Study Of Speech Recognition For Mandarin Digit Based On Characteristics Of Hearing And Neural Networks

Mandarin Digit Speech Recognition is an important branch in the speech recognition field ,the application of which are telephone dialing,access control System,confirmation of ID cards ,etc. But Mandarin Digit Speech Recognition is more difficult because of the fewer syllables which lead to high degree of confusion.Based on the research of the common mandarin speech recognition technology, this thesis is to enhance system performance though improving them. From the generation model, it studies the various components of a whole system including pretreatment,the acoustic model training, recognition algorithm.The main work is bellow:Firstly, in the discussion and comparison of traditional features such as LPCC and MFCC, it extracts the new features on the basis of bark wavelet transform to improve the stability. Then a preliminary study on visual research of life-form to extract speech features PCNN is given and improved.Secondly, in the discussion on the traditional DTW algorithm, PNN based on the estimated density function and Bayesian decision theory is built as recognition model.Thirdly, Training model and algorithms are simulated by software. Training and testing database is established. And the HCI interface is created. Then real-time performance and accuracy of the system is tested and the main findings and conclusions are given.Finally, the further research and improvement direction of this subject is raised.