Significant advances in human-computer interaction will require systems which can exhibit truly cognitive behaviour. This is particularly so in spoken dialogue systems (SDS) where, despite wide deployment and significant investment, current systems are still limited in capability and fragile to changes in environment or application. Recent advances in statistical modelling and machine learning offer the potential for making a significant step forward in SDS. By both exploiting and extending these advances, the CLASSiC project will improve generalization to unexpected situations. By modelling the whole end-to-end system as an integrated statistical process, the CLASSiC project will demonstrate a qualitative leap in the adaptivity, flexibility, robustness, and naturalness of SDS.The CLASSiC partners will develop a modular processing framework with an explicit representation of uncertainty which connects the various sources of uncertainty (understanding errors, ambiguity, etc) to the constraints to be exploited (task, dialogue, and user contexts). This architecture will support a layered hierarchy of supervised learning and reinforcement learning in order to facilitate mathematically principled optimisation and adaptation techniques. The architecture will be developed in close cooperation with our industrial partner in order to ensure that it provides a practical deployment platform as well as a flexible research test-bed.The resulting CLASSiC SDS will be able to adapt autonomously both to the needs of different users and to changing operating environments, and to learn through experience. The data-driven methodology will also enable faster and lower-cost system implementation through automatic optimisation. Overall, the project will demonstrate not only a step-change in the capability of practical spoken dialogue systems, it will also mark a significant step forward in the longer term goal of endowing autonomous systems with truly human-like capabilities.