Study On Speaker-Independent Isolated Words Speech Recognition System

As an interdisciplinary field, speech recognition is theoretically very valued. Speech recognition has become one of the important research fields and a mark of the development of science. Although speech technology has got some achievements, most speech recognition systems are still limited in lab and would have problems if migrated from lab which are much far from practicality. The ultimate reasons for restricting practicality can be classified to two kinds, one is precision for recognition and the other is complexity of the system.Chinese speech recognition technology and its implement have been studied in this paper confronted with problems in both theory and application. The main work and research fruits are as follows.Firstly, introduced endpoint detection method and analyzed the problems of speech signal processing and feature extraction. Then put forward a modified method, that a double restricted value with dynamic length of windows. At last discuss MFCC in details.Secondly, the construction of speech recognition system based on DTW/HMM and its application are studied. Then analyzed some practical problems in HMM, including initialized model choosing, scaling problem and so on.Thirdly, studied a discrete HMM for speaker-independent speech recognition system based on corrective training, and then expanded to continuous HMM field. In the end results of experiments show it overcomes the shortcomings of classical ML algorithm and improves the resolution of HMM system.Fourthly, create a database for training and testing which contains 528 pronunciations from 11 people. And all of the simulations are based on this database.Also the feasibility of mixed C++ with MATLAB has been explored.