Coupling Phonology and Phonetics in a Constraint-Based Gestural Model

9 veriied for German by means of electromagnetic articulography. Such a study should then also incorporate medial clusters as well as consonantal nuclei. Also further linguistic dimensions such as word accent etc. deserve closer investigation as to their innuence on timing. Machine learning of parametric time models from a gesturally segmented and annotated corpus should constitute a promising approach in this respect (cf. van Santen 1993). On the other hand the factual realisation of phonological processes must be studied and modelled. Does nal devoicing possess a gradual nature, being moreover conditioned by pragmatic factors (Port & Crawford 1989)? Can Ger-man /8/ realization (Diphthong dift=8] { diphthongieren : : : 8giqKKn]) be conceived of as intrasegmental variation of gestural phasing relationships? What about the syllable-dependent processes of /K/-allophony (frisch fKiM] { Berg b k] { Bauer bauj ]), of ich/ach alternation combined with /g/ spirantisation 16 and schwa epenthesis? Can the latter be modelled by short, ge-sturally unspeciied temporal gaps, as hypothesized in Walther (1992)? In conjunction with electromagnetic articulography experiments such hyphotheses can be evaluated very concretely in our approach. The approach presented here addresses the problem of phonology{phonetics coupling by assuming a declarative constraint-based phonology component, which is realized in the constraint logic programming language CUF extended by arithmetic constraints (Walther 1992, 1993). In order to couple this component in a simple fashion and without explicit interfaces to the articulatory synthesizer of Krr oger (1993 a,b), the central concept of gestures is employed. Initial practical results in synthesis demonstrate the general capabilities of our approach. Thus we may view our approach as a useful tool in future research directed towards an empirically anchored demarcation of linguistic phenomena: whether a given instance is best described as categorial-phonological or gradual-phonetic may be evaluated relative to an implemented model which can encompass both. Importantly, the typed attribute-value structure in (7) contains all the information that is needed by the articulatory synthesizer in order to compute the sound event it denotes. 13 In particular, the missing explicit interface documents itself by the absence of e.g. an additional phonetics attribute together with a corresponding complex interpretational relationship between phon and phonetics values. 5 Discussion and summary The nite-state and uniication-based approaches of (Bird 1992) and (Bird & Klein 1993) share the concerns of this work in explicitly advancing the enrichment of HPSG by declaratively speciied phonological information, but lack the quantitative-phonetic orientation pursued here. …