A Hybrid Approach For Speech Recognition Based On Combination Of HMM And RBF Neural Network

Speech recognition is a comprehensive technology involved such areas as acoustics, phonetics, linguistics, computer science, information processing and artificial intelligence, which is used widely in such man-machine correspondence aeras as military affairs command, product inspection and date input.The research of speech recognition technology has been focused by the world for a long time. However, the most accurate speech recognition systems are still slow, expensive and inconvenient, which limit its speed, hardware implementation and application.Firstly, the deficiency and virtue of the hidden markov mode (HMM) and neural network(NN) are analyzed in the article.The mandarin digit recognition model based on the CDHMM & RBFNN and rapid trainning algorithms of estimating the state based on RBF are proposed. The observation symbol probability distribution is computed by NN-group composed of many simple RBF in the tranning phase. The average vector sequence of HMM is used to match the unknown utterance in the recognition phase.The minimum distance HMM is selected to normalize out the unknown utterance. And then the uneasy discriminatial recognition results output from the HMM are accurately recognition by the different measure of the RBF.Then, the HMMNN is simulated with the software of MATLAB 6. 5. The influence on the recognition results of some factors as feature parameter, number of training samples,background noise are discussed. The isolated and connected mandarin digit recognition for independent speaker used the CDHMM and HMMNN are tested respectively. Compared with the CDHMM, the HMMNN has the capability of faster learning ,higher recognition rate and potential advantages in automatic speech recognition application. The results show the validity of the new HMMNN applicated in the speech recognition engineering field .