Research Of Kernel Function Of Support Vector Machine And Its Application In Speech Recognition

With the development of computer technology, artificial intelligence becomes a research focus, and has a broad application potential. Speech recognition technology is one of the most frontier technologies of artificial intelligence. Its goal is to make computers understand human language, to achieve human-computer communication, in order to operate more naturally and conveniently.Support vector machine (SVM) is a new pattern recognition method rising from1990s. It can solve some problems in practice, for instance problem with small sample size, non-linear problem, high dimensions, and local minimum points. SVM possesses better generalization performance and classification accuracy than Hidden Markov Model(HMM) and Artificial Neural Networks(ANN); therefore it is suitable for speech recognition.One significant characteristic of support vector machine is the introduction of kernel function, which maps input data in low-dimension space into high-dimension space. Support vector machine classifier is fully described by kernel function and training datasets. The choice of kernel function and the value of its parameter directly affect the performance of SVM classifiers.First through theoretical and experimental analysis, this paper studies the effect of different kernel functions and different parameter values to speech recognition results. It is proven that Gaussian kernel SVM could obtain best classification accuracy in speech recognition system. The best parameter of Gaussian kernel and penalty parameter of SVM are selected during the experiments. Furthermore, this paper applies two kinds of improved Gaussian kernel into speech recognition system, UKF kernel function and Corrected-Gaussian kernel function.Wavelet transform can carry multi-dimensional analysis on speech signal. Multi-dimensional analysis reduces the correlation among the signal components after wavelet transform, and the auto-correlative matrix represents sparse ribbon-shaped distribution. In this paper, combining wavelet technology and SVM kernel function, Morlet wavelet kernel and Mexico wavelet kernel are constructed and applied to speech recognition system.The experiment results show that UKF kernel, Corrected-Gaussian kernel and Mexico kernel perform better recognition accuracy than Gaussian kernel, and do not increase the space or time complexity of the SVM algorithm.