This paper describes the lexical access component which forms part of a complete speech understanding system being developed within the ESPRIT Sundial project (P2218). The objective function of a lexical access component within a dialogue system is discussed. The paper presents results from application-specific tests. Recognition results are given for tests where the lexical access component is used alone and where it used in conjunction with a parser. A method for using lexical access to directly assess the performance of phoneme models is also presented.