Research In Speech Recognition Based On Fuzzy Neural Network

In recent years, speech recognition is a rapid development technology, because of the important theoretical value and extensive potential application, it obtains people's more attention. Speech is a complex nonlinear process, and the limitation of speech recognition method which is based on linear system theory becomes more and more outstanding. With the research and application gradually deepens to artificial neural network, fuzzy logic, particle swarm optimization algorithm and other nonlinear theory, these theories have been started to apply independently or intermeshed in speech recognition region. Therefore, Starting from a typical speech recognition system, the basic principle of speech recognition was introduced in this paper, and several methods common used of feature extraction, pattern matching as well as training were discussed. Based on neural network and fuzzy logic theory, a speech recognition system of fuzzy neural network was constructed, at last particle group optimistic algorithm was introduced to precede speech recognition research.This paper introduced firstly the basic knowledge of fuzzy logic and neural network, then the network structure and algorithm of the standard T-S fuzzy neural network were presented. But if this model is applied in speech recognition directly, it would produce the problems of rule disaster and network ratiocination invalidation. So an improved T-S fuzzy neural network algorithm owing to four layer network structure was presented, and this network structure can be applied to speech recognition system by adding a compensating factor related to input dimension, then simulation experiment of speech recognition system using this algorithm is proceeded. The result showed that the improved T-S fuzzy neural network can be used in speech recognition system, at the same time, the recognition rate of this network is higher than RBF network, and the improved T-S fuzzy neural network owned better robustness. These explained fully that the fuzzy neural network had more strong classification capacity, but the deficiency was that training speed was slower, and it depended on initial value. This defect will be improved in future.Be similar to genetic algorithm, particle swarm optimization algorithm is also an iterative optimize instrument, it can search the global optimum by iterating. But the particle swarm without the inheritance operation such as cross and variation, decided the searching according to its speed, and particle has an important memory character. This paper constructed a model which based on particle swarm optimization algorithm to optimize fuzzy neural network initial value, first of all, some better parameters were primarily selected by particle swarm, then putting these parameters as the initial values of fuzzy neural network, at last, training the improved T-S fuzzy neural network algorithm, obtaining better results of speech recognition.