Speech for a Construction Task Domain
Hans Brandt-Pook, Gernot A. Fink, Sven Wachsmuth, Gerhard Sagerer*
Technische Fakultdt, Angewandte Informatik Universität Bielefeld
Postfach 100131, 33501 Bielefeld, Germany
http://www.TechFak.Uni-Bielefeld. DE/ gernot/
The development of speech processing front-ends for the controlling of complex systems has received more and more interest during the last years. Usually this task is divided in two subtasks. The speech recogniser records the utterance and puts out a corresponding text, and the speech understanding: module tries to extract an internal representation of the meaning of the utterance. As shown in Figure I the recogniser uses an acoustic model as well as a statistical language model, which are both generated from a corpus. The speech understanding uses linguistic and domain specific knowledge designed by experts.
Figure 1: (leneral architecture oil' speech processing systems
Traditionally speech processing systems are organised strictly horizontally under two aspects. Firstly, the knowledge represented in the speech recogniser is completely independent from the linguistic knowledge used in the understanding component. Secondly, within the linguistic analysis the utterances are processed sequentially in several layers of abstraction (syntax, semantics, pragmatics). These general principles can be found in most speech processing projects (see e.g. ( Eckert et al. 93, Lavie et al. 97)). This approach seems to be motivated____________________