Research On Speech Recognition Algorithm Of Speaker-Independent Isolated Words

In recent years,speech recognition has gradually entered the daily life of people from the laboratory.Even more,the speech recognition technology of isolated words has widely been introduced into many varieties of fields.However,isolated speech recognition technology still suffers from many practical problems,such as real-time recognition.In this thesis,the speaker independent isolated speech recognition is taken as the object of study.By further studying and improving the existed commonly used isolated speech recognition algorithms,the time consumption of isolated speech recognition has been greatly reduced.The detailed work of the thesis is as follows:1.In order to solve the problem that DTW algorithm collects isolated word templates slowly,this thesis proposes an improved method for the traditional templates clustering algorithm.The method directly selects several appropriate initial template vectors for the isolated speech template training set,so as to avoid the set division and the increasement of template vectors during templates collecting with traditional DTW algorithm.The experimental results show that improved templates clustering algorithm effectively reduce the template training time.2.In order to increase speech recognition speed of isolated words in DTW algorithm,two improved schemes are proposed.In the first improvement,the global matching path in DTW algorithm is limited in a dynamic parallelogram instead of a static one.In the second,the matching style between voice feature vector for testing and template vector changes from full-length continuous matching to partial-length continuous matching.The experimental results show that the proposed scheme can improve the efficiency of isolated speech recognition while keeping the recognition rate unchanged.3.In order to increase speech recognition speed of isolated words with speech recognition algorithm based on HMM-GMM,this thesis proposes an improved scheme.Firstly the scheme cuts off the initial parts of the speech feature vector and cultivates corresponding probability value in all HMM-GMMs.Then reserve parts of the HMM-GMMs where probability values are relatively larger.In this way,repeatedly reserve HMM-GMMs until models' number reaches only one.The simulation results show that the improved algorithm based on HMM-GMM can significantly reduce the recognition time of isolated speech recognition while keeping the recognition rate unchanged.