The Theory, Analysis And Experiments About The Speaker-Independent Continous Speech Recognition

Speaker-independent continuous mandarin figure speech recognition system has wide application in many fields such as telephone speech gateway, remote control of home appliances, industry control and information inquiry system. It has theoretical and practical importance in both few vocabulary and large vocabulary speech recognition systems. Owing to its characteristics such as shorter pronunciation, easier confusion and many more dialects, mandarin figure speech recognition is more difficult than English figure speech recognition and becomes a challenging topic in the speech recognition field.Hidden Markov Model (HMM) which is based on statistics becomes the main speech recognition technique now days because its strong capability in modeling the dynamic time sequence and flexibility in the selection of the parameters, structures and training techniques. However,HMM not only needs a great deal of samples to train its model which makes the sample collection and training a big work and has low usage rate of sample information, but also can't train the sample of low probability sufficiently which might lead to false recognition.This paper is mainly to study the figure syllable modeling for speaker-independent continuous mandarin figure speech. A new human-computer combined figure syllable feature extraction technique is proposed. The training set and the test set of figure syllable are extracted from the continuous figure speech. Guided by the high dimensional space covering theory the high dimensional space covering neural network for every class of figure syllable is constructed. The recognition rate of the test set reaches more than 97%. This model is applied to a new speaker-independent continuous figure speech recognition technique based on high dimensional space covering dynamic scanning theory.Experiments show the high dimensional space covering neural network outperforms the HMM when the number of samples is few. The high dimensional space covering neural network can describe the sample distribution in high dimensional feature space more reasonably and make better use the sample information.The high dimensional space covering neural network constructed by the former sample set and the false recognized samples has higher recognition rate than the former high dimensional space covering neural...