ODYSSEY 2004 - The Speaker and Language Recognition Workshop
May 31 - June 3, 2004
In this article, it is studied the usefulness of the support vector machines (SVM) algorithm in the active classification of voice records into the sets normal and pathologic. In practice, each one of the samples employed on the classifier training must be manually labelled by an specialist, increasing in this way the training cost. Thus, it is imperative to obtain a classifier with a low generalization error, such that the number of training samples is as low as possible. A model selection technique, namely the Leave-One-Out criterion, was applied for the tuning of the appropriate parameters of the SVM. Also, a Radial Basis Function kernel was employed. The results obtained in the categorization of the aforementioned voice records showed that the number of tagged training samples can be reduced up to a 70% for the same testing error that the one obtained when the whole training set is labelled.
Bibliographic reference. Rengifo, Carlos Lino / Alvarez, Diego Andrés / Henao, Ricardo / Castellanos, Germán / Hurtado, Jorge Eduardo (2004): "Active learning on the classification of voice pathologies", In ODYS-2004, 271-274.