I will describe some of the recent work at CSTR on conversational speech recognition. In particular, I will focus on language modelling and the use of dialogue structure (inferred via intonational cues). The small amount of training data is a particular problem in this field, and some techniques for generating probabilistic language models from small corpora will be described.
The difficulties of integrating information from acoustic, intonational, language model and dialogue structure sources will be examined, and some possible methods proposed.
To download this paper, please return to Proceedings of the 1997 Postgraduate Conference