Projections non linéaires, estimation de la fiabilité d'un capteur et fusion dépendante du contexte pour la reconnaissance audiovisuelle de parole dans le bruit

If we want to exploit fully the auditory and visual complementarity, and realize an audiovisual speech recognition system robust in noise, we have to insert in the sensor fusion process a contextual information about the sensors reliability. Here, we show how efficiently estimate the acoustic sensor reliability thanks to a une "Curvilinear Component Analysis" (CCA), and the way of insert this context estimated into the fusion process in two recognition architectures.