Study On Small Vocabulary And Non-specific Person Of Speech Recognition System

Speech recognition is to make the computer “understand” the human language，andunderstand the contents of the language so it can perform a specific command or task.As the basis of human-computer interaction technology，the ultimate goal of Speechrecognition is to achieve natural communication between human and machine. Theapplication fields of speech recognition are very extensive, within all of these，non-specific，small-vocabulary speech recognition system is an important branch in theapplication fields of speech recognition.This paper makes in-depth study in non-specific， small-vocabulary speechrecognition system. And on the basis of this knowledge，The software of a non-specific，small-vocabulary speech recognition system is written. The major work is described asbelow:(1)Through analysis of the shortcomings of the traditional endpoint detectionalgorithm. An improved endpoint detection algorithm called Adaptive dual-thresholdspeech endpoint detection algorithm is brought forth, and the experiments show that thealgorithm has better performance. And then analyze three different characteristicparameters of speech recognition：LPCC、MFCC、PNCC.(2)Study the current three mainstream speech recognition algorithms: DTW, HMMand SVM, analyze their principles, characteristics and realization process， thelimitations and shortcomings of the three algorithms. Then construct the two smallspeech recognition systems. And analyze how to select the parameters of the HMMmodel through the experiment. Then compare the scope of applicaition the DTW andHMM algorithm.(3) Analyze the Support Vector Machine(SVM) algorithm which is based on thestatistical learning theory. Then introduce the basic principle of SVM and the processhow to use it to realize the speech recognition. After introducing the different kernelfunction of SVM, the RBF kernel function is choosed. Under the condition of the samekernel function，the punish factor and the kernel function parameters can indirectlyinfluence the recognition consequence of the speech recognition system.(4) Because the selection of punish factor and the kernel function parameters havethe great influence on the speech recognition effect. But so far，it has not put forwardthe scientific method to choose. The traditional method is usually selected by repeating the experiment or the human experience. This paper has conducted the preliminarystudy， puts forward PSO—SVM algorithm which using the Particle SwarmOptimization(PSO) to select the punish factor and the kernel function parameters ofSVM. Then realize the PSO—SVM speech recognition system in the form of GUI usingthe GUIDE of MATLAB. Through the results of the experiment, the effectiveness of thenew algorithm is verified. And it also proves if under the condition of the limitedsample，the recognition performance of this system is better than HMM algorithm andSVM algorithm which don’t use the optimization algorithm.(5)To consider the hardware conditions and algorithm characteristics， DTWalgorithm is selected as the hardware speech recognition algorithm.And complete asmall embedded speech recognition system which based on SPCE061A.Give aintroduction to the hardware components and software components.and it proved agood result through testing.