Abstract: Support vector machine (SVM) training is difficult for large-scale data set of speech recognition. A new SVM pre-extracting algorithm was proposed. On the one hand, kernel Fuzzy C-Means clustering was separately performed on each class of original data set. All the cluster centers were as a representative set of each class. On the other hand, according to the geometric distribution of support vectors and combined with the classification strategy of one-versus-one for SVM multi-class classification algorithm, boundary samples were extracted as support vectors for SVM to training and prediction. The algorithm was applied to embedded speech recognition system. Experiments indicate that this method improves the efficiency of training but also maintains the high recognition rate.

Key words: support vector, multi-class classification, kernel fuzzy C-Means clustering, sample pre- extracting, speech recognition system simulation