Publikation

A Speaker Classification Framework for Non-intrusive User Modeling: Speech-based Personalization of In-Car Services

Michael Feld

PhD-Thesis Computer Science Institute, Saarland University 12/2011.

Abstrakt

Speaker Classification, i.e. the automatic detection of certain characteristics of a person based on his or her voice, has a variety of applications in modern computer technology and artificial intelligence: As a non-intrusive source for user modeling, it can be employed for personalization of human-machine interfaces in numerous domains. This dissertation presents a principled approach to the design of a novel Speaker Classification system for automatic age and gender recognition which meets these demands. Based on literature studies, methods and concepts dealing with the underlying pattern recognition task are developed. The final system consists of an incremental GMM-SVM supervector architecture with several optimizations. An extensive data-driven experiment series explores the parameter space and serves as evaluation of the component. Further experiments investigate the language-independence of the approach. As an essential part of this thesis, a framework is developed that implements all tasks associated with the design and evaluation of Speaker Classification in an integrated development environment that is able to generate efficient runtime modules for multiple platforms. Applications from the automotive field and other domains demonstrate the practical benefit of the technology for personalization, e.g. by increasing local danger warning lead time for elderly drivers.

Projekte

Deutsches Forschungszentrum für Künstliche Intelligenz
German Research Center for Artificial Intelligence