The Design Of Isolated Chinese Word Speech Recognition System

Since all kinds of machines play an important role in our produce and life, it’ssignificant to improve the relationship between human and machine. As ahuman-computer interaction media with good experience, the research of speechrecognition has important practical significance.In this paper, a brief introduction of the speech recognition’s theory andtechnology is done, and a detailed discussion and analysis of the main methods isdone, especially the hidden Markov model which is the mainstream developmentdirection. Based on the discussion and research, a design of the isolated Chinese wordrecognition system based on hmm model is proposed. Then complete the software ofsmall vocabulary speaker-independent isolated recognition system.The following work is completed in this thesis:1. Introduce some concepts and methods involving in the speech recognitionsystem. Make a new method of combining power spectral entropy and short-timeaverage magnitude in the part of endpoint detection. Use Mel-frequency cepstrumcoefficients including its dynamic difference coefficients to form the featureparameters vector with thirty-nine dimensions.2. Introduce the basic thought of hmm model and its application in speechrecognition system in detail and derived the detailed formula involved in the systemmain algorithms. Based on the analysis, describe the training of discrete hmm modeland continuous hmm model, including the selection and setting of parameters. Tosave some time, some improvements are done to the model training.3. Complete the software of speaker-independent isolated word speechrecognition programmed in C language on the Windows system and build a speechdatabase with audio files and record messages of ten vocabularies.4. Complete the system’s training and testing base on speech database, verify theresearch about the key technology of system and point the direction of improving thesystem’s performance.