|Steve Young (Cambridge University Engineering Department)||University of Cambridge|
At Interspeech 2002 in Denver, I suggested that it should be possible to build a complete spoken dialogue system in which every component was based on a statistical model with parameters estimated from data. The potential advantages of such a system would include lower development cost, increased robustness to noise and the ability to learn on-line so that performance would continue to improve over time.Eight years later, fully statistical systems have now been built in the laboratory and their potential demonstrated. This talk will review the basic principles of statistical dialogue systems and discuss the major lessons learnt so far. The focus will be on dialogue management and in particular the representation of dialogue state, approaches to belief monitoring, parameter estimation and policy optimisation. Probabilistic components for speech understanding, natural language generation and synthesis will also be covered. The talk will end with a discussion of future challenges and the direction ahead.