Research On Speech Recognition System Based On Neural Network And Wavelet

As the development and application of computer technology and improvement of intelligent.human expect that the machine can understand human language.It is precisely because such expectations and needs makes speech recognition technology rapid development and application. Recently,Speech Recognition is becoming received widespread attention because of very important theoretical research value and broad practical application prospects.now people do more research on it.In this paper,the basic theories of speech recognition are introduced.such as the framework and categories of recognition system,preprocessing,feature extracting training,recognition decision,threshold value comparc,and so on.In summary,Speech recognition technology mainly includes three aspects of feature extraction, pattern matching criteria, and model training techniques.It is to select a certain sound characteristics, then utilize a certain model algorithm to establish a unique library of templates,each template matching of the voice signal, and ultimately get the best recognition results.This paper analyzed aspects of the speech recognition process such as outpoint detection,the characteristic parameter extraction and recognition decision. The main content is probably as follows:1.This paper analyzes endpoint detection of speech recognition,and analyzes a fusion of Double Threshold Energy and Neural Network,which acts as new method for speech signal to detect endpoint.and the experiment shows that the proposed method is more effective and higher recognition rate than other common methods.2.This paper analyzes a variety of commonly used speech feature parameter coefficient, and its extraction algorithm.and studies the advantages and disadvantages of them.On the basis of the existing algorithms,an optimization algorithm is proposed that fuses the wavelet transform,the LPCC parameter and MFCC parameter.In Matlab,Simulation verify the reasonableness and effectiveness of the overall action of the new algorithm.3.This paper analyzes the recognition decision,most commonly used algorithm to identify decision-making is the Viterbi algorithm. However, by the impact of the various aspects, the result will appear error.In this article, with the recognition result obtained by the Viterbi algorithm, and then after a threshold comparison, the key that improves the speech recognition rate is to calculate reasonable threshold,under certain circumstances, the threshold is got through many experiments.4.Finally,this paper put the speech recognition into practice in wheeled robot platform and Android (Android system).The result of experiment shows that it is an efficient, stable, practical, high recognition rate of speech recognition systems.