Application Research Of Support Vector Machines Sample Pre-selection In Speech Recognition

Speech recognition is the key technology of human-computer interaction which has a large and broad application space in the industrial production and social life. Therefore, the speech recognition system is a considerable theoretical and practical value.This article first briefly introduce the basic concepts and principles of the speech recognition system, and analyze the advantages of traditional speech recognition technology and its limitations. Second section describes the paper studies machine learning methods:support vector machine. This part introduces the support vector machine based on statistical learning theory, and support vector machines more commonly used multi-class classification method. This article is designed that the parameter optimization of kernel function will make the additional computation, which results to waste a lot of time, and support vector machines in the training process, a lot of time is wasted in the complex calculations on non-support vectors. Especially for Large-scale speech recognition systems, training time of support vector machine on unnecessary overhead will be more remarkable. In view of the above drawbacks, this paper refers to the kernel function based on Chebyshev polynomial. The Chebyshev kernels avoid to new parameter optimization problems, because of changing the kernel function. Test results show that the generalized Chebyshev kernel approaches to the minimum support vector number for classification in general. Combining with the excellent performance of Gaussian kernel, generalized Chebyshev kernels were properly improved to obtain modified Chebyshev kernels, and the modified Chebyshev kernel approaches the maximum classification performance.At the same time, the paper also proposed a training sample pre-selection algorithm of multi-class classification support vector. Kernel Fuzzy C-means clustering is a typical and dynamic clustering algorithm, and the advantages of kernel is that it can non-linearly map the model space data to high dimensional feature space. This method is based on Kernel Fuzzy C-means clustering and combined the idea of one-versus-one method in multiclass classification support vector machine. According to the established guidelines, it pre-selected the sample data which is likely to belong to the support vectors in training sample set and considerably reduced the training time of support vector machines. This paper takes validation experiments on several different vocabulary and the SNR of the speech database, and the experiment has achieved satisfactory result in speech recognition application. By the method, the learning efficiency and generalization ability of support vector machine classifier can be significant improved.