Research Of Time-Ordered Pattern Recognition Based On ANN/HMM

In Human-Machine Interaction, handwriting input and speech input are two methods which are the most naturally and accustomed ones for human beings. Handwriting and speech recognition are two important research objects in multi-mode Human-Machine Interaction technology. The two objects both have the characteristic of time-ordered. In this paper, we expect to find a new method which can be a common one for time-ordered patterns, based on Artificial Neural Network (ANN) and Hidden Markov Model (HMM). ANN has some features of anti-noise, auto-adaptation, strong capability of learning and high speed, and HMM is capable to deal with time series. So in this paper, HMM is used as the model of the whole pattern to be recognized, and simulates the transition between the states of the time series. ANN is a state emission probability estimator for HMM. It simulates the states of HMM, and is used as primitive model of the pattern to be recognized.In addition, considering the problems of simple structure and weak capability of auto-adaptation, we proposed an auto-split-and-merge method to determine the state number of a model. In this method, states are automatically added or deleted on a proper position according to the training data. We split the states with low modeling precision, delete the redundant ones, and finally achieve a balance. Taking handwriting symbols and speech commands for example, we compare the modeling effects between this method and traditional ones. The results show that, this advanced method can improve the modeling precision and save 25% of system resource.In order to put the research achievement into use, we developed a simple multi-mode Human-Machine Interaction system. In this system, we can write symbols or speak to give orders to the computer in a more natural way. In addition, it has the characteristics of simple structure, short response time and high recognition rate. We can obtain a recognition rate of 98 % for handwriting symbols, and 83.6% for speech commands. This is enough for common use.