Two-level modelling of speech variant rules

Abstract

This paper describes a phonetic knowledge base for German consisting of a set of speech variant rules. These rules have been established on the basis of empirical, corpus-based investigations enriched by linguistic generalisations. Theoretical and computational foundations of speech variant rules are discussed, and their practical application in a linguistic word recognition system (BELLEx3, U Bielefeld) is demonstrated. Although the speech variant rules described in this paper have been established for the purpose of knowledge-based word recognition, their declarative implementation in a two-level transducer enables them to be employed for both recognition and generation of speech variants. Finally, an extension of standard two-level techniques is described whereby two-level transducers defining constraints on mapping relations between input and output forms are integrated with wellformedness-constraints on input forms stated in terms of finite-state automata.