a deep study on them. With the decrease of SNR, the algorithm based on Wavlet Transform for noisy speech endpoint detection keeps relatively ideal effect. Using Mel frequency cepstrum coefficient as feature parameter, it has better denoise robustness. Aiming at the problem of how to overcome the localization of HMM about hypothesis of independence, we introduce the method of ANN, with which HMM is combined. And we propose a hybrid model HMM/SOFMNN, which improve the performance of speech recognition system.The experimentations proved thar the correct recognition ratio of mandarin digit speeker independent speech recognition system which we build with above algorithms can reach 94.5% in condition of SNR is12db. And the system has a good performance with noise.