Research On The Speaker-Independent Speech Recognition System

With the development of the information technology, the technology of speech recognition has been gradually applied to the stage from research into practicality, but how to improve the recognition rate of speaker-independent system is still one of the hot spots in the current speech recognition technology. Various techniques of speaker-independent speech recognition system are studied systemically in the thesis, meanwhile a speech recognition system based on speaker-independent isolated words is completed on the MATLAB.The speech recognition system, the speech signal preprocessing, endpoint detection of the starting and ending points and the extraction of feature parameters, are presented in detail. On this basis, three common algorithms of speech recognition consisting of Dynamic Time Warping(DTW), hidden Markov model(HMM) and neural network model(ANN) are analyzed. The HMM is mainly studied and some effective measures are adopted to solve data overflow existing in the HMM. Then, the filtering algorithm based on Empirical Mode Decomposition(EMD) with differential entropy of Singular Value Decomposition(SVD) and speech endpoint detection based on improved Hilbert-Huang Transform(HHT) are proposed for low SNR noisy environment, and the results of traditional algorithms are analyzed and compared.Meanwhile, the speech recognition system with speaker-independent based on HMM is completed on the MATLAB. The result shows that the recognition rate of the system is improved by the improved algorithm after comparing with the traditional methods of filtering and endpoint detection methods. Finally, a GUI of the speech recognition system on the MATLAB platform is designed with different interfaces and processes of real time speech recognition. The experiment of speech recognition based on the speech library is completed to verify the validity of these methods.