Application Of Improved SVM In Low SNR Speech Recognition

With the development of science and technology, computer becomes anindispensable part of the life bringing great convenience for people’s study andlife. Compared with the mouse、keyboard、touch screen and other equipment，speech recognition becomes one of the more convenient way of thehuman-computer interaction, while speech recognition is an important way toachieve intelligent.Support vector machines is a new machine learning algorithms based onVC dimension theory and structural risk minimization. It can effectively solveproblems with small sample size, non-linear problem and high dimensions. Dueto these advantages, SVM has become a hot research field of machine learningalgorithms. And more and more scholars committed to the research of its basictheory and improved methods.Support vector machines have achieved very good results in speechrecognition, but its performance in the low SNR speech recognition needs to befurther improved. Fuzzy support vector machine has good noise immunity. Inthis paper firstly fuzzy support vector machines is studied, and to take thecontribution made by every feature into account，fuzzy support vector machines is improved by feature weighted. Finally，experiments on Korean speechdatabase showthe effectiveness of the algorithm.AdaBoost, adaptive boosting, one of weighted integration algorithm, hassolid theoretical foundation and has achieved great success in the field ofmachine learing. AdaBoost proposed a new idea for machine learning algorithm.When it is difficult to establish strong classifier directly, multiple weakclassifiers are integrated into strong classifier. In this article, AdaBoost.M2series for multi-classification are combined with SVM, and in order to improvethe performance further, GeesePSO is chosen to optimize the weights tocompose the strong classifier. Finally，experiments on Korean and Aurora-2speech database show the effectiveness of AdaBoost.M2-SVM andAdaBoost.M2-SVM improved by GeesePSO.