Research On Madarin Continuous Speech Recognition And Application In Mobile Robots With It

Mandarin continuous speech recognition has been done for more than 10 years. Although some achievements have been obtained, many significant and difficult problems are not yet solved. Firstly, the context-dependent acoustic modeling must be paid to more attention and efforts to further improve its robustness and accuracy, especially to Mandarin triphone modeling. Secondly, because of the different channel and yawp or speaker's reasons, system recognition rate was depressed. Then it needs us to lucubrate on the research of adaptation. Finally, we also need study the portability of the technologies to shorten the cost of time with the research in new areas. My thesis is mainly to solve the above problems.Firstly, we study the context effects on Mandarin speech recognition and the decision tree based triphone acoustic modeling. We discuss our faced problems in the decision tree based Mandarin triphone modeling, including the selection of Mandarin base phone units, the criterion to design the context-related questions, and the complexity optimization on decision tree. The thesis advance new base phone units which add 6 initial/ final to standard initial/ final sets. The experiment of recognition system on HTK results show that the performance of new system is improved so much.The thesis compares Maximum A Posteriority algorithm with Maximum Likelyhood Linear Regression algorithm in speaker adaptation module and advances a better method which combines the stringendo MAP and fast MLLR, and adapt with new data on real time. By the results of experiment, the optimized adaptation algorithm is better than old ones.Finally, write a application under Microsoft Visual Studio.NET by using ATK toolkit, then save the recognizing results to a variable which used to control the thread direction for a robot. Experimentation of the navigation system make out that this recognition system is successful and recognizing rate reach 85%.