Statistical methods for building robust spoken dialogue systems in an automobile

We investigate the potential of statistical techniques for spoken dialogue systems in an automotive environment. Specifically, we focus on partially observable Markov decision processes (POMDPs), which have recently been proposed as a statistical framework for building dialogue managers (DMs). These statistical DMs have explicit models of uncertainty, which allow alternative recognition hypotheses to be exploited, and dialogue management policies that can be optimised automatically using reinforcement learning. This paper presents a voice-based in-car system for providing information about local amenities (e.g. restaurants). A user trial is described which compares performance of a trained statistical dialogue manager with a conventional handcrafted system. The results demonstrate the differing behaviours of the two systems and the performance advantage obtained when using the statistical approach