Publikation

Two-Layered Speaker Classification Using Dynamic Bayesian Networks

Christian Müller

In: Proceedings of the IASTED International Conference on Computational Intelligence. IASTED International Conference on Computational Intelligence (CI-2006) November 20-22 San Francisco CA United States IASTED 2006.

Abstrakt

This paper describes the two-layered AGENDER speaker classification approach which primarily recognizes the speaker's age and gender but also incorporates novel domain-independent pattern classification aspects which can be applied to acquire other speaker characteristics like emotions or cognitive load. The described approach dis tinguishes itself by means of a special post-processing technique: On the so called Second Layer, multiple post processing problems are solved with one single mecha nism, namely dynamic Bayesian networks (DBNs). While the actual classification aspect " the First Layer " is merely summarized, this paper focuses on the description of the Second Layer. Particularly, examples are provided on how DBNs can be used for: explicitely modeling the classification inherent uncertainty, incorporating top down knowl edge into the decision making process, and fusing the re sults of multiple classifiers. The paper finishes with a sum mary as well as an outlook to future extensions.

Deutsches Forschungszentrum für Künstliche Intelligenz
German Research Center for Artificial Intelligence