Sie befinden Sich nicht im Netzwerk der Universität Paderborn. Der Zugriff auf elektronische Ressourcen ist gegebenenfalls nur via VPN oder Shibboleth (DFN-AAI) möglich. mehr Informationen...
Human speech perception considers both the auditory and visual nature of speech. Speech is more intelligible if the face of the talker can be seen and this is especially so in noisy conditions. More robust automatic speech recognition is possible if visual speech cues can be integrated with traditional acoustic systems.This thesis discusses the problems of visual speech parameterisation from mouth image sequences for use in audio-visual speech recognition. Five new lipreading techniques are evaluated using a hidden Markov model based visual-only recognition task and compared with an enhanced implementation of a previous lip contour tracker.The best methods are tested on two different multi-talker audio-visual databases to compare performance across different tasks. Combined audio-visual performance is tested using both early and late integration schemes.The addition of visual information to automatic speech recognition is found to improve accuracy and this is most pronounced in acoustically noisy conditions.Real-time implementations of two of the proposed methods demonstrate that the extension to audio-visual speech recognition is not impractical using current desktop technology.