Speech recognition has made major progress over the last decade. Nevertheless, state-of-the-art speech recognizers generally perform significantly worth than humans do. Research for improvement in several direction has been done in the framework of COST 249 (http://www.elis.rug.ac.be/ELISgroups/speech/cost249). In this framework, the project ARCOS-G at the ETHZ aimed at a speech recognizer which takes into account, that speech recognition must be based on different kinds of knowledge, e.g. statistical knowledge of phoneme properties vs. discrete linguistical knowledge (lexicon, syntax, etc.). Consequently, the adequate approach to a speech recognizer has, in order to cope with all required knowledge, to include statistical and knowledge-based parts. The ARCOS project has shown the feasibility of such a speech recognizer. The actual system, however, lacks a good search strategy component to efficiently drive the parser to the optimum solution (for further information, please consult the project web page at
http://www.tik.ee.ethz.ch/~spr).