Based on statistical learning theory, [Vapnik, 1995] formulates the Support Vector Machine. SVM claims to guarantee generalisation, i.e. the decision rules reflects the regularities of the training data rather than the incapabilities of the learning machine. It also allows various other learning machines to be constructed under a unified framework, hences simplifying comparisons and promoting understanding.
Recently the Neural Network community has shown great interest in SVM. It has been applied to various tasks including pattern classification and regression estimation. In pattern classification, very good results [Osuna et al., 1997b] have been reported. [Schmidt and Gish, 1996] has applied SVM to a speaker identification task and has reported that SVM gives slightly better performance than the BBN modified Gaussian system on the difficult Switchboard task.
These interests and activities, coupled with the attractive claim in the SVM formulation, has motivated this critical study and investigation of SVM. The aim of this project is to understand SVM and the concepts that form its basis. Following this, a practical implementation of SVM will be developed and will be tested on speech pattern classification tasks.
In Chapter 
, the concepts and principles that form the basis for
SVM are reviewed. Chapter and Chapter detail
the training of SVM and its multi-class extension respectively. The
difficulties in tuning the SVM classifier for good performance will be
discussed in Chapter .