Skip to Main Content
Support vector machine (SVM) is a promising pattern classification technique that implements the structural risk minimization principle (SRM) in statistical learning theory. This paper proposes a new improved SVM, called class fusion support vector machine or FSVM, which has more robustness to noise and outliers than the standard SVM. We present an investigation into the application of FSVM to the out-of-vocabulary (OOV) rejection problem in a DTW based real-time ASR system. The feature vector consisting of parameters such as normalized N-best word scores and their 1st differences are directly derived from the recognition results as input to the OOV rejection process. The performance of the proposed FSVM classifier is compared with the standard SVM and neural networks.
Machine Learning and Cybernetics, 2004. Proceedings of 2004 International Conference on (Volume:6 )
Date of Conference: 26-29 Aug. 2004